WO2015041295A1 - Terrain category classification method, terrain category classification program, and terrain category classification device - Google Patents

Terrain category classification method, terrain category classification program, and terrain category classification device Download PDF

Info

Publication number
WO2015041295A1
WO2015041295A1 PCT/JP2014/074699 JP2014074699W WO2015041295A1 WO 2015041295 A1 WO2015041295 A1 WO 2015041295A1 JP 2014074699 W JP2014074699 W JP 2014074699W WO 2015041295 A1 WO2015041295 A1 WO 2015041295A1
Authority
WO
WIPO (PCT)
Prior art keywords
type
ground surface
vector
position vector
relationship
Prior art date
Application number
PCT/JP2014/074699
Other languages
French (fr)
Japanese (ja)
Inventor
廣瀬　明
方 尚
Original Assignee
国立大学法人東京大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国立大学法人東京大学 filed Critical 国立大学法人東京大学
Priority to JP2015537966A priority Critical patent/JPWO2015041295A1/en
Publication of WO2015041295A1 publication Critical patent/WO2015041295A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
    • G01S13/90—Radar or analogous systems specially adapted for specific applications for mapping or imaging using synthetic aperture techniques, e.g. synthetic aperture radar [SAR] techniques
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
    • G01S7/024—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using polarisation effects
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
    • G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
    • G01S13/90—Radar or analogous systems specially adapted for specific applications for mapping or imaging using synthetic aperture techniques, e.g. synthetic aperture radar [SAR] techniques
    • G01S13/9021—SAR image post-processing techniques
    • G01S13/9027—Pattern recognition for feature extraction
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
    • G01S13/90—Radar or analogous systems specially adapted for specific applications for mapping or imaging using synthetic aperture techniques, e.g. synthetic aperture radar [SAR] techniques
    • G01S13/904—SAR modes
    • G01S13/9076—Polarimetric features in SAR

Definitions

  • the present invention relates to a ground surface classification method, a ground surface classification program, and a ground surface classification device, and more specifically, from an irradiation wave irradiated to a plurality of areas of the ground surface and a scattered wave obtained by irradiation of the irradiation wave.
  • a ground type classification method, a ground type classification program, and a ground type classification that classify ground types of a plurality of regions based on type relationship learning that learns a type relationship that is a relationship between radio wave information and a ground type. Relates to the device.
  • this kind of surface classification method is based on natural information using the polarization information of radio waves emitted from the satellite or aircraft to the ground surface and the polarization information of scattered waves from the ground surface obtained by the emission of these radio waves.
  • a method for determining the terrain and the artificial terrain has been proposed (see, for example, Patent Document 1).
  • This method uses the covariance matrix (Covariance Matrix, hereinafter referred to as “C matrix”) of each variability of the scattering matrix obtained by modeling the scattering of the ground surface into the scattering of the natural terrain and the scattering of the urban area.
  • the natural terrain and the artificial terrain are classified by determining the terrain and the artificial terrain.
  • the ground surface type classification method, the ground surface type classification program, and the ground surface type classification device of the present invention are mainly intended to classify ground surface types with high accuracy.
  • ground surface type classification method employs the following means in order to achieve the main purpose described above.
  • the ground surface classification method of the present invention is: Input radio wave information consisting of an irradiation wave irradiated to a plurality of areas on the ground surface and scattered waves obtained by the irradiation of the irradiation wave, and learn a type relationship that is a relationship between the radio wave information and the surface type.
  • the ground surface type classification method for classifying the ground surface types of the plurality of regions based on type relationship learning
  • the type relationship learning a position vector of Poincare spheres of polarization information for a plurality of regions and a variation vector indicating a variation from a position vector of a neighboring region of the position vector are expanded to a plurality of four quaternions.
  • the gist is to learn the type relationship by using a neural network with an elemental vector as an input value.
  • radio wave information consisting of an irradiation wave irradiated to a plurality of areas of the ground surface and a scattered wave obtained along with the irradiation of the irradiation wave is input, and the radio wave information and the ground surface type are input.
  • the ground surface types of a plurality of regions are classified based on the type relationship learning for learning the type relationship which is Type relationship learning is a method of expanding a quaternion vector obtained by extending a position vector in the Poincare sphere of polarization information for a plurality of regions and a variation vector indicating a variation of the position vector from a neighboring position vector to a quaternion.
  • the type relationship is learned using a neural network as an input value. Since the position vector and the variation vector reflect the polarization information, the type relationship can be properly learned by learning the type relationship using the position vector and the variation vector. Further, since the type relationship is learned using both the position vector and the variation vector, the type relationship can be learned more appropriately as compared with the case where only one of the position vector and the variation vector is used. Furthermore, since the type relationship is learned using a neural network using a plurality of quaternion vectors obtained by expanding the position vector and the variation vector to a quaternion, the type relationship can be learned appropriately. By classifying the ground surface types of a plurality of regions based on the properly learned type relationship in this way, the ground surface types can be classified with high accuracy.
  • the type relationship learning is performed by causing each component of the position vector to correspond to three imaginary parts and the first value to correspond to the real part as the input value.
  • a quaternion position vector obtained by expanding a vector to a quaternion, and each component of the variation vector are made to correspond to three imaginary parts and a second value is made to correspond to a real part to make the variation vector a quaternion.
  • an extended quaternion variation vector is learned more appropriately.
  • the “first value” can be determined as appropriate, and a value of 0 is particularly preferable.
  • the “second value” can be determined as appropriate, and the value of 0 is particularly preferable.
  • the type relationship learning is learning performed using a position vector in a Poincare sphere of polarization information for a region where a predetermined ground surface type is known as the position vector, Based on the type relationship in the type relationship learning and the input radio wave information, the degree of relevance for each surface type of the plurality of regions is calculated, and based on the calculated degree of relevance for each surface type The ground surface types of a plurality of areas may be classified. Since the type relationship learning is performed using the position vector in the Poincare sphere of polarization information for a known region, the type relationship can be learned more appropriately. Since the ground type is classified using the learned type relationship, the ground type can be classified with higher accuracy.
  • the type relationship learning is performed when the irradiation wave is set to one of a horizontally polarized wave and a vertically polarized wave as the input value, and when the irradiation wave is a 45 degree polarized wave and ⁇ 45
  • the three positions in each of the polarization state of one degree of polarization and the polarization state of the left-handed circular polarization and right-handed circular polarization of the irradiation wave It is also possible to include six quaternion vectors obtained by extending the vector and the three variation vectors to quaternions.
  • the type relation learning is a value of 0. It is also possible to input a threshold vector, which is a quaternion in which the third value that does not correspond to the real part and the value 0 corresponds to the three imaginary parts, as the threshold value of the neural network.
  • the “third value” can be determined as appropriate, and is particularly preferably a value of ⁇ 1.
  • the region type includes n different types of area types. It is also possible that the neural network has the n vectors whose inner products are 0 as the output values expected for the input values. In this way, it is possible to classify the surface type more appropriately.
  • the ground surface types may include at least four types of lake, grassland, forest, city, and desert. In this way, it is possible to classify a plurality of areas on the ground surface into at least four types of lakes, grasslands, forests, towns, and deserts.
  • the ground surface classification program of the present invention is A ground type classification program for classifying ground types of a plurality of areas of the ground surface,
  • a radio wave information input module for inputting radio wave information composed of an irradiation wave irradiated to a plurality of regions of the ground surface and a scattered wave obtained by the irradiation of the irradiation wave;
  • Multiple quaternion vectors obtained by extending the position vector of the polarization information for multiple areas in the Poincare sphere and the variation vector indicating the variation from the position vector of the neighboring area of the position vector to a quaternion
  • a type relationship learning module that learns a type relationship that is a relationship between the radio wave information and the surface type using a neural network
  • a ground surface type classification module for classifying the ground surface types of the plurality of regions based on the learned type relationship; It is a summary to provide.
  • radio wave information consisting of irradiation waves irradiated to a plurality of areas of the ground surface and scattered waves obtained by irradiation of the irradiation waves is input, and polarization information for the plurality of areas is input.
  • Type relationship using a neural network with multiple quaternion vectors obtained by expanding the position vector in the Poincare sphere of the quaternion and the variation vector indicating the variation of the position vector from the position vector of the neighboring area into a quaternion To learn. Since the position vector and the variation vector are values that reflect the polarization information, the type relationship can be properly learned by learning the type relationship using the position vector and the variation vector.
  • the type relationship is learned using both the position vector and the variation vector, the type relationship can be learned more appropriately as compared with the case where only one of the position vector and the variation vector is used. Furthermore, since the type relationship is learned using a neural network using a plurality of quaternion vectors obtained by expanding the position vector and the variation vector into a quaternion, the type relationship can be learned more appropriately. Then, the ground surface types of the plurality of regions are classified based on the learned type relationship. Since the ground surface types of a plurality of regions are classified based on the properly learned type relationship, the ground surface types can be classified with high accuracy.
  • the ground surface classification apparatus of the present invention is A ground type classification device that classifies ground types of a plurality of areas of the ground surface,
  • a radio wave information input unit for inputting radio wave information composed of an irradiation wave irradiated to a plurality of areas on the ground surface and a scattered wave obtained by irradiation of the irradiation wave;
  • Multiple quaternion vectors obtained by extending the position vector of the polarization information for multiple areas in the Poincare sphere and the variation vector indicating the variation from the position vector of the neighboring area of the position vector to a quaternion
  • a type relationship learning unit that learns a type relationship that is a relationship between the radio wave information and the surface type using a neural network
  • a ground surface type classification unit for classifying the ground surface types of the plurality of regions based on the learned type relationship; It is a summary to provide.
  • radio wave information including irradiation waves irradiated to a plurality of areas on the ground surface and scattered waves obtained by irradiation of the irradiation waves is input, and polarization information for the plurality of areas is input.
  • Type relationship using a neural network with multiple quaternion vectors obtained by expanding the position vector in the Poincare sphere of the quaternion and the variation vector indicating the variation of the position vector from the position vector of the neighboring area into a quaternion To learn. Since the position vector is the polarization information itself, the type relationship can be properly learned by learning the type relationship using the position vector.
  • the type relationship is learned using both the position vector and the variation vector, the type relationship can be learned more appropriately as compared with the case where only one of the position vector and the variation vector is used. Furthermore, since the type relationship is learned using a neural network using a plurality of quaternion vectors obtained by expanding the position vector and the variation vector into a quaternion, the type relationship can be learned more appropriately. Then, the ground surface types of the plurality of regions are classified based on the learned type relationship. Since the ground surface types of a plurality of regions are classified based on the properly learned type relationship, the ground surface types can be classified with high accuracy.
  • FIG. 12 is a reference diagram illustrating an example of a final output Pout-type for each local table type. It is the schematic diagram which expressed the self-organization map in the feature-value space showing a polarization state. It is the model which expressed the self-organization map in the SOM space showing the connection condition of a neuron. It is a flowchart which shows an example of the surface classification classification program 150 performed by CPU32.
  • FIG. 1 is a configuration diagram showing an outline of a configuration of a synthetic aperture radar system 10 including a ground surface classification device 30 as a first embodiment of the present invention.
  • the synthetic aperture radar system 10 includes an artificial satellite 20 and a ground surface type classification device 30 that inputs various information from the artificial satellite 20 and classifies the ground surface into one of lake, grassland, forest, and city. .
  • the artificial satellite 20 irradiates a plurality of regions on the surface with laser light, detects scattered light of the irradiated laser light, and detects the irradiation position of the laser light, the polarization information of the irradiated laser light, and the polarization of the scattered light.
  • Various kinds of data such as PolSAR (Polarimetric Synthetic Apertture ⁇ Radar) image data including information are output to the ground surface classification device 30.
  • the ground surface classification device 30 communicates information from a well-known CPU 32, ROM 34, RAM 36, hard disk drive (HDD) 38, input / output processing circuit 40, and satellite 20 connected via a bus 31.
  • a receiving circuit 42 is provided.
  • a receiving circuit 42 To the input / output processing circuit 40, a receiving circuit 42, a disk drive device 43, a keyboard 44, and a mouse 45 are connected so that data can be input, and a display 46 that outputs information is connected so that data can be input and output.
  • the HDD 38 stores a ground surface classification program 50 installed as application software.
  • the ground surface type classification program 50 includes a PolSAR image input module 52 that inputs PolSAR image data from the artificial satellite 20, a type relationship learning module 54 that learns a type relationship that is a relationship between the PolSAR image data and the type of the ground surface, It is composed of a ground surface type classification module 56 that classifies the ground surface type of each region of the ground surface based on the type relationship learned by the relationship learning module 54.
  • FIG. 2 is a flowchart showing an example of the ground surface type classification program 50 executed by the CPU 32.
  • the surface classification classification program 50 is written by the CPU 32 at a predetermined address in the RAM 36 when the execution is instructed by the user. Then, the CPU 32 reads and executes the ground surface type classification program 50 written in the RAM 36.
  • the CPU 32 When the ground surface classification program 50 is executed, the CPU 32 first executes a process of inputting PolSAR image data from the artificial satellite 20 (step S100), and a position vector Pv, which will be described later, is changed from the input PolSAR image data. Based on the set position vector Pv, the input vector Xi obtained by expanding the set position vector Pv and the variation vector Vv to a quaternion, and the type relationship learned in advance, the observed area is defined as a lake and grass (grass) ), Forest, town (town) (step S110).
  • the explanation of the ground surface type classification program 50 is interrupted, and the learning of the position vector Pv, the fluctuation vector Vv, the quaternion, and the type relationship will be described.
  • the position vector Pv and the variation vector Vv will be described.
  • the irradiation wave Iw is expressed by using the normalized Jones vector represented by Expression (1)
  • the scattered wave Rw can be expressed by Expression (2) using the scattering matrix S
  • the Jones coherence matrix J is expressed by Expression (3).
  • ⁇ > indicates that they are averaged temporally or spatially.
  • the averaged Stokes vector is calculated by Equation (4) based on the Jones coherence matrix J.
  • the averaged Stokes vector includes polarization information of the scattered wave from the ground surface, and the degree of polarization (Degree of Polarization) DoP is calculated using Equation (5). Since the Jones coherence matrix J is a complex Hermitian semi-definite matrix,
  • the reflected wave is completely polarized when the degree of polarization Dop is 1, and is completely unpolarized when the degree of polarization DoP is 0, and the degree of polarization Dop is greater than 0. When the value is less than 1, it is partially polarized.
  • the averaged Stokes vector can be expressed as one point on the Poincare sphere or in the Poincare sphere, and the position coordinates (x, y, z) of this point can be expressed by Equation (6).
  • the position vector Pv can be obtained for each pixel in the PolSAR image, and indicates an average state of polarization.
  • the variation between the position vector Pv and the position vector Pv of the pixel in the vicinity of the position vector Pv is calculated by the equation (7) as an average deviation.
  • N is the number of pixels used for one calculation
  • x, y, and z are average values of x in a plurality of pixels used for one calculation.
  • the variation vector Vv indicates the distribution of the position vector Pv in the Poincare sphere.
  • the position vector Pv and the variation vector Vv are named “Poincare sphere parameters”.
  • FIG. 3 is an explanatory diagram showing a satellite photograph of the Fuji Susono area and a drawing sketched for each type of satellite photograph.
  • four black squares indicate the observation areas of the lake, grassland, forest, and city (hereinafter referred to as “Group 1”), and four white squares indicate the lake, grassland, forest, and city.
  • Each observation area (hereinafter referred to as “Group 2”) is shown.
  • Each square includes 40 ⁇ 40 observation areas.
  • the Jones coherence matrix J is derived from the above-described equation (3).
  • the Jones coherence matrix J is derived as a spatial average value.
  • the relationship between the window size and the average polarization degree was examined using the PolSAR image data of groups 1 and 2.
  • FIG. 4 shows the relationship between the number of pixels on one side of the observation area and the average polarization degree.
  • the window size is 5 ⁇ 5
  • the window size is 9 ⁇ 9.
  • the average polarization degree becomes a constant value when the window size is larger than 5 ⁇ 5.
  • the window size for obtaining the position vector Pv is set to 5 ⁇ 5.
  • the relationship between the window size and the norm of the variation vector Vv was examined using the PolSAR image data of groups 1 and 2.
  • FIG. 5 shows the relationship between the number of pixels on one side of the observation region and the norm of the variation vector Vv.
  • the magnitude of the variation vector Vv becomes a constant value when the window size is larger than 9 ⁇ 9. Therefore, in the embodiment, the window size for obtaining the variation vector Vv is set to 9 ⁇ 9.
  • FIG. 6 is an explanatory diagram showing the position vector Pv of the scattered wave when the irradiation wave is horizontally polarized
  • FIG. 7 is an explanatory diagram showing the fluctuation vector Vv of the scattered wave when the irradiation wave is horizontal polarization.
  • FIG. 8 is an explanatory diagram showing the position vector Pv of the scattered wave when the irradiation wave is 45-degree polarized wave
  • FIG. 9 shows the fluctuation vector Vv of the scattered wave when the irradiation wave is 45-degree polarized wave
  • FIG. 10 is an explanatory diagram showing a position vector Pv of the scattered wave when the irradiation wave is left-handed circularly polarized wave
  • FIG. 11 is a diagram of the scattered wave when the irradiation wave is left-handed circularly polarized wave. It is explanatory drawing which shows the fluctuation vector Vv.
  • the position vector Pv and the variation vector Vv show different distributions on the graph if the types are different, and the position vector Pv and the variation vector Vv do not have the same distribution. Therefore, by classifying the type of the ground surface using both the position vector Pv and the variation vector Vv, it is possible to classify the ground surface type with higher accuracy than that using either the position vector Pv or the variation vector Vv.
  • the quaternion p is a four-dimensional number composed of one real part and three imaginary parts of bases i, j, and k that are orthogonal to each other. Can be expressed as a vector of In Equation (8), p e , p i , p j, and p k are real numbers.
  • the quaternion bases i, j, k follow the Hamilton rule shown in equation (9).
  • FIG. 12 is a block diagram for explaining the configuration of a quaternion neural network used for type relationship learning.
  • a supervised learning is performed using a neural network to learn the relationship between the Poincare sphere parameters and the type of the ground surface.
  • the number of hidden layer neurons is 8
  • the number of output layer neurons is 4
  • the threshold nodes of the input layer and hidden layer are ( ⁇ 1, 0, 0), respectively. , 0).
  • FIG. 13 is a flowchart illustrating an example of the type relationship learning routine.
  • the teacher learns from the PolSAR image data of four regions whose types of lake, grassland, forest, and town are known when the irradiation wave is horizontally polarized wave, 45 degree polarized wave, and left-handed circularly polarized wave.
  • Input data X in i-lake, X in i-grass, X in i-forest, X in i- which is obtained by selecting data and connecting the position vector Pv and variation vector Vv of the selected teacher data as components.
  • the process of setting to “down” (hereinafter, one of these four may be described as “X in i-type”) is executed (step S200).
  • i indicates i-th teacher data for various types of lakes, grasslands, forests, and towns.
  • the input vector X in i-type represents each component of the position vector Pv when the irradiation wave is a horizontally polarized wave, 45 degree polarized wave, and left-handed circularly polarized wave.
  • a quaternion position vector x H , x 45 °, x lc obtained by extending the position vector Pv to a quaternion by corresponding to three imaginary parts and a value 0 corresponding to one real part, and a variation vector Vv
  • the quaternion variation vectors ⁇ H , ⁇ 45 °, ⁇ lc, in which the variation vector Vv is expanded to a quaternion by associating each component with three imaginary parts and the value 0 with one real part, are connected. It was supposed to be a thing.
  • X in i-type [x H , ⁇ H , x 45 °, ⁇ 45 °, x lc , ⁇ lc ] (14)
  • x H (0, x, y, z) H
  • ⁇ H (0, ⁇ x, ⁇ y, ⁇ z) H
  • x 45 ° (0, x, y, z) 45 ° ...
  • ⁇ 45 ° (0, ⁇ x, ⁇ y, ⁇ z) 45 ° ...
  • x lc (0, x, y, z) lc (19)
  • ⁇ lc (0, ⁇ x, ⁇ y, ⁇ z) lc (20)
  • a coupling load w described later is set to a predetermined initial value and expected output vectors D lake , D grass , D forest, D town is set (step S210).
  • the expected output vectors D lake , D grass , D forest, and D town are expressed by equation (21) so that any two of the expected output vectors D lake , D grass , D forest, and D town are orthogonal to each other. It was supposed to be set.
  • equation (21) so that any two of the expected output vectors D lake , D grass , D forest, and D town are orthogonal to each other. It was supposed to be set.
  • the expected output vectors D lake , D grass , D forest, and D town it is possible to classify the surface type more appropriately.
  • the processing of steps S220 to S250 is executed for each type of lake, grassland, forest, and town, and processing for adjusting the coupling load wi of the neural network is executed.
  • the output vector Y is calculated using equation (22) with the i-th data among the teacher data of a certain type (for example, lake) as the input vector Xi (step S220).
  • wi is a connection weight in the neural network for connecting the input vector Xi and the output vector Y.
  • s is the initial state
  • N is the number of input nodes
  • an error value E is calculated using the equation (24), and the error value E is compared with a threshold err (for example, 5 ⁇ 10 ⁇ 4 ) (step S230) and step It is determined whether or not the number Nf indicating the number of times S220 has been executed is 100 (step S240).
  • the number of times Nf is set to a value 0 as an initial value, and is a number that increases by 1 every time step S220 is executed.
  • the value w old of the combined load wi is updated to the value w new using the equations (17) to (22) (step S250).
  • the processes in steps S220 to S250 are repeated until the error value E becomes smaller than the threshold value err or the number of times Nf becomes 100.
  • a description of subscripts and the like used in Expression (24) to Expression (28) is shown in FIG.
  • the coupling load w is adjusted so that the output vector Y is close to the expected output vectors D lake , D grass , D forest, D town with respect to the input vector Xin whose type is known.
  • Step S260 it is determined whether or not the processing of steps S220 to S250 is completed for the four types of lake, grassland, forest, and town ( (Step S260) If the processing of steps S220 to S250 is not completed for the four types, the type is changed (step S270) (for example, the type is changed from lake to grassland) and steps S220 to S250 are performed. Execute the process.
  • the average Serr of the error values E obtained for the four types is calculated using the equation (29), It is checked whether or not the average Serr is equal to or greater than the threshold err (step S280).
  • E i-lake , E i-grass , E i-forest , and E i-town are the error values E for each type used for the determination in step S230.
  • step S280 When the average Serr is greater than or equal to the threshold err (step S280), when the target teacher data is not the last (Nth) data (when i ⁇ N), the processing of steps S220 to S260 is performed for the next teacher data (The value 1 is added to the integer i and the integer i is updated).
  • the first teacher data is on the other hand
  • step S290 the processing of steps S220 to S260 is executed (step S290).
  • the processing of steps S220 to S290 is executed until the average Serr is less than the threshold err, and when the average Serr is less than the threshold err, this routine is terminated.
  • the output vector Y of equation (22) is a value near the expected output vector D lake
  • the output vector of equation (22) is a value near the expected output vector D grass
  • the output vector of equation (22) is a value reflecting the ground surface type.
  • the connection load wi associates the input vector Xi based on the Poincare sphere parameter with the output vector Y reflecting the ground surface type, and the connection load wi is learned in the type relationship learning process. Therefore, the type relationship learning process is a process of learning a type relationship that is a relationship between the radio wave information and the ground surface type. Since the position vector Pv and the variation vector Vv reflect the polarization information directly obtained from the PolSAR image data, the type relationship is more appropriately learned by learning the type relationship using the position vector Pv and the variation vector Vv. Can learn. Further, since both the position vector Pv and the variation vector Vv are used, more accurate learning can be performed. Furthermore, since the type relationship is learned using a neural network with the input vector X in i-type having the component of the position vector Pv and the variation vector Vv as input values, the type relationship is more appropriately learned. Can do.
  • the description returns to the ground surface classification program 50 illustrated in FIG.
  • PolSAR image data whose type is unknown from the artificial satellite 20 is input (step S100)
  • the position vector Pv and variation vector of the PolSAR image data when the irradiation wave is horizontally polarized wave, 45 degree polarized wave, and left-handed circularly polarized wave.
  • the region irradiated with the laser based on the input vector Xi having Vv as a component and the type relationship learned by the type relationship learning routine shown in FIG. 13 is one of lake, grassland, forest, and town. (Step S110).
  • the classification of the type is based on the input vector Xi obtained from the PolSAR image data (the component is the same as X in i-type in equation (14)) and the connection weight wi obtained by learning shown in FIG. Is used to calculate the output vector Y according to the above equation (22), and using the output vector Y and the expected outputs D lake , D grass , D forest, D town , the error value Etype ( (type, lake, glass, forest, down) is calculated, and the final output Pout-type is calculated by the equation (30) using the error value Etype.
  • the pixel is of a type close to the value 1, and if it is close to the value 0, it is not classified as that type. For example, if the final output Pout-lake is close to the value 1 and the final outputs Pout-glass, Pout-forest, Pout-town are close to the value 0, the pixel is classified as “lake” and the corresponding observation area is Classify as “Lake”.
  • FIG. 15 is a diagram illustrating the values of the final outputs Pout-lake, Pout-glass, Pout-forest, and Pout-town, which are the results of classifying the surface type using the method of the embodiment for the Fuji Susono area illustrated in FIG. FIG.
  • the upper left is the value of the final output Pout-lake in each region
  • the upper right is the value of the final output Pout-glass in each region
  • the lower left is the value of the final output Pout-forest in each region
  • the lower right is the Pout in each region.
  • -Town values are shown, and each value is expressed by shading so that the color becomes darker as the value 1 is approached.
  • the ground surface type classification method of the embodiment can favorably classify the ground surface type.
  • a type relationship is learned using a neural network using a plurality of quaternion vectors obtained by extending the position vector Pv and the variation vector Vv to a quaternion as an input vector Xi, and the ground surface type is determined using the learning result.
  • the ground surface type can be classified with higher accuracy.
  • a type relationship is established using a neural network with a plurality of quaternion vectors obtained by extending the position vector Pv and the variation vector Vv to a quaternion as an input vector Xi.
  • the type relationship can be learned more appropriately. Thereby, it is possible to classify the ground surface type with higher accuracy.
  • the ground surface is classified into four types, and expected outputs D lake , D grass , D forest, and D town whose inner products are 0 values are expected output values.
  • the ground surface is classified into five types, and five expected output vectors D lake , D grass , D forest, and D town whose inner products have a value of 0 may be used as expected output values.
  • the inner products may not have a value of 0.
  • the threshold node of the neural network illustrated in FIG. 12 is set to ( ⁇ 1, 0, 0, 0), but the threshold node takes into consideration the degree of load of the arithmetic processing. Can be determined as appropriate.
  • the ground surface type classification apparatus 130 as a second embodiment of the present invention will be described.
  • the ground type classification program 50 of the first embodiment performs supervised learning using a neural network
  • the ground type classification program 150 uses a neural network.
  • the configuration is the same as that of the ground surface type classification apparatus 30 of the first embodiment except that unsupervised learning is performed. Therefore, in the ground surface classification device 130, the same components as those of the ground surface classification device 30 are denoted by the same reference numerals, and the description thereof is omitted.
  • the ground type classification program 150 of the second embodiment self-organizes a plurality of quaternion vectors obtained by extending the position vector Pv and the variation vector Vv into quaternions as input vectors Xi, X in i-type as a neural network. Perform unsupervised learning using a map (self-organizig map: SOM) to classify pixel types.
  • FIG. 16 is a schematic diagram expressing the self-organizing map in a feature amount space representing a polarization state
  • FIG. 17 is a schematic diagram expressing the self-organizing map in an SOM space representing the connection state of neurons.
  • FIG. 18 is a flowchart showing an example of the ground surface type classification program 150 executed by the CPU 32.
  • the ground classification classification program 150 is executed, first, a process of inputting the PolSAR image data of the observation area is executed (step S300), and the combined loads wlake, wgrass, wforest for various types of lakes, grasslands, forests, and towns are executed.
  • Wtown is set to a predetermined initial value, and an initialization process is performed to set a value 1 to an integer i indicating the i-th pixel in the target Poincare sphere (step S310).
  • the created input vector Xi is input (step S320).
  • the input vector Xi includes three components of position vectors Pv H , Pv 45 °, and Pv lc in each polarization state of horizontal polarization, 45 degree polarization, and left-hand circular polarization.
  • a quaternion position vector x H , x 45 °, x lc is obtained by expanding the position vector Pv H , Pv 45 °, Pv lc to a quaternion by making the value 0 correspond to one real part, corresponding to the imaginary part.
  • each component of the fluctuation vectors Vv H , Vv 45 °, and Vv lc of each polarization state of horizontal polarization, 45 degree polarization, and left-hand circular polarization correspond to three imaginary parts and a value 0 is one real number. It is assumed that the quaternion variation vectors ⁇ H , ⁇ 45 °, and ⁇ lc obtained by extending the variation vectors Vv H , Vv 45 °, and Vv lc to quaternions are made to correspond to each other. Since the integer i is now set to 1, the input vector Xi created from the PolSAR image data of the first pixel is input in the process of step S320.
  • the distance (Xi, wc) between the input vector Xi shown in the equation (32) and each of the coupling loads wlake, wgrass, wforest, wtown is calculated, and the distance (Xi, wc) is calculated.
  • the neuron having the smallest connection weights wlake, wgrass, wforest, wtown (closest to the input vector Xi) is determined as the winner neuron, and the i-th pixel is classified into the winner class cw (step S330).
  • connection weight wlake is closest to the input vector Xi
  • the neuron having the connection weight wlake is determined as the winner neuron, and the i-th pixel is classified as a lake.
  • c is any one of rake, glass, forest, and town.
  • step S340 the weights of the neuron wcw of the winner class cw and the two neurons in the vicinity of the neuron wc (neighboring neurons in the SOM space) wcw ⁇ 1 are updated using the equations (33) to (35) (step S340). . Then, it is checked whether or not the processing of steps S320 to S340 has been completed for all the observation pixels (step S350). If the processing of steps S320 to S340 has not been completed for all the pixels, the integer i is set. The integer i is updated to the value obtained by adding 1 (step S360), and the process returns to step S310.
  • steps S320 to S340 when the processing of steps S320 to S340 has not been completed for all pixels, the processing of steps S320 to S340 is performed for the next pixel. Now, since the first pixel is considered, when the processing of steps S320 to S340 is completed for the first pixel, the processing of steps S320 to S340 is executed for the second pixel.
  • wcw wcw + ⁇ (Xi ⁇ wcw) (33)
  • wcw ⁇ 1 wcw ⁇ 1 + ⁇ (Xi ⁇ wcw) (34) 0 ⁇ ⁇ ⁇ ⁇ ⁇ 1 (35)
  • step S350 When the processing of S320 to S340 is executed for all such pixels (step S350), the processing of S320 to S350 is performed for all the pixels Nitr times (one or more times. For example, 50 times, 100 times, 150 times) It is determined whether or not the process has been repeated (step S370). When the processing of S320 to S350 has not been repeated for all pixels for Nitr times, the integer i is set to a value 1 (step S380), and the processing of steps S320 to S350 is repeated from the first pixel to all pixels. On the other hand, when the processes of S320 to S350 are repeated Nitr times (step S370), this routine is terminated.
  • PolSAR image data can be classified into one of the types of lake, grassland, forest, and city, that is, the surface area is classified into any type. Compared with those that are reasonably decomposed, it is possible to classify with higher accuracy.
  • a neural network is used with an input vector Xi obtained by expanding a position vector Pv and a variation vector Vv obtained from PolSAR image data into a quaternion as an input value.
  • an input vector Xi obtained by expanding a position vector Pv and a variation vector Vv obtained from PolSAR image data into a quaternion as an input value.
  • step S330 it is determined that the neuron having the connection weights wlake, wglass, wforest, wtown that is the closest to the input vector Xi is the winner neuron. 36) or the similarity (Xi, wc) between the input vector Xi shown in the equation (37) and each of the coupling weights wlake, wgrass, wforest, wtown, and the coupling weight wlake, with the largest similarity (Xi, wc).
  • a neuron having wgrass, wforest, wtown may be determined as a winner neuron.
  • each component of the position vector Pv corresponds to three imaginary parts in the input vectors Xi and X in i-type, and the value 0 is one real part.
  • the position vector Pv is expanded to a quaternion by making it correspond to, but each component of the position vector Pv is made to correspond to two imaginary parts and one real part, and the value 0 is made to correspond to the remaining imaginary part.
  • a value different from the value 0 may be associated with the real part.
  • each component of the variation vector Vv corresponds to three imaginary parts and a value 0 corresponds to one real part in the input vector X in i-type.
  • the variation vector Vv is expanded to a quaternion by performing the above, but each component of the variation vector Vv is associated with two imaginary parts and one real part, and the value 0 is associated with the remaining imaginary part.
  • the real part may be associated with a value different from 0.
  • the input vector X in i-type is a position vector Pv when the irradiation wave is horizontally polarized wave, 45 degree polarized wave, and left circularly polarized wave
  • the variation vector Vv is extended to a quaternion
  • a vertical polarization may be used instead of a horizontal polarization
  • a ⁇ 45 degree polarization may be used instead of a 45 degree polarization
  • right-handed circular polarization may be used instead of left-handed circular polarization.
  • the ground surface is classified into one of lake, grassland, forest, and town. It may be classified into at least four of forest, city, and desert, and the surface type may be classified as one of five among lake, grassland, forest, city, and desert. Also good.
  • the light (wave source) emitted from the artificial satellite 20 is a microwave, but the wave source is not limited to the microwave, and millimeter waves, terahertz waves, light waves, etc. are used as the wave source. It doesn't matter.
  • the ground type classification method in the embodiment, learning by the processing in steps S200 to S290 of the type relationship learning routine illustrated in FIG. 13 corresponds to “type relationship learning”.
  • the PolSAR image input module 52 corresponds to the “radio wave information input module”
  • the classification relation learning module 54 corresponds to the “class relation learning module”
  • the ground classification classification module 56 is “ Corresponds to the “Surface type classification module”.
  • the CPU 32 that executes the process of step S100 in FIG. 2 corresponds to the “radio wave information input unit”
  • the CPU 32 that corresponds to the “learning unit” and executes the process of step S110 in FIG. 2 corresponds to the “ground type classification unit”.
  • the present invention can be used for the manufacturing industry of the ground classification classification program and the ground classification classification apparatus.

Landscapes

  • Engineering & Computer Science (AREA)
  • Remote Sensing (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Physics & Mathematics (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • General Physics & Mathematics (AREA)
  • Electromagnetism (AREA)
  • Radar Systems Or Details Thereof (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

　Processing for inputting PolSAR image data from an artificial satellite is executed (step S100), a position vector (Pv) and variation vector (Vv) are set from the inputted PolSAR image data, and an observed region is classified into any one of the categories of lake, grass, forest, and town on the basis of a learned type function and an input vector (Xi) in which the set position vector (Pv) and variation vector (Vv) are expanded to a quaternion (step S110). A terrain type can thereby be classified with good precision.

Description

地表種別分類方法および地表種別分類プログラム並びに地表種別分類装置Ground surface classification method, ground surface classification program, and ground surface classification device
　本発明は、地表種別分類方法および地表種別分類プログラム並びに地表種別分類装置に関し、詳しくは、地表の複数の領域に対して照射した照射波とこの照射波の照射に伴って得られる散乱波とからなる電波情報を入力し、電波情報と地表種別との関係である種別関係を学習する種別関係学習に基づいて複数の領域の地表種別を分類する地表種別分類方法および地表種別分類プログラム並びに地表種別分類装置に関する。 The present invention relates to a ground surface classification method, a ground surface classification program, and a ground surface classification device, and more specifically, from an irradiation wave irradiated to a plurality of areas of the ground surface and a scattered wave obtained by irradiation of the irradiation wave. A ground type classification method, a ground type classification program, and a ground type classification that classify ground types of a plurality of regions based on type relationship learning that learns a type relationship that is a relationship between radio wave information and a ground type. Relates to the device.
　従来、この種の地表種別分類方法としては、人工衛星や航空機から地表に発射される電波の偏波情報とこの電波の発射に伴って得られる地表からの散乱波の偏波情報を用いて自然地形と人工地形とを判定するものが提案されている（例えば、特許文献１参照）。この方法では、地表の散乱を自然地形の散乱と都市部の散乱とにモデル化して得られた散乱行列の各変量の共分散行列（Covariance Matrix、以下「Ｃ行列」という）を用いて自然地形と人工地形との判定を行なうことにより、自然地形と人工地形とを分類している。 Conventionally, this kind of surface classification method is based on natural information using the polarization information of radio waves emitted from the satellite or aircraft to the ground surface and the polarization information of scattered waves from the ground surface obtained by the emission of these radio waves. A method for determining the terrain and the artificial terrain has been proposed (see, for example, Patent Document 1). This method uses the covariance matrix (Covariance Matrix, hereinafter referred to as “C matrix”) of each variability of the scattering matrix obtained by modeling the scattering of the ground surface into the scattering of the natural terrain and the scattering of the urban area. The natural terrain and the artificial terrain are classified by determining the terrain and the artificial terrain.
特開２００５－１４０６０７号公報JP 2005-140607 A
　しかしながら、上述の地表種別分類方法では、地表の種類毎の散乱をどのようにモデル化するかによって、分類の精度が大きく変わってしまう。精度良く分類するためには、地表の散乱を表す行列を尤もらしく分解する必要があるが、一般に、こうした分解を一意的に行なうことが困難であるため、精度良く地表の種別を分類することができない場合がある。 However, in the above-described ground type classification method, the accuracy of classification varies greatly depending on how the scattering for each type of ground is modeled. In order to classify accurately, it is necessary to reasonably decompose the matrix representing the scattering of the ground surface, but in general, it is difficult to uniquely perform such decomposition, so it is possible to classify the type of the ground surface with high accuracy. There are cases where it is not possible.
　本発明の地表種別分類方法および地表種別分類プログラム並びに地表種別分類装置は、地表種別を精度良く分類することを主目的とする。 The ground surface type classification method, the ground surface type classification program, and the ground surface type classification device of the present invention are mainly intended to classify ground surface types with high accuracy.
　本発明の地表種別分類方法および地表種別分類プログラム並びに地表種別分類装置は、上述の主目的を達成するために以下の手段を採った。 The ground surface type classification method, the ground surface type classification program, and the ground surface type classification device of the present invention employ the following means in order to achieve the main purpose described above.
　本発明の地表種別分類方法は、
　地表の複数の領域に対して照射した照射波と該照射波の照射に伴って得られる散乱波とからなる電波情報を入力し、前記電波情報と地表種別との関係である種別関係を学習する種別関係学習に基づいて前記複数の領域の地表種別を分類する地表種別分類方法において、
　前記種別関係学習は、複数の領域に対する偏波情報のポアンカレ球における位置ベクトルと、該位置ベクトルの近隣の領域の位置ベクトルからの変動を示す変動ベクトルと、を四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて前記種別関係を学習する
　ことを要旨とする。 The ground surface classification method of the present invention is:
Input radio wave information consisting of an irradiation wave irradiated to a plurality of areas on the ground surface and scattered waves obtained by the irradiation of the irradiation wave, and learn a type relationship that is a relationship between the radio wave information and the surface type. In the ground surface type classification method for classifying the ground surface types of the plurality of regions based on type relationship learning,
In the type relationship learning, a position vector of Poincare spheres of polarization information for a plurality of regions and a variation vector indicating a variation from a position vector of a neighboring region of the position vector are expanded to a plurality of four quaternions. The gist is to learn the type relationship by using a neural network with an elemental vector as an input value.
　この本発明の地表種別分類方法では、地表の複数の領域に対して照射した照射波と該照射波の照射に伴って得られる散乱波とからなる電波情報を入力し、電波情報と地表種別との関係である種別関係を学習する種別関係学習に基づいて複数の領域の地表種別を分類する。種別関係学習は、複数の領域に対する偏波情報のポアンカレ球における位置ベクトルと、位置ベクトルの近隣の位置ベクトルからの変動を示す変動ベクトルと、を四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて種別関係を学習する。位置ベクトルと変動ベクトルとは偏波情報を反映しているから、位置ベクトルと変動ベクトルとを用いて種別関係を学習することにより、適正に種別関係を学習することができる。また、位置ベクトルと変動ベクトルとの双方を用いて種別関係を学習するから、位置ベクトルおよび変動ベクトルのいずれか一方のみを用いるものと比較すると、より適正に種別関係を学習することができる。さらに、位置ベクトルと変動ベクトルとを四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて種別関係を学習するから、適正に種別関係を学習することができる。こうして適正に学習された種別関係に基づいて複数の領域の地表種別を分類することにより、地表種別を精度良く分類することができる。 In this ground surface type classification method of the present invention, radio wave information consisting of an irradiation wave irradiated to a plurality of areas of the ground surface and a scattered wave obtained along with the irradiation of the irradiation wave is input, and the radio wave information and the ground surface type are input. The ground surface types of a plurality of regions are classified based on the type relationship learning for learning the type relationship which is Type relationship learning is a method of expanding a quaternion vector obtained by extending a position vector in the Poincare sphere of polarization information for a plurality of regions and a variation vector indicating a variation of the position vector from a neighboring position vector to a quaternion. The type relationship is learned using a neural network as an input value. Since the position vector and the variation vector reflect the polarization information, the type relationship can be properly learned by learning the type relationship using the position vector and the variation vector. Further, since the type relationship is learned using both the position vector and the variation vector, the type relationship can be learned more appropriately as compared with the case where only one of the position vector and the variation vector is used. Furthermore, since the type relationship is learned using a neural network using a plurality of quaternion vectors obtained by expanding the position vector and the variation vector to a quaternion, the type relationship can be learned appropriately. By classifying the ground surface types of a plurality of regions based on the properly learned type relationship in this way, the ground surface types can be classified with high accuracy.
　こうした本発明の地表種別分類方法において、前記種別関係学習は、前記入力値として、前記位置ベクトルの各成分を３つの虚数部に対応させると共に第１の値を実数部に対応させることによって前記位置ベクトルを四元数に拡張した四元数位置ベクトルと、前記変動ベクトルの各成分を３つの虚数部に対応させると共に第２の値を実数部に対応させることによって前記変動ベクトルを四元数に拡張した四元数変動ベクトルと、を含むものとすることもできる。こうすれば、より適正に種別関係を学習することができる。ここで、「第１の値」としては、適宜定めることができ、特に値０とするのが好ましい。また、「第２の値」としては、適宜定めることができ、特に値０とするのが好ましい。 In such a ground surface type classification method of the present invention, the type relationship learning is performed by causing each component of the position vector to correspond to three imaginary parts and the first value to correspond to the real part as the input value. A quaternion position vector obtained by expanding a vector to a quaternion, and each component of the variation vector are made to correspond to three imaginary parts and a second value is made to correspond to a real part to make the variation vector a quaternion. And an extended quaternion variation vector. In this way, the type relationship can be learned more appropriately. Here, the “first value” can be determined as appropriate, and a value of 0 is particularly preferable. Further, the “second value” can be determined as appropriate, and the value of 0 is particularly preferable.
　さらに、本発明の地表種別分類方法において、前記種別関係学習は、前記位置ベクトルとして予め定めた地表種別が既知の領域に対する偏波情報のポアンカレ球における位置ベクトルを用いて行なわれる学習であり、前記種別関係学習における前記種別関係と前記入力された電波情報とに基づいて前記複数の領域の地表種別毎の関連性の程度を演算し、該演算した地表種別毎の関連性の程度に基づいて前記複数の領域の地表種別を分類するものとすることもできる。種別関係学習は、既知の領域に対する偏波情報のポアンカレ球における位置ベクトルを用いて行なわれるから、より適正に種別関係を学習することができる。こうして学習した種別関係を用いて地表種別を分類するから、より精度よく地表種別を分類することができる。この場合において、前記種別関係学習は、前記入力値として、前記照射波を水平偏波および垂直偏波のいずれか一方の偏波状態にしたときと、前記照射波を４５度偏波および－４５度偏波のいずれか一方の偏波状態にしたときと、前記照射波を左旋円偏波および右旋円偏波のいずれか一方の偏波状態にしたときと、のそれぞれにおける３つの前記位置ベクトルおよび３つの前記変動ベクトルを四元数に拡張した６つの四元数ベクトルを含むものとすることもできる。 Further, in the ground surface type classification method of the present invention, the type relationship learning is learning performed using a position vector in a Poincare sphere of polarization information for a region where a predetermined ground surface type is known as the position vector, Based on the type relationship in the type relationship learning and the input radio wave information, the degree of relevance for each surface type of the plurality of regions is calculated, and based on the calculated degree of relevance for each surface type The ground surface types of a plurality of areas may be classified. Since the type relationship learning is performed using the position vector in the Poincare sphere of polarization information for a known region, the type relationship can be learned more appropriately. Since the ground type is classified using the learned type relationship, the ground type can be classified with higher accuracy. In this case, the type relationship learning is performed when the irradiation wave is set to one of a horizontally polarized wave and a vertically polarized wave as the input value, and when the irradiation wave is a 45 degree polarized wave and −45 The three positions in each of the polarization state of one degree of polarization and the polarization state of the left-handed circular polarization and right-handed circular polarization of the irradiation wave It is also possible to include six quaternion vectors obtained by extending the vector and the three variation vectors to quaternions.
　位置ベクトルとして予め定めた地表種別が既知の領域に対する偏波情報のポアンカレ球における位置ベクトルを用いて種別関係学習を行なう態様の本発明の地表種別分類方法において、前記種別関係学習は、値０ではない第３の値を実数部に対応させると共に値０を３つの虚数部に対応させた四元数である閾値ベクトルを前記ニューラルネットワークの閾値として入力するものとすることもできる。「第３の値」としては、適宜定めることができ、特に値－１とするのが好ましい。 In the ground type classification method of the present invention in which the type relation learning is performed using the position vector in the Poincare sphere of the polarization information for the region whose surface type is predetermined as the position vector, the type relation learning is a value of 0. It is also possible to input a threshold vector, which is a quaternion in which the third value that does not correspond to the real part and the value 0 corresponds to the three imaginary parts, as the threshold value of the neural network. The “third value” can be determined as appropriate, and is particularly preferably a value of −1.
　位置ベクトルとして予め定めた地表種別が既知の領域に対する偏波情報のポアンカレ球における位置ベクトルを用いて種別関係学習を行なう態様の本発明の地表種別分類方法において、前記地域種別は、異なるｎ個の種別であり、前記ニューラルネットワークは、互いに内積が値０となる前記ｎ個のベクトルを前記入力値に対して期待される出力値とするものとすることもできる。こうすれば、より適正に地表種別を分類することができる。 In the ground type classification method of the present invention in which the type relation learning is performed using the position vector in the Poincare sphere of polarization information for a region whose ground type is predetermined as a position vector, the region type includes n different types of area types. It is also possible that the neural network has the n vectors whose inner products are 0 as the output values expected for the input values. In this way, it is possible to classify the surface type more appropriately.
　本発明の地表種別分類方法において、前記地表種別は、湖、草地、森、街、砂漠のうちの少なくとも４つの種別を含むものとすることもできる。こうすれば、地表の複数の領域を湖、草地、森、街、砂漠のうちの少なくとも４つの種別に分類することができる。 In the method for classifying ground surface types according to the present invention, the ground surface types may include at least four types of lake, grassland, forest, city, and desert. In this way, it is possible to classify a plurality of areas on the ground surface into at least four types of lakes, grasslands, forests, towns, and deserts.
　本発明の地表種別分類プログラムは、
　地表の複数の領域の地表種別を分類する地表種別分類プログラムであって、
　地表の複数の領域に対して照射した照射波と該照射波の照射に伴って得られる散乱波とからなる電波情報を入力する電波情報入力モジュールと、
　複数の領域に対する偏波情報のポアンカレ球における位置ベクトルと、該位置ベクトルの近隣の領域の位置ベクトルからの変動を示す変動ベクトルと、を四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて前記電波情報と地表種別との関係である種別関係を学習する種別関係学習モジュールと、
　前記学習した種別関係に基づいて前記複数の領域の地表種別を分類する地表種別分類モジュールと、
　を備えることを要旨とする。 The ground surface classification program of the present invention is
A ground type classification program for classifying ground types of a plurality of areas of the ground surface,
A radio wave information input module for inputting radio wave information composed of an irradiation wave irradiated to a plurality of regions of the ground surface and a scattered wave obtained by the irradiation of the irradiation wave;
Multiple quaternion vectors obtained by extending the position vector of the polarization information for multiple areas in the Poincare sphere and the variation vector indicating the variation from the position vector of the neighboring area of the position vector to a quaternion A type relationship learning module that learns a type relationship that is a relationship between the radio wave information and the surface type using a neural network,
A ground surface type classification module for classifying the ground surface types of the plurality of regions based on the learned type relationship;
It is a summary to provide.
　この本発明の地表種別分類プログラムでは、地表の複数の領域に対して照射した照射波と照射波の照射に伴って得られる散乱波とからなる電波情報を入力し、複数の領域に対する偏波情報のポアンカレ球における位置ベクトルと、位置ベクトルの近隣の領域の位置ベクトルからの変動を示す変動ベクトルと、を四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて種別関係を学習する。位置ベクトルおよび変動ベクトルは偏波情報を反映した値であるから、位置ベクトルおよび変動ベクトルを用いて種別関係を学習することにより、適正に種別関係を学習することができる。また、位置ベクトルと変動ベクトルとの双方を用いて種別関係を学習するから、位置ベクトルおよび変動ベクトルのいずれか一方のみを用いるものと比較すると、より適正に種別関係を学習することができる。さらに、位置ベクトルと変動ベクトルとを四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて種別関係を学習するから、より適正に種別関係を学習することができる。そして、学習した種別関係に基づいて複数の領域の地表種別を分類する。適正に学習された種別関係に基づいて複数の領域の地表種別を分類するから、地表種別を精度良く分類することができる。 In this ground classification classification program of the present invention, radio wave information consisting of irradiation waves irradiated to a plurality of areas of the ground surface and scattered waves obtained by irradiation of the irradiation waves is input, and polarization information for the plurality of areas is input. Type relationship using a neural network with multiple quaternion vectors obtained by expanding the position vector in the Poincare sphere of the quaternion and the variation vector indicating the variation of the position vector from the position vector of the neighboring area into a quaternion To learn. Since the position vector and the variation vector are values that reflect the polarization information, the type relationship can be properly learned by learning the type relationship using the position vector and the variation vector. Further, since the type relationship is learned using both the position vector and the variation vector, the type relationship can be learned more appropriately as compared with the case where only one of the position vector and the variation vector is used. Furthermore, since the type relationship is learned using a neural network using a plurality of quaternion vectors obtained by expanding the position vector and the variation vector into a quaternion, the type relationship can be learned more appropriately. Then, the ground surface types of the plurality of regions are classified based on the learned type relationship. Since the ground surface types of a plurality of regions are classified based on the properly learned type relationship, the ground surface types can be classified with high accuracy.
　本発明の地表種別分類装置は、
　地表の複数の領域の地表種別を分類する地表種別分類装置であって、
　地表の複数の領域に対して照射した照射波と該照射波の照射に伴って得られる散乱波とからなる電波情報を入力する電波情報入力部と、
　複数の領域に対する偏波情報のポアンカレ球における位置ベクトルと、該位置ベクトルの近隣の領域の位置ベクトルからの変動を示す変動ベクトルと、を四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて前記電波情報と地表種別との関係である種別関係を学習する種別関係学習部と、
　前記学習した種別関係に基づいて前記複数の領域の地表種別を分類する地表種別分類部と、
　を備えることを要旨とする。 The ground surface classification apparatus of the present invention is
A ground type classification device that classifies ground types of a plurality of areas of the ground surface,
A radio wave information input unit for inputting radio wave information composed of an irradiation wave irradiated to a plurality of areas on the ground surface and a scattered wave obtained by irradiation of the irradiation wave;
Multiple quaternion vectors obtained by extending the position vector of the polarization information for multiple areas in the Poincare sphere and the variation vector indicating the variation from the position vector of the neighboring area of the position vector to a quaternion A type relationship learning unit that learns a type relationship that is a relationship between the radio wave information and the surface type using a neural network,
A ground surface type classification unit for classifying the ground surface types of the plurality of regions based on the learned type relationship;
It is a summary to provide.
　この本発明の地表種別分類装置では、地表の複数の領域に対して照射した照射波と照射波の照射に伴って得られる散乱波とからなる電波情報を入力し、複数の領域に対する偏波情報のポアンカレ球における位置ベクトルと、位置ベクトルの近隣の領域の位置ベクトルからの変動を示す変動ベクトルと、を四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて種別関係を学習する。位置ベクトルは偏波情報そのものであるから、位置ベクトルを用いて種別関係を学習することにより、適正に種別関係を学習することができる。また、位置ベクトルと変動ベクトルとの双方を用いて種別関係を学習するから、位置ベクトルおよび変動ベクトルのいずれか一方のみを用いるものと比較すると、より適正に種別関係を学習することができる。さらに、位置ベクトルと変動ベクトルとを四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて種別関係を学習するから、より適正に種別関係を学習することができる。そして、学習した種別関係に基づいて複数の領域の地表種別を分類する。適正に学習された種別関係に基づいて複数の領域の地表種別を分類するから、地表種別を精度良く分類することができる。 In the ground surface classification apparatus according to the present invention, radio wave information including irradiation waves irradiated to a plurality of areas on the ground surface and scattered waves obtained by irradiation of the irradiation waves is input, and polarization information for the plurality of areas is input. Type relationship using a neural network with multiple quaternion vectors obtained by expanding the position vector in the Poincare sphere of the quaternion and the variation vector indicating the variation of the position vector from the position vector of the neighboring area into a quaternion To learn. Since the position vector is the polarization information itself, the type relationship can be properly learned by learning the type relationship using the position vector. Further, since the type relationship is learned using both the position vector and the variation vector, the type relationship can be learned more appropriately as compared with the case where only one of the position vector and the variation vector is used. Furthermore, since the type relationship is learned using a neural network using a plurality of quaternion vectors obtained by expanding the position vector and the variation vector into a quaternion, the type relationship can be learned more appropriately. Then, the ground surface types of the plurality of regions are classified based on the learned type relationship. Since the ground surface types of a plurality of regions are classified based on the properly learned type relationship, the ground surface types can be classified with high accuracy.
本発明の第１実施例としての地表種別分類装置３０を備える合成開口レーダシステム１０の構成の概略を示す構成図である。It is a block diagram which shows the outline of a structure of the synthetic aperture radar system 10 provided with the ground surface classification apparatus 30 as 1st Example of this invention. ＣＰＵ３２により実行される地表種別分類プログラム５０の一例を示すフローチャートである。It is a flowchart which shows an example of the surface classification classification program 50 performed by CPU32. 富士裾野地区の衛星写真と衛星写真を種別毎にスケッチした図面とを示す説明図である。It is explanatory drawing which shows the satellite photograph of Fuji Susono area, and the drawing which sketched the satellite photograph for every classification. 観測領域の１辺のピクセル数と平均偏波度との関係を示す。The relationship between the number of pixels on one side of the observation area and the average polarization degree is shown. 観測領域の１辺のピクセル数と変動ベクトルＶｖのノルムとの関係を示す。The relationship between the number of pixels on one side of the observation area and the norm of the variation vector Vv is shown. 照射波を水平偏波としたときの散乱波の位置ベクトルＰｖを示す説明図である。It is explanatory drawing which shows the position vector Pv of a scattered wave when an irradiation wave is made into a horizontal polarization. 照射波を水平偏波としたときの散乱波の変動ベクトルＶｖを示す説明図である。It is explanatory drawing which shows the fluctuation vector Vv of a scattered wave when an irradiation wave is made into a horizontal polarization. 照射波を４５度偏波としたときの散乱波の位置ベクトルＰｖを示す説明図である。It is explanatory drawing which shows the position vector Pv of a scattered wave when an irradiation wave is made into 45 degree | times polarization. 照射波を４５度偏波としたときの散乱波の変動ベクトルＶｖを示す説明図である。It is explanatory drawing which shows the fluctuation vector Vv of a scattered wave when an irradiation wave is made into 45 degree | times polarization. 照射波を左旋円偏波としたときの散乱波の位置ベクトルＰｖを示す説明図である。It is explanatory drawing which shows the position vector Pv of a scattered wave when an irradiation wave is left-handed circularly polarized wave. 照射波を左旋円偏波としたときの散乱波の変動ベクトルＶｖを示す説明図である。It is explanatory drawing which shows the fluctuation vector Vv of a scattered wave when an irradiation wave is left-handed circularly polarized wave. 種別関係学習に用いられる四元数のニューラルネットワークの構成を説明するための構成図である。It is a block diagram for demonstrating the structure of the quaternion neural network used for classification relationship learning. 種別関係学習ルーチンの一例を示すフローチャートである。It is a flowchart which shows an example of a classification relationship learning routine. 式（２４）～式（２８）に用いられる添え字等の説明を示す表である。10 is a table showing explanations of subscripts and the like used in Expression (24) to Expression (28). 各地表種別に対する最終出力Ｐｏｕｔ－ｔｙｐｅの一例を示す参考図である。FIG. 12 is a reference diagram illustrating an example of a final output Pout-type for each local table type. 自己組織化マップを偏波状態を表す特徴量空間で表現した模式図である。It is the schematic diagram which expressed the self-organization map in the feature-value space showing a polarization state. 自己組織化マップをニューロンの繋がり具合を表すＳＯＭ空間で表現した模式図である。It is the model which expressed the self-organization map in the SOM space showing the connection condition of a neuron. ＣＰＵ３２により実行される地表種別分類プログラム１５０の一例を示すフローチャートである。It is a flowchart which shows an example of the surface classification classification program 150 performed by CPU32.
　次に、本発明を実施するための形態を実施例を用いて説明する。 Next, modes for carrying out the present invention will be described using examples.
　図１は、本発明の第１実施例としての地表種別分類装置３０を備える合成開口レーダシステム１０の構成の概略を示す構成図である。合成開口レーダシステム１０は、人工衛星２０と、人工衛星２０から各種情報を入力し地表を湖、草地、森、街のいずれかの種別に分類する地表種別分類装置３０と、から構成されている。 FIG. 1 is a configuration diagram showing an outline of a configuration of a synthetic aperture radar system 10 including a ground surface classification device 30 as a first embodiment of the present invention. The synthetic aperture radar system 10 includes an artificial satellite 20 and a ground surface type classification device 30 that inputs various information from the artificial satellite 20 and classifies the ground surface into one of lake, grassland, forest, and city. .
　人工衛星２０は、地表の複数の領域に対してレーザ光を照射すると共に照射したレーザ光の散乱光を検出し、レーザ光の照射位置や照射したレーザ光の偏波情報，散乱光の偏波情報を含むＰｏｌＳＡＲ（Polarimetric Synthetic Apertture Radar）画像データなど各種データを地上の地表種別分類装置３０に出力する。 The artificial satellite 20 irradiates a plurality of regions on the surface with laser light, detects scattered light of the irradiated laser light, and detects the irradiation position of the laser light, the polarization information of the irradiated laser light, and the polarization of the scattered light. Various kinds of data such as PolSAR (Polarimetric Synthetic Apertture を Radar) image data including information are output to the ground surface classification device 30.
　地表種別分類装置３０は、図示するように、バス３１を介して接続された周知のＣＰＵ３２，ＲＯＭ３４，ＲＡＭ３６，ハードディスクドライブ（ＨＤＤ）３８，入出力処理回路４０，人工衛星２０からの情報を通信を介して入力する受信回路４２を備える。入出力処理回路４０には、受信回路４２やディスクドライブ装置４３，キーボード４４，マウス４５がデータを入力可能に接続され、情報を出力するディスプレイ４６がデータを入出力可能に接続されている。 As shown in the figure, the ground surface classification device 30 communicates information from a well-known CPU 32, ROM 34, RAM 36, hard disk drive (HDD) 38, input / output processing circuit 40, and satellite 20 connected via a bus 31. A receiving circuit 42 is provided. To the input / output processing circuit 40, a receiving circuit 42, a disk drive device 43, a keyboard 44, and a mouse 45 are connected so that data can be input, and a display 46 that outputs information is connected so that data can be input and output.
　ＨＤＤ３８には、アプリケーションソフトウェアとしてインストールされた地表種別分類プログラム５０が記憶されている。地表種別分類プログラム５０は、人工衛星２０からのＰｏｌＳＡＲ画像データを入力するＰｏｌＳＡＲ画像入力モジュール５２と、ＰｏｌＳＡＲ画像データと地表の種別との関係である種別関係を学習する種別関係学習モジュール５４と、種別関係学習モジュール５４が学習した種別関係に基づいて地表の各領域の地表種別を分類する地表種別分類モジュール５６と、から構成されている。 The HDD 38 stores a ground surface classification program 50 installed as application software. The ground surface type classification program 50 includes a PolSAR image input module 52 that inputs PolSAR image data from the artificial satellite 20, a type relationship learning module 54 that learns a type relationship that is a relationship between the PolSAR image data and the type of the ground surface, It is composed of a ground surface type classification module 56 that classifies the ground surface type of each region of the ground surface based on the type relationship learned by the relationship learning module 54.
　次に、こうして構成される地表種別分類装置３０の動作、特にユーザによるキーボード４４やマウス４５の操作により、地表種別分類プログラム５０が実行されたときの動作について説明する。図２は、ＣＰＵ３２により実行される地表種別分類プログラム５０の一例を示すフローチャートである。この地表種別分類プログラム５０は、ユーザにより実行が指示されたときに、ＣＰＵ３２により、ＲＡＭ３６の所定のアドレスに書き込まれる。そして、ＣＰＵ３２は、ＲＡＭ３６に書き込まれた地表種別分類プログラム５０を読み込んで実行する。 Next, the operation of the ground surface classification device 30 configured as described above, particularly the operation when the ground surface classification program 50 is executed by the operation of the keyboard 44 and the mouse 45 by the user will be described. FIG. 2 is a flowchart showing an example of the ground surface type classification program 50 executed by the CPU 32. The surface classification classification program 50 is written by the CPU 32 at a predetermined address in the RAM 36 when the execution is instructed by the user. Then, the CPU 32 reads and executes the ground surface type classification program 50 written in the RAM 36.
　地表種別分類プログラム５０が実行されると、ＣＰＵ３２は、まず、人工衛星２０からのＰｏｌＳＡＲ画像データを入力する処理を実行し（ステップＳ１００）、入力されたＰｏｌＳＡＲ画像データから後述の位置ベクトルＰｖ，変動ベクトルＶｖを設定し、設定した位置ベクトルＰｖ，変動ベクトルＶｖを四元数に拡張した入力ベクトルＸｉと予め学習されている種別関係とに基づいて、観測した領域を湖（ｌａｋｅ），草地（ｇｒａｓｓ），森（ｆｏｒｅｓｔ），街（ｔｏｗｎ）のうちのいずれか１つの種別に分類する（ステップＳ１１０）。ここで、地表種別分類プログラム５０の説明を中断して、位置ベクトルＰｖ，変動ベクトルＶｖ，四元数、種別関係の学習について説明する。 When the ground surface classification program 50 is executed, the CPU 32 first executes a process of inputting PolSAR image data from the artificial satellite 20 (step S100), and a position vector Pv, which will be described later, is changed from the input PolSAR image data. Based on the set position vector Pv, the input vector Xi obtained by expanding the set position vector Pv and the variation vector Vv to a quaternion, and the type relationship learned in advance, the observed area is defined as a lake and grass (grass) ), Forest, town (town) (step S110). Here, the explanation of the ground surface type classification program 50 is interrupted, and the learning of the position vector Pv, the fluctuation vector Vv, the quaternion, and the type relationship will be described.
　まずは、位置ベクトルＰｖおよび変動ベクトルＶｖについて説明する。一般に、照射波Ｉｗを式（１）で示す規格化ジョーンズベクトルを用いて表現すると、散乱波Ｒｗは、散乱行列Ｓを用いて式（２）により表現でき、ジョーンズコヒーレンス行列Ｊは、式（３）により得られる。式（３）中、＜＞は、時間的または空間的に平均化されていることを示している。 First, the position vector Pv and the variation vector Vv will be described. In general, when the irradiation wave Iw is expressed by using the normalized Jones vector represented by Expression (1), the scattered wave Rw can be expressed by Expression (2) using the scattering matrix S, and the Jones coherence matrix J is expressed by Expression (3). ). In the formula (3), <> indicates that they are averaged temporally or spatially.
Figure JPOXMLDOC01-appb-M000001
Figure JPOXMLDOC01-appb-M000001
Figure JPOXMLDOC01-appb-M000002
Figure JPOXMLDOC01-appb-M000002
Figure JPOXMLDOC01-appb-M000003
Figure JPOXMLDOC01-appb-M000003
　平均化されたストークスベクトルは、ジョーンズコヒーレンス行列Ｊに基づいて式（４）により計算される。平均化されたストークスベクトルは、地表からの散乱波の偏波情報を含んでおり、偏波度(Degree of Polarization）ＤｏＰは、式（５）を用いて計算される。ジョーンズコヒーレンス行列Ｊは、複素エルミート半正定値行列であるから、｜Ｊ｜は値０以上である。すなわち、偏波度ＤｏＰは、値０以上値１以下となっている。ここで、反射波は、偏波度Ｄｏｐが値１であるときには完全偏波となっており、偏波度ＤｏＰが値０のときには完全無偏波であり、偏波度Ｄｏｐが値０より大きく値１未満であるときには部分偏波となっている。 The averaged Stokes vector is calculated by Equation (4) based on the Jones coherence matrix J. The averaged Stokes vector includes polarization information of the scattered wave from the ground surface, and the degree of polarization (Degree of Polarization) DoP is calculated using Equation (5). Since the Jones coherence matrix J is a complex Hermitian semi-definite matrix, | J | is 0 or more. That is, the polarization degree DoP is a value of 0 or more and 1 or less. Here, the reflected wave is completely polarized when the degree of polarization Dop is 1, and is completely unpolarized when the degree of polarization DoP is 0, and the degree of polarization Dop is greater than 0. When the value is less than 1, it is partially polarized.
Figure JPOXMLDOC01-appb-M000004
Figure JPOXMLDOC01-appb-M000004
Figure JPOXMLDOC01-appb-M000005
Figure JPOXMLDOC01-appb-M000005
　平均化されたストークスベクトルは、ポアンカレ球上またはポアンカレ球内の１つの点として表現することができ、この点の位置座標（ｘ，ｙ，ｚ）は、式（６）により表すことができる。実施例では、位置座標（ｘ，ｙ，ｚ）を成分とする３次元のベクトルを、位置ベクトルＰｖ（＝（ｘ，ｙ，ｚ））とする。上述したように、位置ベクトルＰｖは、ＰｏｌＳＡＲ画像における各ピクセルについて得ることができ、偏波の平均的な状態を示している。位置ベクトルＰｖと位置ベクトルＰｖの近傍のピクセルの位置ベクトルＰｖとの変動は、平均偏差として式（７）により計算する。式（７）中、Ｎは１回の計算に用いられるピクセルの数であり、ｘ、ｙ、ｚは１回の計算に用いられる複数のピクセルにおけるｘの平均値である。式（７）におけるσｘ，σｙ，σｚを成分とする３次元のベクトルを、変動ベクトルＶｖ（＝（σｘ，σｙ，σｚ））とする。変動ベクトルＶｖは、ポアンカレ球における位置ベクトルＰｖの分布を示している。実施例では、こうした位置ベクトルＰｖおよび変動ベクトルＶｖを「ポアンカレ球パラメータ」と名付けるものとした。 The averaged Stokes vector can be expressed as one point on the Poincare sphere or in the Poincare sphere, and the position coordinates (x, y, z) of this point can be expressed by Equation (6). In the embodiment, a three-dimensional vector having position coordinates (x, y, z) as components is a position vector Pv (= (x, y, z)). As described above, the position vector Pv can be obtained for each pixel in the PolSAR image, and indicates an average state of polarization. The variation between the position vector Pv and the position vector Pv of the pixel in the vicinity of the position vector Pv is calculated by the equation (7) as an average deviation. In Expression (7), N is the number of pixels used for one calculation, and x, y, and z are average values of x in a plurality of pixels used for one calculation. A three-dimensional vector whose components are σx, σy, and σz in Expression (7) is a variation vector Vv (= (σx, σy, σz)). The variation vector Vv indicates the distribution of the position vector Pv in the Poincare sphere. In the embodiment, the position vector Pv and the variation vector Vv are named “Poincare sphere parameters”.
Figure JPOXMLDOC01-appb-M000006
Figure JPOXMLDOC01-appb-M000006
Figure JPOXMLDOC01-appb-M000007
Figure JPOXMLDOC01-appb-M000007
　ここで、富士裾野地区に対するＡＬＯＳのＬバンドＰＡＬＳＡＲ１．１レベルの実験データを用いてポアンカレ球パラメータと湖，草地，森，街の４つの種別との関係について説明する。図３は、富士裾野地区の衛星写真と衛星写真を種別毎にスケッチした図面とを示す説明図である。衛星写真において、４つの黒塗りの正方形は、湖、草地、森、街の各観測領域（以下、「グループ１」という）を示し、４つの白塗りの正方形は、湖、草地、森、街の各観測領域（以下、「グループ２」という）を示している。各正方形内には、４０×４０個の観測領域が含まれている。最初に、上述した式（３）によりジョーンズコヒーレンス行列Ｊを導出する。ここでは、空間平均値としてジョーンズコヒーレンス行列Ｊを導出するものとした。空間平均値を演算するために最適なウィンドウサイズ（１観測領域におけるピクセル数）を求めるため、グループ１，２のＰｏｌＳＡＲ画像データを用いてウィンドウサイズと平均偏波度との関係を調べた。図４に、観測領域の１辺のピクセル数と平均偏波度との関係を示す。図において、横軸が値５であるときにはウィンドウサイズは５×５であり、横軸が値９であるときにはウィンドウサイズは９×９とする。平均偏波度は、図示するように、ウィンドウサイズが５×５より大きくなると一定値になる。したがって、実施例では、位置ベクトルＰｖを求めるときのウィンドウサイズを５×５とした。同様に、上述の式（７）を用いて変動ベクトルＶｖの空間平均値を求めるため、グループ１，２のＰｏｌＳＡＲ画像データを用いてウィンドウサイズと変動ベクトルＶｖのノルムとの関係を調べた。図５は、観測領域の１辺のピクセル数と変動ベクトルＶｖのノルムとの関係を示す。変動ベクトルＶｖの大きさは、図示するように、ウィンドウサイズが９×９より大きくなると一定値となる。したがって、実施例では、変動ベクトルＶｖを求めるためのウィンドウサイズを９×９とした。 Here, the relationship between the Poincare sphere parameters and the four types of lakes, grasslands, forests, and towns will be described using the ALOS L-band PALSAR 1.1 level experimental data for the Fuji Susono area. FIG. 3 is an explanatory diagram showing a satellite photograph of the Fuji Susono area and a drawing sketched for each type of satellite photograph. In the satellite image, four black squares indicate the observation areas of the lake, grassland, forest, and city (hereinafter referred to as “Group 1”), and four white squares indicate the lake, grassland, forest, and city. Each observation area (hereinafter referred to as “Group 2”) is shown. Each square includes 40 × 40 observation areas. First, the Jones coherence matrix J is derived from the above-described equation (3). Here, the Jones coherence matrix J is derived as a spatial average value. In order to obtain the optimum window size (number of pixels in one observation region) for calculating the spatial average value, the relationship between the window size and the average polarization degree was examined using the PolSAR image data of groups 1 and 2. FIG. 4 shows the relationship between the number of pixels on one side of the observation area and the average polarization degree. In the figure, when the horizontal axis is the value 5, the window size is 5 × 5, and when the horizontal axis is the value 9, the window size is 9 × 9. As shown in the figure, the average polarization degree becomes a constant value when the window size is larger than 5 × 5. Therefore, in the embodiment, the window size for obtaining the position vector Pv is set to 5 × 5. Similarly, in order to obtain the spatial average value of the variation vector Vv using the above equation (7), the relationship between the window size and the norm of the variation vector Vv was examined using the PolSAR image data of groups 1 and 2. FIG. 5 shows the relationship between the number of pixels on one side of the observation region and the norm of the variation vector Vv. As shown in the drawing, the magnitude of the variation vector Vv becomes a constant value when the window size is larger than 9 × 9. Therefore, in the embodiment, the window size for obtaining the variation vector Vv is set to 9 × 9.
　こうしてウィンドウサイズを決定したら、続いて、水平偏波［１，０］T，４５度偏波[１／√２，１／√２]T，左旋円偏波[１／√２，１／√２ｉ]Tのマイクロ波をグループ１に属する観測領域に照射したときの位置ベクトルＰｖと変動ベクトルＶｖとを計算する。図６は照射波を水平偏波としたときの散乱波の位置ベクトルＰｖを示す説明図であり、図７は照射波を水平偏波としたときの散乱波の変動ベクトルＶｖを示す説明図であり、図８は照射波を４５度偏波としたときの散乱波の位置ベクトルＰｖを示す説明図であり、図９は照射波を４５度偏波としたときの散乱波の変動ベクトルＶｖを示す説明図であり、図１０は照射波を左旋円偏波としたときの散乱波の位置ベクトルＰｖを示す説明図であり、図１１は照射波を左旋円偏波としたときの散乱波の変動ベクトルＶｖを示す説明図である。位置ベクトルＰｖ，変動ベクトルＶｖは、図示するように、グラフ上では種別が異なると異なる分布を示し、且つ、位置ベクトルＰｖと変動ベクトルＶｖとで同じ分布をしていない。したがって、位置ベクトルＰｖ，変動ベクトルＶｖの両方を用いて地表の種別の分類を行なうことにより、位置ベクトルＰｖおよび変動ベクトルＶｖのいずれか一方を用いるものより、精度よく地表種別を分類できる。 After determining the window size in this way, subsequently, horizontal polarization [1, 0] T , 45 degree polarization [1 / √2, 1 / √2] T , counterclockwise circular polarization [1 / √2, 1 / √ 2i] A position vector Pv and a variation vector Vv when the observation region belonging to the group 1 is irradiated with T microwaves are calculated. FIG. 6 is an explanatory diagram showing the position vector Pv of the scattered wave when the irradiation wave is horizontally polarized, and FIG. 7 is an explanatory diagram showing the fluctuation vector Vv of the scattered wave when the irradiation wave is horizontal polarization. FIG. 8 is an explanatory diagram showing the position vector Pv of the scattered wave when the irradiation wave is 45-degree polarized wave, and FIG. 9 shows the fluctuation vector Vv of the scattered wave when the irradiation wave is 45-degree polarized wave. FIG. 10 is an explanatory diagram showing a position vector Pv of the scattered wave when the irradiation wave is left-handed circularly polarized wave, and FIG. 11 is a diagram of the scattered wave when the irradiation wave is left-handed circularly polarized wave. It is explanatory drawing which shows the fluctuation vector Vv. As shown in the drawing, the position vector Pv and the variation vector Vv show different distributions on the graph if the types are different, and the position vector Pv and the variation vector Vv do not have the same distribution. Therefore, by classifying the type of the ground surface using both the position vector Pv and the variation vector Vv, it is possible to classify the ground surface type with higher accuracy than that using either the position vector Pv or the variation vector Vv.
　続いて、四元数ベクトルについて説明する。四元数ｐは、１つの実数部と，互いに直交する基底ｉ，ｊ，ｋの３つの虚数部と，から構成される４次元の数であり、式（８）に示すように、４次元のベクトルとして表すことができる。式（８）中、ｐe，ｐi，ｐj,ｐkを実数とした。四元数の基底ｉ，ｊ，ｋは、式（９）に示すハミルトンの規則に従っている。 Next, the quaternion vector will be described. The quaternion p is a four-dimensional number composed of one real part and three imaginary parts of bases i, j, and k that are orthogonal to each other. Can be expressed as a vector of In Equation (8), p e , p i , p j, and p k are real numbers. The quaternion bases i, j, k follow the Hamilton rule shown in equation (9).
Figure JPOXMLDOC01-appb-M000008
Figure JPOXMLDOC01-appb-M000008
Figure JPOXMLDOC01-appb-M000009
Figure JPOXMLDOC01-appb-M000009
　２つの四元数ｐ（＝（ｐe，ｐi，ｐj,ｐk）），ｑ（＝（ｑe，ｑi，ｑj,ｑk））との和および差、外積，内積，ノルムは、式（１０）～（１３）に示すように定義される。 Two quaternions p (= (p e, p i, p j, p k)), the sum and the difference between q (= (q e, q i, q j, q k)), the outer product, inner product, The norm is defined as shown in equations (10) to (13).
Figure JPOXMLDOC01-appb-M000010
Figure JPOXMLDOC01-appb-M000010
Figure JPOXMLDOC01-appb-M000011
Figure JPOXMLDOC01-appb-M000011
Figure JPOXMLDOC01-appb-M000012
Figure JPOXMLDOC01-appb-M000012
Figure JPOXMLDOC01-appb-M000013
Figure JPOXMLDOC01-appb-M000013
　次に、種別関係学習について説明する。図１２は、種別関係学習に用いられる四元数のニューラルネットワークの構成を説明するための構成図である。種別関係学習は、ニューラルネットワークを用いて教師あり学習を行なって、ポアンカレ球パラメータと地表の種別との関係を学習する。実施例のニューラルネットワークでは、図示するように、隠れ層のニューロンの数を８つ，出力層のニューロンの数を４つとし、入力層、隠れ層の閾値ノードをそれぞれ（－１，０，０，０）とした。 Next, type relationship learning will be described. FIG. 12 is a block diagram for explaining the configuration of a quaternion neural network used for type relationship learning. In the type relationship learning, a supervised learning is performed using a neural network to learn the relationship between the Poincare sphere parameters and the type of the ground surface. In the neural network of the embodiment, as shown in the figure, the number of hidden layer neurons is 8, the number of output layer neurons is 4, and the threshold nodes of the input layer and hidden layer are (−1, 0, 0), respectively. , 0).
　図１３は、種別関係学習ルーチンの一例を示すフローチャートである。種別関係学習ルーチンでは、最初に、照射波を水平偏波、４５度偏波、左旋円偏波にしたときの湖、草地、森、街の種別が既知の４つの領域のＰｏｌＳＡＲ画像データから教師データを選択し、選択した教師データの位置ベクトルＰｖと変動ベクトルＶｖとを連ねたものを成分とする入力ベクトルＸinｉ－ｌａｋｅ，Ｘinｉ－ｇｒａｓｓ，Ｘinｉ－ｆｏｒｅｓｔ，Ｘinｉ－ｔｏｗｎ（以下、これらの４つのうちの１つを「Ｘinｉ－ｔｙｐｅ」と記載することもある）に設定する処理を実行する（ステップＳ２００）。ここで、ｉは、湖、草地、森、街の各種別のｉ番目の教師データであることを示している。入力ベクトルＸinｉ－ｔｙｐｅは、式（１４）～式（２０）に示すように、照射波を水平偏波、４５度偏波、左旋円偏波にしたときの位置ベクトルＰｖの各成分を３つの虚数部に対応させると共に値０を１つの実数部に対応させることによって位置ベクトルＰｖを四元数に拡張した四元数位置ベクトルｘH，ｘ45°，ｘlcと、変動ベクトルＶｖの各成分を３つの虚数部に対応させると共に値０を１つの実数部に対応させることによって変動ベクトルＶｖを四元数に拡張した四元数変動ベクトルσH，σ45°，σlcを連ねたものであるものとした。 FIG. 13 is a flowchart illustrating an example of the type relationship learning routine. In the type-related learning routine, first, the teacher learns from the PolSAR image data of four regions whose types of lake, grassland, forest, and town are known when the irradiation wave is horizontally polarized wave, 45 degree polarized wave, and left-handed circularly polarized wave. Input data X in i-lake, X in i-grass, X in i-forest, X in i-, which is obtained by selecting data and connecting the position vector Pv and variation vector Vv of the selected teacher data as components. The process of setting to “down” (hereinafter, one of these four may be described as “X in i-type”) is executed (step S200). Here, i indicates i-th teacher data for various types of lakes, grasslands, forests, and towns. As shown in the equations (14) to (20), the input vector X in i-type represents each component of the position vector Pv when the irradiation wave is a horizontally polarized wave, 45 degree polarized wave, and left-handed circularly polarized wave. A quaternion position vector x H , x 45 °, x lc obtained by extending the position vector Pv to a quaternion by corresponding to three imaginary parts and a value 0 corresponding to one real part, and a variation vector Vv The quaternion variation vectors σ H , σ 45 °, σ lc, in which the variation vector Vv is expanded to a quaternion by associating each component with three imaginary parts and the value 0 with one real part, are connected. It was supposed to be a thing.
Ｘinｉ－ｔｙｐｅ＝［ｘH，σH，ｘ45°，σ45°，ｘlc，σlc］・・・（１４）
ｘH＝（０，ｘ，ｙ，ｚ）H                                ・・・（１５）
σH＝（０，σｘ，σｙ，σｚ）H                          ・・・（１６）
ｘ45°＝（０，ｘ，ｙ，ｚ）45°                             ・・・（１７）
σ45°＝（０，σｘ，σｙ，σｚ）45°                       ・・・（１８）
ｘlc＝（０，ｘ，ｙ，ｚ）lc                               ・・・（１９）
σlc＝（０，σｘ，σｙ,σｚ）lc                         ・・・（２０） X in i-type = [x H , σ H , x 45 °, σ 45 °, x lc , σ lc ] (14)
x H = (0, x, y, z) H (15)
σ H = (0, σx, σy, σz) H (16)
x 45 ° = (0, x, y, z) 45 °                               ... (17)
σ 45 ° = (0, σx, σy, σz) 45 °                         ... (18)
x lc = (0, x, y, z) lc (19)
σ lc = (0, σx, σy, σz) lc (20)
　こうして入力ベクトルＸinｉ－ｔｙｐｅを設定したら、続いて、後述する結合荷重ｗを予め定めた初期値に設定すると共に各種別に期待される出力である期待出力ベクトルＤlake，Ｄgrass，Ｄforest,Ｄtownを設定する（ステップＳ２１０）。期待出力ベクトルＤlake，Ｄgrass，Ｄforest,Ｄtownは、期待出力ベクトルＤlake，Ｄgrass，Ｄforest,Ｄtownのうちのいずれか２つのベクトル同士が直交するよう式（２１）のように設定するものとした。このように、期待出力ベクトルＤlake，Ｄgrass，Ｄforest,Ｄtownを設定することにより、より適正に地表種別を分類することができる。 After the input vector X in i-type is set in this way, subsequently, a coupling load w described later is set to a predetermined initial value and expected output vectors D lake , D grass , D forest, D town is set (step S210). The expected output vectors D lake , D grass , D forest, and D town are expressed by equation (21) so that any two of the expected output vectors D lake , D grass , D forest, and D town are orthogonal to each other. It was supposed to be set. Thus, by setting the expected output vectors D lake , D grass , D forest, and D town , it is possible to classify the surface type more appropriately.
Figure JPOXMLDOC01-appb-M000014
Figure JPOXMLDOC01-appb-M000014
　続いて、湖、草地、森、街の各種別毎に後述するステップＳ２２０～Ｓ２５０の処理を実行して、ニューラルネットワークの結合荷重ｗｉを調整する処理を実行する。まず、ある種別（例えば、湖）の教師データのうちｉ番目のデータを入力ベクトルXｉとして式（２２）を用いて出力ベクトルYを演算する（ステップＳ２２０）。式（２２）中、ｗｉは、入力ベクトルXｉと出力ベクトルYとを結合させるためのニューラルネットワークにおける結合荷重である。なお、式（２２）中、ｓを初期状態とし、Ｎを入力ノードの数とし、非線形関数ｆ（ｓ）を式（２３）で示されるものとし、Ｘ0＝（－１，０，０，０）を閾値ノードの入力値とした。 Subsequently, the processing of steps S220 to S250, which will be described later, is executed for each type of lake, grassland, forest, and town, and processing for adjusting the coupling load wi of the neural network is executed. First, the output vector Y is calculated using equation (22) with the i-th data among the teacher data of a certain type (for example, lake) as the input vector Xi (step S220). In Expression (22), wi is a connection weight in the neural network for connecting the input vector Xi and the output vector Y. In equation (22), s is the initial state, N is the number of input nodes, and the nonlinear function f (s) is represented by equation (23), where X 0 = (− 1, 0, 0, 0) is the input value of the threshold node.
Figure JPOXMLDOC01-appb-M000015
Figure JPOXMLDOC01-appb-M000015
Figure JPOXMLDOC01-appb-M000016
Figure JPOXMLDOC01-appb-M000016
　こうして出力ベクトルYを計算したら、続いて、式（２４）を用いてエラー値Ｅを計算し、エラー値Ｅを閾値ｅｒｒ（例えば、５×１０-4）とを比較する（ステップＳ２３０）と共にステップＳ２２０が実行された回数を示す回数Ｎｆが１００回であるか否かを判定する（ステップＳ２４０）。ここで、回数Ｎｆは、初期値として値０が設定されており、ステップＳ２２０が実行される毎に値１ずつ大きくなる数とした。 When the output vector Y is calculated in this way, subsequently, an error value E is calculated using the equation (24), and the error value E is compared with a threshold err (for example, 5 × 10 −4 ) (step S230) and step It is determined whether or not the number Nf indicating the number of times S220 has been executed is 100 (step S240). Here, the number of times Nf is set to a value 0 as an initial value, and is a number that increases by 1 every time step S220 is executed.
Figure JPOXMLDOC01-appb-M000017
Figure JPOXMLDOC01-appb-M000017
　エラー値Ｅが閾値ｅｒｒ以上であり且つ回数Ｎｆが値１００でないときには、式（１７）～（２２）を用いて結合荷重ｗｉの値ｗoldを値ｗnewに更新して（ステップＳ２５０）、ステップＳ２２０に戻り、エラー値Ｅが閾値ｅｒｒより小さくなるか回数Ｎｆが値１００となるまで、ステップＳ２２０～Ｓ２５０の処理を繰り返す。式（２４）～式（２８）に用いられる添え字等の説明を図１４に示した。ステップＳ２２０～Ｓ２５０の処理により、種別が既知の入力ベクトルXｉｎに対して、出力ベクトルYが期待出力ベクトルＤlake，Ｄgrass，Ｄforest,Ｄtownに近くなるよう結合荷重ｗが調整される。 When the error value E is equal to or greater than the threshold err and the number of times Nf is not 100, the value w old of the combined load wi is updated to the value w new using the equations (17) to (22) (step S250). Returning to S220, the processes in steps S220 to S250 are repeated until the error value E becomes smaller than the threshold value err or the number of times Nf becomes 100. A description of subscripts and the like used in Expression (24) to Expression (28) is shown in FIG. By the processing in steps S220 to S250, the coupling load w is adjusted so that the output vector Y is close to the expected output vectors D lake , D grass , D forest, D town with respect to the input vector Xin whose type is known.
Figure JPOXMLDOC01-appb-M000018
Figure JPOXMLDOC01-appb-M000018
Figure JPOXMLDOC01-appb-M000019
Figure JPOXMLDOC01-appb-M000019
Figure JPOXMLDOC01-appb-M000020
Figure JPOXMLDOC01-appb-M000020
Figure JPOXMLDOC01-appb-M000021
Figure JPOXMLDOC01-appb-M000021
　エラー値Ｅが閾値ｅｒｒより小さくなるか回数Ｎｆが値１００となったときには、湖、草地、森、街の４つ種別に対してステップＳ２２０～Ｓ２５０の処理が終了したか否かを判定し（ステップＳ２６０）、４つ種別に対してステップＳ２２０～Ｓ２５０の処理が終了していないときには、種別を変更して（ステップＳ２７０）（例えば、種別を湖から草地に変更して）ステップＳ２２０～Ｓ２５０の処理を実行する。こうして湖、草地、森、街の４つ種別に対してステップＳ２２０～Ｓ２５０の処理が終了したときには、４つの種別について得られたエラー値Ｅの平均Ｓｅｒｒを式（２９）を用いて計算し、平均Ｓｅｒｒが閾値ｅｒｒ以上であるか否かを調べる（ステップＳ２８０）。式（２９）中、Ｅi-lake，Ｅi-grass，Ｅi-forest，Ｅi-townは、それぞれステップＳ２３０の判定に用いられる種別毎のエラー値Ｅとする。 When the error value E is smaller than the threshold value err or the number Nf is 100, it is determined whether or not the processing of steps S220 to S250 is completed for the four types of lake, grassland, forest, and town ( (Step S260) If the processing of steps S220 to S250 is not completed for the four types, the type is changed (step S270) (for example, the type is changed from lake to grassland) and steps S220 to S250 are performed. Execute the process. Thus, when the processing of steps S220 to S250 is completed for the four types of lake, grassland, forest, and town, the average Serr of the error values E obtained for the four types is calculated using the equation (29), It is checked whether or not the average Serr is equal to or greater than the threshold err (step S280). In equation (29), E i-lake , E i-grass , E i-forest , and E i-town are the error values E for each type used for the determination in step S230.
　Serr＝（Ei-lake+Ei-grass+Ei-forest+Ei-town)/4   ・・・（２９） Serr = (E i-lake + E i-grass + E i-forest + E i-town ) / 4 (29)
　平均Ｓｅｒｒが閾値ｅｒｒ以上であるときには（ステップＳ２８０）、対象としている教師データが最後（Ｎ番目）のデータでないとき（ｉ＜Ｎのとき）には、次の教師データについてステップＳ２２０～Ｓ２６０の処理を実行し（整数ｉに値１を加えて整数ｉを更新し）、対象としている教師データが最後（Ｎ番目）のデータであるとき（ｉ＝Ｎのとき）には、最初の教師データに対してステップＳ２２０～Ｓ２６０の処理を実行する（ステップＳ２９０）。こうして、平均Ｓｅｒｒが閾値ｅｒｒ未満になるまでステップＳ２２０～Ｓ２９０の処理を実行し、平均Ｓｅｒｒが閾値ｅｒｒ未満になったときに、本ルーチンを終了する。このように、種別関係学習処理では、教師データを入力ベクトルXｉとして出力ベクトルYが教師データの種別の期待出力ベクトルＤｔｙｐｅに近づくよう結合荷重ｗｉを調整する教師あり学習を行なうから、例えば、ポアンカレ球パラメータが湖の領域に対するものであるときには、式（２２）の出力ベクトルYは期待出力ベクトルＤlake近傍の値となり、ポアンカレ球パラメータが草地の領域に対するものであるときには、式（２２）の出力ベクトルYは期待出力ベクトルＤgrass近傍の値となる。このように、出力ベクトルYは、地表種別を反映した値となっている。このように、結合荷重ｗｉは、ポアンカレ球パラメータに基づく入力ベクトルXｉと地表種別を反映する出力ベクトルYとを関係づけるものであり、種別関係学習処理では、結合荷重ｗｉを学習する。したがって、種別関係学習処理は、電波情報と地表種別との関係である種別関係を学習する処理となっている。位置ベクトルＰｖ，変動ベクトルＶｖは、ＰｏｌＳＡＲ画像データから直接得られた偏波情報を反映しているから、位置ベクトルＰｖ，変動ベクトルＶｖを用いて種別関係を学習することにより、より適正に種別関係を学習することができる。また、位置ベクトルＰｖ，変動ベクトルＶｖの両方を用いるから、より精度の良い学習ができる。さらに、位置ベクトルＰｖ，変動ベクトルＶｖとを連ねたものを成分とする入力ベクトルＸinｉ－ｔｙｐｅを入力値としてニューラルネットワークを用いて種別関係を学習するから、より適正に種別関係について学習することができる。 When the average Serr is greater than or equal to the threshold err (step S280), when the target teacher data is not the last (Nth) data (when i <N), the processing of steps S220 to S260 is performed for the next teacher data (The value 1 is added to the integer i and the integer i is updated). When the target teacher data is the last (Nth) data (when i = N), the first teacher data is On the other hand, the processing of steps S220 to S260 is executed (step S290). Thus, the processing of steps S220 to S290 is executed until the average Serr is less than the threshold err, and when the average Serr is less than the threshold err, this routine is terminated. In this way, in the type relation learning process, supervised learning is performed in which the joint data wi is adjusted so that the output vector Y approaches the expected output vector Dtype of the type of the teacher data with the teacher data as the input vector Xi. When the parameter is for the lake region, the output vector Y of equation (22) is a value near the expected output vector D lake , and when the Poincare sphere parameter is for the grassland region, the output vector of equation (22). Y is a value near the expected output vector D grass . Thus, the output vector Y is a value reflecting the ground surface type. As described above, the connection load wi associates the input vector Xi based on the Poincare sphere parameter with the output vector Y reflecting the ground surface type, and the connection load wi is learned in the type relationship learning process. Therefore, the type relationship learning process is a process of learning a type relationship that is a relationship between the radio wave information and the ground surface type. Since the position vector Pv and the variation vector Vv reflect the polarization information directly obtained from the PolSAR image data, the type relationship is more appropriately learned by learning the type relationship using the position vector Pv and the variation vector Vv. Can learn. Further, since both the position vector Pv and the variation vector Vv are used, more accurate learning can be performed. Furthermore, since the type relationship is learned using a neural network with the input vector X in i-type having the component of the position vector Pv and the variation vector Vv as input values, the type relationship is more appropriately learned. Can do.
　ここで、図２に例示した地表種別分類プログラム５０の説明に戻る。人工衛星２０からの種別が未知のＰｏｌＳＡＲ画像データを入力したら（ステップＳ１００）、照射波を水平偏波、４５度偏波、左旋円偏波にしたときのＰｏｌＳＡＲ画像データの位置ベクトルＰｖと変動ベクトルＶｖとを連ねたものを成分とする入力ベクトルＸｉと図１３に示した種別関係学習ルーチンにより学習された種別関係とに基づいてレーザが照射された領域を湖、草地、森、街のいずれかの種別に分類する（ステップＳ１１０）。種別の分類は、具体的には、ＰｏｌＳＡＲ画像データから得られる入力ベクトルＸｉ（式（１４）のＸinｉ－ｔｙｐｅと成分は同じ）と図１３に示した学習により得られた結合荷重ｗｉを用いて上述した式（２２）により出力ベクトルYを演算し、出力ベクトルYと期待出力Ｄlake，Ｄgrass，Ｄforest,Ｄtownとを用いて式（２４）より各地表種別におけるエラー値Ｅｔｙｐｅ（ｔｙｐｅは、ｌａｋｅ，ｇｒａｓｓ，ｆｏｒｅｓｔ，ｔｏｗｎ）を計算し、エラー値Ｅｔｙｐｅを用いて式（３０）により最終出力Ｐｏｕｔ－ｔｙｐｅを計算する。そして、最終出力Ｐｏｕｔ－ｔｙｐｅが値１に近ければ、そのピクセルは値１に近い種別であるものとし、値０に近ければその種別には分類されないものとする。例えば、最終出力Ｐｏｕｔ－ｌａｋｅが値１に近く、且つ、最終出力Ｐｏｕｔ－ｇｒａｓｓ，Ｐｏｕｔ－ｆｏｒｅｓｔ，Ｐｏｕｔ－ｔｏｗｎが値０に近ければ、そのピクセルは「湖」に分類し、対応する観測領域を「湖」に分類する。 Here, the description returns to the ground surface classification program 50 illustrated in FIG. When PolSAR image data whose type is unknown from the artificial satellite 20 is input (step S100), the position vector Pv and variation vector of the PolSAR image data when the irradiation wave is horizontally polarized wave, 45 degree polarized wave, and left-handed circularly polarized wave. The region irradiated with the laser based on the input vector Xi having Vv as a component and the type relationship learned by the type relationship learning routine shown in FIG. 13 is one of lake, grassland, forest, and town. (Step S110). Specifically, the classification of the type is based on the input vector Xi obtained from the PolSAR image data (the component is the same as X in i-type in equation (14)) and the connection weight wi obtained by learning shown in FIG. Is used to calculate the output vector Y according to the above equation (22), and using the output vector Y and the expected outputs D lake , D grass , D forest, D town , the error value Etype ( (type, lake, glass, forest, down) is calculated, and the final output Pout-type is calculated by the equation (30) using the error value Etype. If the final output Pout-type is close to the value 1, it is assumed that the pixel is of a type close to the value 1, and if it is close to the value 0, it is not classified as that type. For example, if the final output Pout-lake is close to the value 1 and the final outputs Pout-glass, Pout-forest, Pout-town are close to the value 0, the pixel is classified as “lake” and the corresponding observation area is Classify as “Lake”.
　図１５は、図３に例示した富士裾野地区について実施例の方法を用いて地表種別を分類した結果である最終出力Ｐｏｕｔ－ｌａｋｅ，Ｐｏｕｔ－ｇｒａｓｓ，Ｐｏｕｔ－ｆｏｒｅｓｔ，Ｐｏｕｔ－ｔｏｗｎの値を示す説明図である。図１５において、左上は各領域における最終出力Ｐｏｕｔ－ｌａｋｅの値、右上は各領域における最終出力Ｐｏｕｔ－ｇｒａｓｓの値、左下は各領域における最終出力Ｐｏｕｔ－ｆｏｒｅｓｔの値、右下は各領域におけるＰｏｕｔ－ｔｏｗｎの値を示しており、各値は、値１に近づくほど色が濃くなるよう濃淡で表現されている。図１５と図３とから明らかなように、実施例の地表種別分類方法では、良好に地表種別を分類することができる。このように、位置ベクトルＰｖと変動ベクトルＶｖと、を四元数に拡張した複数の四元数ベクトルを入力ベクトルXｉとしてニューラルネットワークを用いて種別関係を学習し、学習結果を用いて地表種別を分類することにより、より精度良く地表種別を分類することができる。 FIG. 15 is a diagram illustrating the values of the final outputs Pout-lake, Pout-glass, Pout-forest, and Pout-town, which are the results of classifying the surface type using the method of the embodiment for the Fuji Susono area illustrated in FIG. FIG. In FIG. 15, the upper left is the value of the final output Pout-lake in each region, the upper right is the value of the final output Pout-glass in each region, the lower left is the value of the final output Pout-forest in each region, and the lower right is the Pout in each region. -Town values are shown, and each value is expressed by shading so that the color becomes darker as the value 1 is approached. As is clear from FIG. 15 and FIG. 3, the ground surface type classification method of the embodiment can favorably classify the ground surface type. In this way, a type relationship is learned using a neural network using a plurality of quaternion vectors obtained by extending the position vector Pv and the variation vector Vv to a quaternion as an input vector Xi, and the ground surface type is determined using the learning result. By classifying, the ground surface type can be classified with higher accuracy.
Figure JPOXMLDOC01-appb-M000022
Figure JPOXMLDOC01-appb-M000022
　以上説明した第１実施例の地表種別分類装置３０では、位置ベクトルＰｖと変動ベクトルＶｖと、を四元数に拡張した複数の四元数ベクトルを入力ベクトルＸｉとしてニューラルネットワークを用いて種別関係を学習することにより、より適正に種別関係を学習することができる。これにより、より精度良く地表種別を分類することができる。 In the ground surface type classification apparatus 30 of the first embodiment described above, a type relationship is established using a neural network with a plurality of quaternion vectors obtained by extending the position vector Pv and the variation vector Vv to a quaternion as an input vector Xi. By learning, the type relationship can be learned more appropriately. Thereby, it is possible to classify the ground surface type with higher accuracy.
　第１実施例の地表種別分類装置３０では、地表を４つの種別に分類するものとし、互いに内積が値０となる期待出力Ｄlake，Ｄgrass，Ｄforest,Ｄtownを期待される出力値としたが、地表を５つの種別に分類するものとし、互いに内積が値０となる５つの期待出力ベクトルＤlake，Ｄgrass，Ｄforest,Ｄtownを期待される出力値としてもよく、期待出力ベクトルを互いに内積が値０とならないものとしてもよい。 In the ground surface type classification device 30 of the first embodiment, the ground surface is classified into four types, and expected outputs D lake , D grass , D forest, and D town whose inner products are 0 values are expected output values. However, the ground surface is classified into five types, and five expected output vectors D lake , D grass , D forest, and D town whose inner products have a value of 0 may be used as expected output values. The inner products may not have a value of 0.
　第１実施例の地表種別分類装置３０では、図１２に例示したニューラルネットワークの閾値ノードを（－１，０，０，０）としたが、閾値ノードは演算処理の負荷の程度を考慮して、適宜定めることができる。 In the ground surface classification apparatus 30 of the first embodiment, the threshold node of the neural network illustrated in FIG. 12 is set to (−1, 0, 0, 0), but the threshold node takes into consideration the degree of load of the arithmetic processing. Can be determined as appropriate.
　次に、本発明の第２実施例としての地表種別分類装置１３０について説明する。第２実施例の地表種別分類装置１３０は、第１実施例の地表種別分類プログラム５０がニューラルネットワークを用いて教師あり学習を行なうものであるのに対して地表種別分類プログラム１５０がニューラルネットワークを用いて教師なし学習を行なうものである点を除いて、第１実施例の地表種別分類装置３０と同一の構成となっている。したがって、地表種別分類装置１３０において、地表種別分類装置３０と同一の構成には同一の符号を付し、その説明を省略する。 Next, the ground surface type classification apparatus 130 as a second embodiment of the present invention will be described. In the ground type classification device 130 of the second embodiment, the ground type classification program 50 of the first embodiment performs supervised learning using a neural network, whereas the ground type classification program 150 uses a neural network. The configuration is the same as that of the ground surface type classification apparatus 30 of the first embodiment except that unsupervised learning is performed. Therefore, in the ground surface classification device 130, the same components as those of the ground surface classification device 30 are denoted by the same reference numerals, and the description thereof is omitted.
　第２実施例の地表種別分類プログラム１５０は、位置ベクトルＰｖと変動ベクトルＶｖとを四元数に拡張した複数の四元数ベクトルを入力ベクトルＸｉ，Ｘinｉ－ｔｙｐｅとしてニューラルネットワークとして自己組織化マップ（self-organizig map:SOM）を用いて教師なし学習を行なってピクセルの種別を分類する。図１６は、自己組織化マップを偏波状態を表す特徴量空間で表現した模式図であり、図１７は、自己組織化マップをニューロンの繋がり具合を表すＳＯＭ空間で表現した模式図である。 The ground type classification program 150 of the second embodiment self-organizes a plurality of quaternion vectors obtained by extending the position vector Pv and the variation vector Vv into quaternions as input vectors Xi, X in i-type as a neural network. Perform unsupervised learning using a map (self-organizig map: SOM) to classify pixel types. FIG. 16 is a schematic diagram expressing the self-organizing map in a feature amount space representing a polarization state, and FIG. 17 is a schematic diagram expressing the self-organizing map in an SOM space representing the connection state of neurons.
　図１８は、ＣＰＵ３２により実行される地表種別分類プログラム１５０の一例を示すフローチャートである。地表種別分類プログラム１５０が実行されると、最初に、観測領域のＰｏｌＳＡＲ画像データを入力する処理を実行し（ステップＳ３００）、湖、草地、森、街の各種別の結合荷重ｗｌａｋｅ，ｗｇｒａｓｓ，ｗｆｏｒｅｓｔ，ｗｔｏｗｎに予め定めた初期値を設定すると共に対象とするポアンカレ球においてｉ番目のピクセルであることを示す整数ｉに値１を設定する初期化処理を実行し（ステップＳ３１０）、ＰｏｌＳＡＲ画像データから作成した入力ベクトルＸｉを入力する（ステップＳ３２０）。入力ベクトルＸｉは、式（３１）に示すように、水平偏波，４５度偏波、左旋円偏波の各偏波状態の位置ベクトルＰｖH，Ｐｖ45°，Ｐｖlcの各成分を３つの虚数部に対応させると共に値０を１つの実数部に対応させることによって位置ベクトルＰｖH，Ｐｖ45°，Ｐｖlcを四元数に拡張した四元数位置ベクトルｘH，ｘ45°，ｘlcと、水平偏波，４５度偏波、左旋円偏波の各偏波状態の変動ベクトルＶｖH，Ｖｖ45°，Ｖｖlcの各成分を３つの虚数部に対応させると共に値０を１つの実数部に対応させることによって変動ベクトルＶｖH，Ｖｖ45°，Ｖｖlcを四元数に拡張した四元数変動ベクトルσH，σ45°，σlcを連ねたものであるものとした。今、整数ｉは値１に設定されているから、ステップＳ３２０の処理では、１番目のピクセルのＰｏｌＳＡＲ画像データから作成した入力ベクトルＸｉが入力されることになる。 FIG. 18 is a flowchart showing an example of the ground surface type classification program 150 executed by the CPU 32. When the ground classification classification program 150 is executed, first, a process of inputting the PolSAR image data of the observation area is executed (step S300), and the combined loads wlake, wgrass, wforest for various types of lakes, grasslands, forests, and towns are executed. , Wtown is set to a predetermined initial value, and an initialization process is performed to set a value 1 to an integer i indicating the i-th pixel in the target Poincare sphere (step S310). The created input vector Xi is input (step S320). As shown in Expression (31), the input vector Xi includes three components of position vectors Pv H , Pv 45 °, and Pv lc in each polarization state of horizontal polarization, 45 degree polarization, and left-hand circular polarization. A quaternion position vector x H , x 45 °, x lc is obtained by expanding the position vector Pv H , Pv 45 °, Pv lc to a quaternion by making the value 0 correspond to one real part, corresponding to the imaginary part. And each component of the fluctuation vectors Vv H , Vv 45 °, and Vv lc of each polarization state of horizontal polarization, 45 degree polarization, and left-hand circular polarization correspond to three imaginary parts and a value 0 is one real number. It is assumed that the quaternion variation vectors σ H , σ 45 °, and σ lc obtained by extending the variation vectors Vv H , Vv 45 °, and Vv lc to quaternions are made to correspond to each other. Since the integer i is now set to 1, the input vector Xi created from the PolSAR image data of the first pixel is input in the process of step S320.
X ≡ [ｘH　σH　ｘ45° σ45°　ｘlc σlc ]T 　　・・・（３１） X ≡ [x H σ H x 45 °   σ 45 ° x lc σ lc ] T (31)
　こうして入力ベクトルＸｉを入力したら、続いて、式（３２）に示す入力ベクトルＸｉと各結合荷重ｗｌａｋｅ，ｗｇｒａｓｓ，ｗｆｏｒｅｓｔ，ｗｔｏｗｎとの距離（Ｘｉ，ｗｃ）を演算し、距離（Ｘｉ，ｗｃ）が最も小さい（入力ベクトルＸｉに最も距離が近い）結合荷重ｗｌａｋｅ，ｗｇｒａｓｓ，ｗｆｏｒｅｓｔ，ｗｔｏｗｎを持つニューロンを勝者ニューロンであると決定し、ｉ番目のピクセルを勝者クラスｃｗに分類する（ステップＳ３３０）。例えば、結合荷重ｗｌａｋｅが入力ベクトルＸｉに最も近いときには、結合荷重ｗｌａｋｅを持つニューロンが勝者ニューロンと決定され、ｉ番目のピクセルは湖（ｌａｋｅ）に分類される。なお、式（３２）中、「ｃ」は、ｌａｋｅ，ｇｒａｓｓ，ｆｏｒｅｓｔ，ｔｏｗｎのうちのいずれか１つの種別とした。 When the input vector Xi is input in this manner, the distance (Xi, wc) between the input vector Xi shown in the equation (32) and each of the coupling loads wlake, wgrass, wforest, wtown is calculated, and the distance (Xi, wc) is calculated. The neuron having the smallest connection weights wlake, wgrass, wforest, wtown (closest to the input vector Xi) is determined as the winner neuron, and the i-th pixel is classified into the winner class cw (step S330). For example, when the connection weight wlake is closest to the input vector Xi, the neuron having the connection weight wlake is determined as the winner neuron, and the i-th pixel is classified as a lake. In the equation (32), “c” is any one of rake, glass, forest, and town.
　距離（Ｘｉ，ｗｃ）≡ ｜｜Ｘｉ－ｗｃ｜｜  ・・・（３２） Distance (Xi, wc) = || Xi-wc ||    ... (32)
　続いて、式（３３）～式（３５）を用いて勝者クラスｃｗのニューロンｗｃｗとニューロンｗｃの近傍の２つのニューロン（ＳＯＭ空間で隣のニューロン）ｗｃｗ±１の荷重を更新する（ステップＳ３４０）。そして、全ての観測ピクセルに対してステップＳ３２０～Ｓ３４０の処理が終了したか否かを調べ（ステップＳ３５０）、全てのピクセルに対してステップＳ３２０～Ｓ３４０の処理が終了していないときには、整数ｉを整数ｉに値１を加えたものに更新して（ステップＳ３６０）、ステップＳ３１０に戻る。すなわち、全てのピクセルに対してステップＳ３２０～Ｓ３４０の処理が終了していないときには、次のピクセルに対してステップＳ３２０～Ｓ３４０の処理を行なうのである。今、１番目のピクセルについて考えているから、１番目のピクセルに対してステップＳ３２０～Ｓ３４０の処理が終了したときには、２番目のピクセルに対してステップＳ３２０～Ｓ３４０の処理を実行する。 Subsequently, the weights of the neuron wcw of the winner class cw and the two neurons in the vicinity of the neuron wc (neighboring neurons in the SOM space) wcw ± 1 are updated using the equations (33) to (35) (step S340). . Then, it is checked whether or not the processing of steps S320 to S340 has been completed for all the observation pixels (step S350). If the processing of steps S320 to S340 has not been completed for all the pixels, the integer i is set. The integer i is updated to the value obtained by adding 1 (step S360), and the process returns to step S310. That is, when the processing of steps S320 to S340 has not been completed for all pixels, the processing of steps S320 to S340 is performed for the next pixel. Now, since the first pixel is considered, when the processing of steps S320 to S340 is completed for the first pixel, the processing of steps S320 to S340 is executed for the second pixel.
　ｗｃｗ＝ｗｃｗ＋α（Ｘｉ－ｗｃｗ）　      ・・・（３３）
　ｗｃｗ±１＝ｗｃｗ±１＋β（Ｘｉ－ｗｃｗ）・・・（３４）
　０≦β≦α≦1                             ・・・（３５） wcw = wcw + α (Xi−wcw) (33)
wcw ± 1 = wcw ± 1 + β (Xi−wcw) (34)
0 ≦ β ≦ α ≦ 1 (35)
　こうした全てのピクセルに対しＳ３２０～Ｓ３４０の処理を実行したら（ステップＳ３５０）、全てのピクセルに対してＳ３２０～Ｓ３５０の処理をＮｉｔｒ回（１回以上の回数。例えば、５０回，１００回，１５０回など）繰り返したか否かを判定する（ステップＳ３７０）。全てのピクセルに対してＳ３２０～Ｓ３５０の処理をＮｉｔｒ回繰り返していないときには、整数ｉに値１を設定して（ステップＳ３８０）、最初のピクセルからステップＳ３２０～Ｓ３５０の処理を繰り返し、全てのピクセルに対してＳ３２０～Ｓ３５０の処理をＮｉｔｒ回繰り返したときには（ステップＳ３７０）、本ルーチンを終了する。こうした処理により、ＰｏｌＳＡＲ画像データを湖，草地，森，街のうちのいずれかの種別に分類することができ、即ち、地表の領域をいずれかの種別に分類するから、地表の散乱の種類を尤もらしく分解するものと比較すると、より精度良く分類することができる。 When the processing of S320 to S340 is executed for all such pixels (step S350), the processing of S320 to S350 is performed for all the pixels Nitr times (one or more times. For example, 50 times, 100 times, 150 times) It is determined whether or not the process has been repeated (step S370). When the processing of S320 to S350 has not been repeated for all pixels for Nitr times, the integer i is set to a value 1 (step S380), and the processing of steps S320 to S350 is repeated from the first pixel to all pixels. On the other hand, when the processes of S320 to S350 are repeated Nitr times (step S370), this routine is terminated. By such processing, PolSAR image data can be classified into one of the types of lake, grassland, forest, and city, that is, the surface area is classified into any type. Compared with those that are reasonably decomposed, it is possible to classify with higher accuracy.
　以上説明した第２実施例の地表種別分類装置１３０によれば、ＰｏｌＳＡＲ画像データから得られる位置ベクトルＰｖと変動ベクトルＶｖとを四元数に拡張した入力ベクトルＸｉを入力値としてニューラルネットワークを用いて結合荷重ｗｃを学習することにより、より精度良く地表種別を分類することができる。 According to the ground surface type classification device 130 of the second embodiment described above, a neural network is used with an input vector Xi obtained by expanding a position vector Pv and a variation vector Vv obtained from PolSAR image data into a quaternion as an input value. By learning the combined load wc, it is possible to classify the ground surface type with higher accuracy.
　第２実施例の地表種別分類装置１３０では、ステップＳ３３０の処理で、入力ベクトルＸｉに最も距離が近い結合荷重ｗｌａｋｅ，ｗｇｒａｓｓ，ｗｆｏｒｅｓｔ，ｗｔｏｗｎを持つニューロンを勝者ニューロンであると決定したが、式（３６）または式（３７）に示す入力ベクトルＸｉと各結合荷重ｗｌａｋｅ，ｗｇｒａｓｓ，ｗｆｏｒｅｓｔ，ｗｔｏｗｎとの類似度（Ｘｉ，ｗｃ）を演算し、最も類似度（Ｘｉ，ｗｃ）が大きい結合荷重ｗｌａｋｅ，ｗｇｒａｓｓ，ｗｆｏｒｅｓｔ，ｗｔｏｗｎを持つニューロンを勝者ニューロンであると決定するものとしてもよい。 In the ground surface type classification device 130 of the second embodiment, in the process of step S330, it is determined that the neuron having the connection weights wlake, wglass, wforest, wtown that is the closest to the input vector Xi is the winner neuron. 36) or the similarity (Xi, wc) between the input vector Xi shown in the equation (37) and each of the coupling weights wlake, wgrass, wforest, wtown, and the coupling weight wlake, with the largest similarity (Xi, wc). A neuron having wgrass, wforest, wtown may be determined as a winner neuron.
Figure JPOXMLDOC01-appb-M000023
Figure JPOXMLDOC01-appb-M000023
Figure JPOXMLDOC01-appb-M000024
Figure JPOXMLDOC01-appb-M000024
　第１，第２実施例の地表種別分類装置３０，１３０では、入力ベクトルＸｉ，Ｘinｉ－ｔｙｐｅにおいて、位置ベクトルＰｖの各成分を３つの虚数部に対応させると共に値０を１つの実数部に対応させることによって位置ベクトルＰｖを四元数に拡張するものとしたが、位置ベクトルＰｖの各成分を２つの虚数部と１つの実数部に対応させて値０を残余の虚数部に対応させてもよいし、実数部には値０と異なる値を対応させてもよい。 In the ground surface classification devices 30 and 130 of the first and second embodiments, each component of the position vector Pv corresponds to three imaginary parts in the input vectors Xi and X in i-type, and the value 0 is one real part. The position vector Pv is expanded to a quaternion by making it correspond to, but each component of the position vector Pv is made to correspond to two imaginary parts and one real part, and the value 0 is made to correspond to the remaining imaginary part. Alternatively, a value different from the value 0 may be associated with the real part.
　第１，第２実施例の地表種別分類装置３０，１３０では、入力ベクトルＸinｉ－ｔｙｐｅにおいて、変動ベクトルＶｖの各成分を３つの虚数部に対応させると共に値０を１つの実数部に対応させることによって変動ベクトルＶｖを四元数に拡張するものとしたが、変動ベクトルＶｖの各成分を２つの虚数部と１つの実数部に対応させて値０を残余の虚数部に対応させてもよいし、実数部には値０と異なる値を対応させてもよい。 In the ground surface classification devices 30 and 130 of the first and second embodiments, each component of the variation vector Vv corresponds to three imaginary parts and a value 0 corresponds to one real part in the input vector X in i-type. The variation vector Vv is expanded to a quaternion by performing the above, but each component of the variation vector Vv is associated with two imaginary parts and one real part, and the value 0 is associated with the remaining imaginary part. The real part may be associated with a value different from 0.
　第１，第２実施例の地表種別分類装置３０，１３０では、入力ベクトルＸinｉ－ｔｙｐｅは、照射波を水平偏波、４５度偏波、左旋円偏波にしたときの位置ベクトルＰｖ，変動ベクトルＶｖを四元数に拡張したものを含むとしたが、水平偏波に代えて垂直偏波を用いてもよいし、４５度偏波に代えて－４５度偏波を用いてもよいし、左旋円偏波に代えて右旋回円偏波を用いてもよい。 In the ground type classification devices 30 and 130 of the first and second embodiments, the input vector X in i-type is a position vector Pv when the irradiation wave is horizontally polarized wave, 45 degree polarized wave, and left circularly polarized wave, Although the variation vector Vv is extended to a quaternion, a vertical polarization may be used instead of a horizontal polarization, or a −45 degree polarization may be used instead of a 45 degree polarization. However, right-handed circular polarization may be used instead of left-handed circular polarization.
　第１，第２実施例の地表種別分類装置３０，１３０では、地表の種別を湖，草地，森，街のうちのいずれか１つに分類するものとしたが、地表の種別を湖、草地、森、街、砂漠のうちの少なくとも４つのいずれかに分類するものとしてもよいし、地表の種別を湖、草地、森、街、砂漠の５つのうちのいずれか１つに分類するものとしてもよい。 In the ground surface classification devices 30 and 130 according to the first and second embodiments, the ground surface is classified into one of lake, grassland, forest, and town. It may be classified into at least four of forest, city, and desert, and the surface type may be classified as one of five among lake, grassland, forest, city, and desert. Also good.
　第１，第２実施例では、人工衛星２０から照射する光（波源）をマイクロ波としたが、波源はマイクロ波に限定されるものではなく、ミリ波、テラヘルツ波、光波などを波源として用いても構わない。 In the first and second embodiments, the light (wave source) emitted from the artificial satellite 20 is a microwave, but the wave source is not limited to the microwave, and millimeter waves, terahertz waves, light waves, etc. are used as the wave source. It doesn't matter.
　実施例の主要な要素と発明の概要の欄に記載した考案の主要な要素との対応関係について説明する。地表種別分類方法としては、実施例では、図１３に例示した種別関係学習ルーチンのステップＳ２００～Ｓ２９０の処理による学習が「種別関係学習」に相当する。地表種別分類プログラムとしては、実施例では、ＰｏｌＳＡＲ画像入力モジュール５２が「電波情報入力モジュール」に相当し、種別関係学習モジュール５４が「種別関係学習モジュール」に相当し、地表種別分類モジュール５６が「地表種別分類モジュール」に相当する。地表種別分類装置としては、実施例では、図２におけるステップＳ１００の処理を実行するＣＰＵ３２が「電波情報入力部」に相当し、図１３に例示する種別関係学習ルーチンを実行するＣＰＵ３２が「種別関係学習部」に相当し、図２におけるステップＳ１１０の処理を実行するＣＰＵ３２が「地表種別分類部」に相当する。 The correspondence between the main elements of the embodiment and the main elements of the device described in the Summary of Invention will be described. As the ground type classification method, in the embodiment, learning by the processing in steps S200 to S290 of the type relationship learning routine illustrated in FIG. 13 corresponds to “type relationship learning”. In the embodiment, as the ground classification classification program, the PolSAR image input module 52 corresponds to the “radio wave information input module”, the classification relation learning module 54 corresponds to the “class relation learning module”, and the ground classification classification module 56 is “ Corresponds to the “Surface type classification module”. In the embodiment, in the embodiment, the CPU 32 that executes the process of step S100 in FIG. 2 corresponds to the “radio wave information input unit”, and the CPU 32 that executes the type relationship learning routine illustrated in FIG. The CPU 32 that corresponds to the “learning unit” and executes the process of step S110 in FIG. 2 corresponds to the “ground type classification unit”.
　なお、実施例の主要な要素と発明の概要の欄に記載した考案の主要な要素との対応関係は、実施例が発明の概要の欄に記載した考案を実施するための形態を具体的に説明するための一例であることから、発明の概要の欄に記載した考案の要素を限定するものではない。即ち、発明の概要の欄に記載した考案についての解釈はその欄の記載に基づいて行なわれるべきものであり、実施例は発明の概要の欄に記載した考案の具体的な一例に過ぎないものである。 The correspondence between the main elements of the embodiment and the main elements of the device described in the summary column of the invention is a concrete form of the embodiment for carrying out the device described in the summary column of the invention. Since this is an example for explanation, the elements of the invention described in the summary section of the invention are not limited. That is, the invention described in the summary column should be interpreted based on the description in the column, and the embodiments are only specific examples of the invention described in the summary column. It is.
　以上、本発明を実施するための形態について実施例を用いて説明したが、本発明はこうした実施例に何等限定されるものではなく、本発明の要旨を逸脱しない範囲内において、種々なる形態で実施し得ることは勿論である。 As mentioned above, although the form for implementing this invention was demonstrated using the Example, this invention is not limited at all to such an Example, In the range which does not deviate from the summary of this invention, it is with various forms. Of course, it can be implemented.
　本発明は、地表種別分類プログラムや地表種別分類装置の製造業などに利用可能である。 The present invention can be used for the manufacturing industry of the ground classification classification program and the ground classification classification apparatus.

Claims (9)

  1. 　地表の複数の領域に対して照射した照射波と該照射波の照射に伴って得られる散乱波とからなる電波情報を入力し、前記電波情報と地表種別との関係である種別関係を学習する種別関係学習に基づいて前記複数の領域の地表種別を分類する地表種別分類方法において、
    　前記種別関係学習は、複数の領域に対する偏波情報のポアンカレ球における位置ベクトルと、該位置ベクトルの近隣の領域の位置ベクトルからの変動を示す変動ベクトルと、を四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて前記種別関係を学習する
    　地表種別分類方法。     Input radio wave information consisting of an irradiation wave irradiated to a plurality of areas on the ground surface and scattered waves obtained by the irradiation of the irradiation wave, and learn a type relationship that is a relationship between the radio wave information and the surface type. In the ground surface type classification method for classifying the ground surface types of the plurality of regions based on type relationship learning,
    In the type relationship learning, a position vector of Poincare spheres of polarization information for a plurality of regions and a variation vector indicating a variation from a position vector of a neighboring region of the position vector are expanded to a plurality of four quaternions. A method for classifying a ground surface type, wherein a neural network is used as an input value for a numerator vector to learn the type relationship.
  2. 　請求項１記載の地表種別分類方法であって、
    　前記種別関係学習は、前記入力値として、前記位置ベクトルの各成分を３つの虚数部に対応させると共に第１の値を実数部に対応させることによって前記位置ベクトルを四元数に拡張した四元数位置ベクトルと、前記変動ベクトルの各成分を３つの虚数部に対応させると共に第２の値を実数部に対応させることによって前記変動ベクトルを四元数に拡張した四元数変動ベクトルと、を含む
    　地表種別分類方法。     The ground surface type classification method according to claim 1,
    In the type relationship learning, as the input value, each component of the position vector is made to correspond to three imaginary parts and the first value is made to correspond to the real part so that the position vector is expanded to a quaternion. A number position vector, and a quaternion variation vector obtained by expanding each variation vector into a quaternion by corresponding each component of the variation vector to three imaginary parts and corresponding a second value to the real part. Include ground type classification method.
  3. 　請求項１または２記載の地表種別分類方法であって、
    　前記種別関係学習は、前記位置ベクトルとして予め定めた地表種別が既知の領域に対する偏波情報のポアンカレ球における位置ベクトルを用いて行なわれる学習であり、
    　前記種別関係学習における前記種別関係と前記入力された電波情報とに基づいて前記複数の領域の地表種別毎の関連性の程度を演算し、該演算した地表種別毎の関連性の程度に基づいて前記複数の領域の地表種別を分類する
    　地表種別分類方法。     The ground surface type classification method according to claim 1 or 2,
    The type relationship learning is learning performed using a position vector in the Poincare sphere of polarization information for an area where a predetermined ground surface type is known as the position vector,
    Based on the type relationship in the type relationship learning and the input radio wave information, the degree of relevance for each surface type of the plurality of regions is calculated, and based on the calculated degree of relevance for each surface type A ground surface type classification method for classifying the ground surface types of the plurality of regions.
  4. 　請求項３記載の地表種別分類方法であって、
    　前記種別関係学習は、前記入力値として、前記照射波を水平偏波および垂直偏波のいずれか一方の偏波状態にしたときと、前記照射波を４５度偏波および－４５度偏波のいずれか一方の偏波状態にしたときと、前記照射波を左旋円偏波および右旋円偏波のいずれか一方の偏波状態にしたときと、のそれぞれにおける３つの前記位置ベクトルおよび３つの前記変動ベクトルを四元数に拡張した６つの四元数ベクトルを含む
    　地表種別分類方法。     The ground surface classification method according to claim 3,
    In the type relationship learning, as the input value, when the irradiation wave is in a polarization state of one of horizontal polarization and vertical polarization, the irradiation wave is converted into 45 degree polarization and −45 degree polarization. The three position vectors and three in each of the polarization state when the irradiation wave is set to one of the left-handed circular polarization state and the right-handed circular polarization state A ground surface type classification method including six quaternion vectors obtained by extending the variation vector to a quaternion.
  5. 　請求項３または４記載の地表種別分類方法であって、
    　前記種別関係学習は、値０ではない第３の値を実数部に対応させると共に値０を３つの虚数部に対応させた四元数である閾値ベクトルを前記ニューラルネットワークの閾値として入力する
    　地表種別分類方法。     The ground surface classification method according to claim 3 or 4,
    In the type relation learning, a threshold vector that is a quaternion in which a third value that is not 0 is associated with a real part and a value 0 is associated with three imaginary parts is input as a threshold value of the neural network. Classification method.
  6. 　請求項３ないし５のいずれか１つの請求項に記載の地表種別分類方法であって、
    　前記地域種別は、異なるｎ個の種別であり、
    　前記ニューラルネットワークは、互いに内積が値０となる前記ｎ個のベクトルを前記入力値に対して期待される出力値とする
    　地表種別分類方法。     A ground surface type classification method according to any one of claims 3 to 5,
    The region types are n different types,
    The neural network uses the n vectors whose inner products have a value of 0 as output values expected for the input values.
  7. 　請求項１ないし６のいずれか１つの請求項に記載の地表種別分類方法であって、
    　前記地表種別は、湖、草地、森、街、砂漠のうちの少なくとも４つの種別を含む
    　地表種別分類方法。     The ground surface type classification method according to any one of claims 1 to 6,
    The surface type includes at least four types of lake, grassland, forest, city, and desert.
  8. 　地表の複数の領域の地表種別を分類する地表種別分類プログラムであって、
    　地表の複数の領域に対して照射した照射波と該照射波の照射に伴って得られる散乱波とからなる電波情報を入力する電波情報入力モジュールと、
    　複数の領域に対する偏波情報のポアンカレ球における位置ベクトルと、該位置ベクトルの近隣の領域の位置ベクトルからの変動を示す変動ベクトルと、を四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて前記電波情報と地表種別との関係である種別関係を学習する種別関係学習モジュールと、
    　前記学習した種別関係に基づいて前記複数の領域の地表種別を分類する地表種別分類モジュールと、
    　を備える地表種別分類プログラム。     A ground type classification program for classifying ground types of a plurality of areas of the ground surface,
    A radio wave information input module for inputting radio wave information composed of an irradiation wave irradiated to a plurality of regions of the ground surface and a scattered wave obtained by the irradiation of the irradiation wave;
    Multiple quaternion vectors obtained by extending the position vector of the polarization information for multiple areas in the Poincare sphere and the variation vector indicating the variation from the position vector of the neighboring area of the position vector to a quaternion A type relationship learning module that learns a type relationship that is a relationship between the radio wave information and the surface type using a neural network,
    A ground surface type classification module for classifying the ground surface types of the plurality of regions based on the learned type relationship;
    A ground classification classification program.
  9. 　地表の複数の領域の地表種別を分類する地表種別分類装置であって、
    　地表の複数の領域に対して照射した照射波と該照射波の照射に伴って得られる散乱波とからなる電波情報を入力する電波情報入力部と、
    　複数の領域に対する偏波情報のポアンカレ球における位置ベクトルと、該位置ベクトルの近隣の領域の位置ベクトルからの変動を示す変動ベクトルと、を四元数に拡張した複数の四元数ベクトルを入力値としてニューラルネットワークを用いて前記電波情報と地表種別との関係である種別関係を学習する種別関係学習部と、
    　前記学習した種別関係に基づいて前記複数の領域の地表種別を分類する地表種別分類部と、
    　を備える地表種別分類装置。     A ground type classification device that classifies ground types of a plurality of areas of the ground surface,
    A radio wave information input unit for inputting radio wave information composed of an irradiation wave irradiated to a plurality of areas on the ground surface and a scattered wave obtained by irradiation of the irradiation wave;
    Multiple quaternion vectors obtained by extending the position vector of the polarization information for multiple areas in the Poincare sphere and the variation vector indicating the variation from the position vector of the neighboring area of the position vector to a quaternion A type relationship learning unit that learns a type relationship that is a relationship between the radio wave information and the surface type using a neural network,
    A ground surface type classification unit for classifying the ground surface types of the plurality of regions based on the learned type relationship;
    A ground surface type classification apparatus comprising:
PCT/JP2014/074699 2013-09-18 2014-09-18 Terrain category classification method, terrain category classification program, and terrain category classification device WO2015041295A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2015537966A JPWO2015041295A1 (en) 2013-09-18 2014-09-18 Ground surface classification method, ground surface classification program, and ground surface classification device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US201361879348P 2013-09-18 2013-09-18
US61/879,348 2013-09-18

Publications (1)

Publication Number Publication Date
WO2015041295A1 true WO2015041295A1 (en) 2015-03-26

Family

ID=52688942

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2014/074699 WO2015041295A1 (en) 2013-09-18 2014-09-18 Terrain category classification method, terrain category classification program, and terrain category classification device

Country Status (2)

Country Link
JP (1) JPWO2015041295A1 (en)
WO (1) WO2015041295A1 (en)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102015117712A1 (en) 2015-10-16 2017-04-20 Friedrich-Alexander-Universität Erlangen-Nürnberg Imaging polarimetry
JP2017224258A (en) * 2016-06-15 2017-12-21 正仁 櫨田 Method of calculating quaternion in real number by cpu of computer
DE102016213330A1 (en) 2016-07-21 2018-03-08 Friedrich-Alexander-Universität Erlangen-Nürnberg Method for locating a mobile object and transponder for this purpose
CN107918776A (en) * 2017-11-01 2018-04-17 中国科学院深圳先进技术研究院 A kind of plan for land method, system and electronic equipment based on machine vision
CN108171273A (en) * 2018-01-12 2018-06-15 西安电子科技大学 Classification of Polarimetric SAR Image method based on K-SVD and depth heap stack network
CN108564006A (en) * 2018-03-26 2018-09-21 西安电子科技大学 Based on the polarization SAR terrain classification method from step study convolutional neural networks
CN109116350A (en) * 2017-06-22 2019-01-01 波音公司 Synthetic aperture radar mapping and registration arrangement and method
CN109145870A (en) * 2018-09-12 2019-01-04 中国农业大学 Two stages salt-soda soil monitoring method based on SAR polarization decomposing and convolutional neural networks
JP2019035598A (en) * 2017-08-10 2019-03-07 株式会社パスコ Land cover learning data generation device, land cover learning data, land cover classification prediction device, and land cover learning data generation program
CN109784413A (en) * 2019-01-23 2019-05-21 西安电子科技大学 A kind of Classification of Polarimetric SAR Image method based on long short-term memory circulation nerve net
CN110414566A (en) * 2019-07-01 2019-11-05 武汉大学 A kind of Land cover types change detecting method based on time series PolSAR image
JPWO2021199245A1 (en) * 2020-03-31 2021-10-07

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112966779A (en) * 2021-03-29 2021-06-15 安徽大学 PolSAR image semi-supervised classification method

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2005078175A (en) * 2003-08-28 2005-03-24 Toa Corp Data processor and data processing method
JP2006140952A (en) * 2004-11-15 2006-06-01 Toa Corp Image processor and image processing method

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2005078175A (en) * 2003-08-28 2005-03-24 Toa Corp Data processor and data processing method
JP2006140952A (en) * 2004-11-15 2006-06-01 Toa Corp Image processor and image processing method

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
FANG SHANG ET AL.: "Classification features in phase components of mechanism vectors in PolInSAR optimization", ANTENNAS AND PROPAGATION (ISAP), 2012 INTERNATIONAL SYMPOSIUM ON, 29 October 2012 (2012-10-29), pages 114 - 117 *
FANG SHANG ET AL.: "Polarimetric-basis transformation for land classification in PolInSAR", ELECTRONICS LETTERS, vol. 49, no. 1, 3 January 2013 (2013-01-03), pages 69 - 71 *
FANG SHANG ET AL.: "PolSAR land classification by using quaternion-valued neural networks", SYNTHETIC APERTURE RADAR (APSAR), 2013 ASIA- PACIFIC CONFERENCE ON, 23 September 2013 (2013-09-23), pages 593 - 596 *
FANG SHANG ET AL.: "Use of Poincare sphere parameters for fast supervised PolSAR land classification", GEOSCIENCE AND REMOTE SENSING SYMPOSIUM (IGARSS, 21 July 2013 (2013-07-21), pages 3175 - 3178 *
YOSHIO YAMAGUCHI, RADAR POLARIMETRY FROM BASICS TO APPLICATIONS: RADAR REMOTE SENSING USING POLARIMETRIC INFORMATION, 25 December 2007 (2007-12-25), pages 140 - 151 *

Cited By (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017063894A1 (en) 2015-10-16 2017-04-20 Friedrich-Alexander-Universität Erlangen-Nürnberg Imaging polarimetry
US11112487B2 (en) 2015-10-16 2021-09-07 Symeo Gmbh Method for performing imaging polarimetry, transponder, and system for performing imaging polarimetry
DE102015117712A1 (en) 2015-10-16 2017-04-20 Friedrich-Alexander-Universität Erlangen-Nürnberg Imaging polarimetry
JP2017224258A (en) * 2016-06-15 2017-12-21 正仁 櫨田 Method of calculating quaternion in real number by cpu of computer
DE102016213330A1 (en) 2016-07-21 2018-03-08 Friedrich-Alexander-Universität Erlangen-Nürnberg Method for locating a mobile object and transponder for this purpose
CN109116350A (en) * 2017-06-22 2019-01-01 波音公司 Synthetic aperture radar mapping and registration arrangement and method
CN109116350B (en) * 2017-06-22 2023-08-22 波音公司 System and method for synthetic aperture radar
JP2019035598A (en) * 2017-08-10 2019-03-07 株式会社パスコ Land cover learning data generation device, land cover learning data, land cover classification prediction device, and land cover learning data generation program
JP7045819B2 (en) 2017-08-10 2022-04-01 株式会社パスコ Land cover learning data generator, land cover learning data, land cover classification prediction device and land cover learning data generation program
CN107918776B (en) * 2017-11-01 2022-03-22 中国科学院深圳先进技术研究院 Land planning method and system based on machine vision and electronic equipment
CN107918776A (en) * 2017-11-01 2018-04-17 中国科学院深圳先进技术研究院 A kind of plan for land method, system and electronic equipment based on machine vision
CN108171273A (en) * 2018-01-12 2018-06-15 西安电子科技大学 Classification of Polarimetric SAR Image method based on K-SVD and depth heap stack network
CN108171273B (en) * 2018-01-12 2022-04-19 西安电子科技大学 Polarimetric SAR image classification method based on K-SVD and depth stack network
CN108564006A (en) * 2018-03-26 2018-09-21 西安电子科技大学 Based on the polarization SAR terrain classification method from step study convolutional neural networks
CN108564006B (en) * 2018-03-26 2021-10-29 西安电子科技大学 Polarized SAR terrain classification method based on self-learning convolutional neural network
CN109145870A (en) * 2018-09-12 2019-01-04 中国农业大学 Two stages salt-soda soil monitoring method based on SAR polarization decomposing and convolutional neural networks
CN109145870B (en) * 2018-09-12 2021-06-25 中国农业大学 Two-stage saline-alkali soil monitoring method based on SAR polarization decomposition and convolutional neural network
CN109784413A (en) * 2019-01-23 2019-05-21 西安电子科技大学 A kind of Classification of Polarimetric SAR Image method based on long short-term memory circulation nerve net
CN110414566A (en) * 2019-07-01 2019-11-05 武汉大学 A kind of Land cover types change detecting method based on time series PolSAR image
WO2021199245A1 (en) * 2020-03-31 2021-10-07 日本電気株式会社 Analysis device, analysis method, and storage medium
JPWO2021199245A1 (en) * 2020-03-31 2021-10-07
JP7375915B2 (en) 2020-03-31 2023-11-08 日本電気株式会社 Analytical equipment, analytical methods and programs

Also Published As

Publication number Publication date
JPWO2015041295A1 (en) 2017-03-02

Similar Documents

Publication Publication Date Title
WO2015041295A1 (en) Terrain category classification method, terrain category classification program, and terrain category classification device
US11915501B2 (en) Object detection method and apparatus, electronic device, and storage medium
Wen et al. Directionally constrained fully convolutional neural network for airborne LiDAR point cloud classification
CN109685152B (en) Image target detection method based on DC-SPP-YOLO
Wen et al. Airborne LiDAR point cloud classification with global-local graph attention convolution neural network
JP2022101650A (en) Data discriminator training method, data discriminator training device, program and training method
CN110555811A (en) SAR image data enhancement method and device and storage medium
CN110276377A (en) A kind of confrontation sample generating method based on Bayes&#39;s optimization
Lee et al. Perceptual organization of 3D surface points
CN113033520A (en) Tree nematode disease wood identification method and system based on deep learning
Liu et al. Where should we place lidars on the autonomous vehicle?-an optimal design approach
JP2020042503A (en) Three-dimensional symbol generation system
CN111458691A (en) Building information extraction method and device and computer equipment
He et al. Human segmentation of infrared image for mobile robot search
CN111965645B (en) Multi-baseline vegetation height inversion method and device considering geometric constraints
CN116933141B (en) Multispectral laser radar point cloud classification method based on multicore graph learning
CN113160117A (en) Three-dimensional point cloud target detection method under automatic driving scene
CN112926052A (en) Deep learning model security vulnerability testing and repairing method, device and system based on genetic algorithm
CN116612097A (en) Method and system for predicting internal section morphology of wood based on surface defect image
WO2023159073A1 (en) Methods and systems of sensor fusion in cooperative perception systems
Phadke et al. An analysis of trends in UAV swarm implementations in current research: Simulation versus hardware
Roos et al. A method for evaluation and optimization of automotive camera systems based on simulated raw sensor data
Soemantoro et al. An AI automated self-organising, feature imitating approach for point cloud data reduction
CN111507938A (en) Human body dangerous article detection method and system
Cheng et al. Research on Calibration Method of Camera Stabilizer Based on Binocular Vision

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 14845096

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2015537966

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 14845096

Country of ref document: EP

Kind code of ref document: A1