CN108734122B - Hyperspectral urban water body detection method based on self-adaptive sample selection - Google Patents

Hyperspectral urban water body detection method based on self-adaptive sample selection Download PDF

Info

Publication number
CN108734122B
CN108734122B CN201810471321.8A CN201810471321A CN108734122B CN 108734122 B CN108734122 B CN 108734122B CN 201810471321 A CN201810471321 A CN 201810471321A CN 108734122 B CN108734122 B CN 108734122B
Authority
CN
China
Prior art keywords
water body
area
original
suspected
image
Prior art date
Legal status (The legal status 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 status listed.)
Active
Application number
CN201810471321.8A
Other languages
Chinese (zh)
Other versions
CN108734122A (en
Inventor
唐林波
王文正
邓宸伟
冯帆
赵保军
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
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 Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Priority to CN201810471321.8A priority Critical patent/CN108734122B/en
Publication of CN108734122A publication Critical patent/CN108734122A/en
Application granted granted Critical
Publication of CN108734122B publication Critical patent/CN108734122B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10048—Infrared image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/30—Assessment of water resources

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Quality & Reliability (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Investigating Or Analysing Materials By Optical Means (AREA)

Abstract

The invention provides a hyperspectral urban water body detection method based on self-adaptive sample selection, in the aspect of preprocessing of hyperspectral near-infrared spectral band images, noise band images are removed by a quality evaluation SSIM method, noise is further removed by adopting twice average operation, more stable mean images of the near-infrared spectral band can be obtained, and compared with the traditional single-band threshold segmentation method, the method does not need to manually select images to be segmented; the method extracts a suspected water body area by an unsupervised threshold segmentation method, and then eliminates ground objects such as building shadows, building roof asphalt and the like which are similar to a water body in a near infrared spectrum band from the suspected water body area through supervised feature learning and classifier training; therefore, the unsupervised threshold segmentation method and the supervised feature learning and classifier training method are combined, the self-adaptive capacity of urban observation scenes is achieved, the actual acquisition and the actual processing of actual measurement scene data can be realized, and the false alarm rate is low.

Description

Hyperspectral urban water body detection method based on self-adaptive sample selection
Technical Field
The invention belongs to the technical field of remote sensing image processing, and particularly relates to a hyperspectral urban water body detection method based on adaptive sample selection.
Background
With the rapid development of the hyperspectral remote sensing technology, the application of hyperspectrum in military and civil life is more and more extensive, so that the requirement of hyperspectral data processing aiming at specific application environments is higher and higher. The urban water area detection of the hyperspectral remote sensing image is an important research direction, and the method has important significance in aspects of urban hydrological monitoring, urban water network planning, urban ecology, environmental monitoring and the like. However, urban background ground objects are various and complex in types, which brings challenges to the accurate extraction of urban water bodies.
The existing hyperspectral urban water area detection method directly applies the existing deep learning method, converts urban water body detection into the classification problem of ground object types for processing, ignores prior information of urban background ground object types, has extremely long and tedious training time, needs to consume massive training samples and depends on large-scale parallel computing resources, and is not fully proved in theory, and the black box attribute is still obvious. In addition, the categories of the image ground objects in the hyperspectral urban area are complex and various, due to the influence of factors such as shooting angles, environments, the phenomenon of spectrum difference of the same object and the like, the identification of the ground objects by directly adopting original hyperspectral data is usually low in detection rate, and the ground objects such as building shadows, building roof asphalt and the like have spectral properties which are similar to those of a water body in a near-infrared spectral band, so that the problem that the false alarm is too high easily occurs in the existing hyperspectral urban water area detection method.
Disclosure of Invention
In order to solve the problems, the invention provides a hyperspectral urban water body detection method based on adaptive sample selection, which can remove the interference of factors such as asphalt, shadow and the like and reduce the detection false alarm rate.
A hyperspectral urban water body automatic detection method comprises the following steps:
s1: acquiring a mean value image of all near infrared spectrum band images in an original hyperspectral image, and taking the mean value image as a first mean value image;
s2: respectively acquiring SSIM values of the first mean image and all near infrared spectrum band images, eliminating near infrared spectrum band images corresponding to the SSIM value smaller than a preset threshold value T1, and acquiring mean images of the rest near infrared spectrum band images again, wherein the mean images are used as second mean images;
s3: performing threshold segmentation on the second mean value image to obtain a suspected water body area;
s4: respectively performing expansion operation and corrosion operation on the suspected water body area, taking an original spectrum corresponding to the result of the expansion operation in the original hyperspectral image as a positive sample, and taking an original spectrum corresponding to the result of the corrosion operation in the original hyperspectral image as a negative sample; the original spectrum is a spectrum formed in a full-wave band range by taking the wavelength as a variable according to the pixel value of each pixel point in the original hyperspectral image;
s5: training a spectral feature classifier according to the spectral features of the positive sample and the spectral features of the negative sample by adopting a machine learning method;
s6: and (4) rechecking the original spectrum of the suspected water body area obtained in the step (S3) in the original hyperspectral image by adopting the spectral feature classifier, so as to obtain the urban water body area.
Further, the method for obtaining the positive sample comprises the following steps:
performing expansion operation on the suspected water body area, and removing the suspected water body area with the area smaller than a preset threshold value T2, so that the remaining suspected water body area is used as a water body sample extraction area;
and acquiring an original spectrum corresponding to the water body central area from an original hyperspectral image according to the coordinates of the water body central area in the water body sample extraction area, wherein the original spectrum is the positive sample.
Further, the method for obtaining the negative sample comprises the following steps:
carrying out corrosion operation on the suspected water body area, expanding the suspected water body area, wherein the expanded part is a peripheral area of the suspected water body area, and taking the peripheral area as a non-water body sample extraction area;
and acquiring an original spectrum corresponding to the non-water body sample extraction area from an original hyperspectral image according to the coordinates of the non-water body sample extraction area, wherein the original spectrum is the negative sample.
Further, the method for acquiring the spectral characteristics of the positive sample and the spectral characteristics of the negative sample comprises the following steps:
randomly selecting part of positive samples and part of negative samples to train a spectral feature extraction network model by adopting a deep learning method, and obtaining the spectral feature extraction network model;
and extracting a network model according to the spectral characteristics, and extracting the spectral characteristics of the remaining positive samples and the remaining negative samples.
Further, the method for acquiring the original spectrum of the suspected water body area corresponding to the original hyperspectral image comprises the following steps:
acquiring coordinates of pixel points where the suspected water body area is located;
and in the full-wave band range of the original hyperspectral image, taking the pixel value corresponding to the pixel point with the same coordinate as the suspected water body area as the original spectrum corresponding to the suspected water body area in the original hyperspectral image.
Has the advantages that:
1. the invention provides a hyperspectral urban water body detection method based on self-adaptive sample selection, in the aspect of preprocessing of hyperspectral near-infrared spectral band images, noise band images are removed by a quality evaluation SSIM method, noise is further removed by adopting twice average operation, more stable mean images of the near-infrared spectral band can be obtained, and compared with the traditional single-band threshold segmentation method, the method does not need to manually select images to be segmented;
the method extracts a suspected water body area by an unsupervised threshold segmentation method, and then eliminates ground objects such as building shadows, building roof asphalt and the like which are similar to a water body in a near infrared spectrum band from the suspected water body area through supervised feature learning and classifier training; therefore, the unsupervised threshold segmentation method and the supervised feature learning and classifier training method are combined, the self-adaptive capacity of urban observation scenes is achieved, the actual acquisition and the actual processing of actual measurement scene data can be realized, and the false alarm rate is low; meanwhile, the invention selects the positive sample and the negative sample through the expansion operation and the corrosion operation respectively, and the sample selection method is self-adaptive and does not need to select the sample manually.
2. Because the water body is different from the space geometric characteristics of the ground objects which are easy to cause false alarm of the water body, such as shadow, asphalt and the like, the urban water bodies, such as riverways, landscape lakes and the like, have larger areas, and the areas of building shadow and roof asphalt are smaller; according to the method, the prior information of the urban observation area is fully utilized, the suspected water body area with a small area is removed through expansion operation, the large-area suspected water body area is reserved, so that a positive sample is obtained, the problem that the suspected water body area possibly contains the water body peripheral area caused by the error of the threshold segmentation method can be solved, meanwhile, the range of the suspected water body area is expanded through corrosion operation, the suspected water body peripheral area is also contained in the range of the suspected water body area, so that a negative sample is obtained, and the threshold segmentation error is avoided; therefore, the method is based on the prior information of the urban observation area, and is beneficial to automatic selection of urban water body samples.
3. According to the coordinates of the suspected water body area, extracting an original spectrum in a full-wave band range corresponding to the suspected water body area from the original hyperspectral image; compared with the traditional method for urban water body detection only by using data of a hyperspectral near-infrared spectrum band, the method disclosed by the invention has the advantages that the original spectrum in a full-wave band range is input into the spectral feature classifier for secondary water body detection, namely, the interference of factors such as asphalt and shadow is removed by using all spectral information, and the detection false alarm rate can be effectively reduced.
Drawings
FIG. 1 is a flow chart of a hyperspectral urban water body detection method based on adaptive sample selection according to the invention;
FIG. 2 is a flow chart of hyperspectral near-infrared noise spectrum band rejection based on image quality evaluation provided by the invention;
FIG. 3 is a flow chart of a positive and negative sample acquisition method provided by the present invention;
fig. 4 is a flowchart of the spectral feature extraction and classifier training provided by the present invention.
Detailed Description
In order to make the technical solutions better understood by those skilled in the art, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.
Referring to fig. 1, the figure is a flowchart of a hyperspectral urban water body detection method based on adaptive sample selection according to this embodiment. A hyperspectral urban water body detection method based on adaptive sample selection comprises the following steps:
s1: and acquiring a mean value image of all near infrared spectrum band images in the original hyperspectral image, and taking the mean value image as a first mean value image.
Specifically, an average value is obtained for a near-infrared spectrum image in the acquired original hyperspectral image, so that a first average value image is acquired.
S2: respectively obtaining the SSIM values of the first average value image and all near infrared spectrum band images, eliminating near infrared spectrum band images corresponding to the SSIM value smaller than a preset threshold value T1, and obtaining average value images of the rest near infrared spectrum band images again, wherein the average value images are used as second average value images.
It should be noted that the SSIM (Structural Similarity Index) value is an Index for measuring the Similarity between two images, wherein the SSIM value ranges from-1 to 1, and when two images are identical, the SSIM value is 1. Because various ground objects in the hyperspectral observation image have stability in the absorption and reflection characteristics in a certain spectral range, the quality of the observation image in a single spectral range has similarity under ideal conditions, and particularly, the similarity difference of adjacent spectral range images is smaller. Therefore, whether the observation image in the near infrared spectral band contains a noise image can be judged by using the image quality evaluation SSIM as a criterion for the hyperspectral image in the spectral band range.
Referring to fig. 2, the figure is a flow chart of the hyperspectral near-infrared noise spectrum band rejection based on image quality evaluation provided in this embodiment. And taking the first mean image as a reference image for image quality evaluation. Since the near infrared spectrum contains a plurality of spectral segments, the averaging process can reduce the influence of noise spectrum segments to some extent. And adopting an image quality evaluation criterion, taking the first mean value image as a reference image, and evaluating the quality of all the near-infrared spectrum band images, namely calculating the SSIM values of the first mean value image and all the near-infrared spectrum band images.
The method for calculating the SSIM value based on the image quality evaluation criterion is described below:
Figure BDA0001663327060000061
wherein X is a first mean image, Y is a near infrared spectrum image, muXIs the mean value of the first mean image, μYIs the mean value, sigma, of the near infrared spectral band imageXIs the covariance, σ, of the first mean imageYIs the covariance of the near infrared spectral band image, C1、C2Respectively, constants for maintaining stability.
The theoretical basis for removing the noise spectrum by using the image quality evaluation SSIM is as follows: the noise spectrum image has a larger difference compared with the first mean image due to the existence of noise, so that the quality is poorer, and the quality of the non-noise spectrum image is closer to that of the first mean image. After the noise spectrum is removed, the average image is obtained again for the rest near-infrared spectrum images, and the noise spectrum is removed, namely after the near-infrared spectrum image corresponding to the SSIM value smaller than the preset threshold T1 is removed, the second average image can better reflect the strong absorption characteristics of the ground objects such as the water body and the like in the near-infrared spectrum, and has stronger stability.
S3: and performing threshold segmentation on the second mean value image to obtain a suspected water body area.
It should be noted that the OTSU algorithm may be used to search for a suitable segmentation threshold by maximizing the inter-class variance, so as to implement the adaptive segmentation of the second mean image. And the OTSU threshold segmentation divides the second mean image into a background part and a foreground part according to the gray characteristic of the near-infrared second mean image. The larger the inter-class variance between the background and the foreground is, the larger the difference between the two parts constituting the second mean image is, and when part of the foreground is mistaken for the background or part of the background is mistaken for the foreground, the difference between the two parts is reduced.
Thus, the OTSU algorithm can minimize the probability of a false score for the second mean image segmentation of the near infrared spectral band. The water body, the asphalt and the shadow have strong absorption characteristics in a near-infrared spectrum band, and the gray values of the water body, the asphalt and the shadow areas reflected on the second mean image are smaller, so that stable and accurate extraction of the water body, the asphalt and the shadow can be realized by adopting an OTSU algorithm, and a suspected water body area is obtained; that is, the suspected water body area may contain asphalt or architectural shadows, etc., in addition to the water body.
S4: respectively performing expansion operation and corrosion operation on the suspected water body area, taking an original spectrum corresponding to the result of the expansion operation in the original hyperspectral image as a positive sample, and taking an original spectrum corresponding to the result of the corrosion operation in the original hyperspectral image as a negative sample; the original spectrum is a spectrum formed in a full-wave band range by taking the wavelength as a variable according to the pixel value of each pixel point in the original hyperspectral image.
It should be noted that the urban water body area is often a river channel (width is generally greater than 5 meters), a landscape lake (general area is greater than 100 square meters) and the like with a large volume, and the area of the building shadow and the roof asphalt area (generally less than 50 square meters) is small, so that the urban water body area and the area where the ground object which is easy to cause the false alarm of water body detection meet certain space geometric characteristics.
According to the calculation of the resolution ratio (1.5 m) of the existing airborne hyperspectral observation image, the width of a river channel is generally larger than 3 pixels, the area of a landscape lake is larger than 36 pixels, and the areas of building shadows, asphalt and the like are generally smaller than 10 pixels, so that the areas which are possibly the building shadows and the roof asphalt can be screened out through the area size, and water body areas such as the river channel, the landscape lake and the like with larger areas are reserved.
Referring to fig. 3, it is a flowchart of a positive sample and a negative sample obtaining method provided in this embodiment. A method for obtaining positive and negative examples is described below, comprising the steps of:
s401: and performing expansion operation on the suspected water body area, and removing the suspected water body area with the area smaller than a preset threshold value T2, so that the remaining suspected water body area is used as a water body sample extraction area.
It should be noted that the expansion operation is to enhance and expand the area with large gray value, i.e. the visually brighter non-water body area, so as to communicate the suspected water body area with the area smaller than the preset threshold T2 with the non-water body area.
Optionally, an expansion window with a size of 3 × 3 pixels is set, the suspected water body area obtained in step S2 is expanded, the suspected water body area with an area smaller than a preset threshold T2 is removed, and a large-area suspected water body area is reserved.
It should be noted that the dilation operation may eliminate the problem that the suspected water body area may contain the water body edge area due to the error of the threshold segmentation in step S2, and ensure the accuracy of the selection of the spectral sample (positive sample) of the water body area.
S402: and acquiring an original spectrum corresponding to the water body central area from an original hyperspectral image according to the coordinates of the water body central area in the water body sample extraction area, wherein the original spectrum is the positive sample.
It should be noted that, because the central area of the water body is relatively pure, there is no interference from other surface feature types, and in order to ensure that the selected positive sample is the water body spectrum data, the embodiment selects the original spectrum corresponding to the central area of the water body such as a river, a landscape lake, and the like as the positive sample.
Optionally, the centroid of the water body sample extraction area is obtained by counting the coordinates of the water body sample extraction area, and the coordinates of the centroid are used as the coordinates of the water body central area.
It should be noted that, in order to reduce the computational complexity, part of the water center area may be randomly selected, and the full-band original spectrum corresponding to the randomly selected part of the water center area in the original hyperspectral data is used as a positive sample.
S403: and carrying out corrosion operation on the suspected water body area, expanding the suspected water body area, wherein the expanded part is a peripheral area of the suspected water body area, and taking the peripheral area as a non-water body sample extraction area.
It should be noted that, in the corrosion operation, the region with a small gray value, i.e., the visually darker water body region, is enhanced and expanded to remove the brighter noise, so that the peripheral region of the water body region is also drawn into the water body region. In order to further expand the range of the suspected water body area and avoid a threshold segmentation error, the embodiment performs a corrosion operation on the suspected water body area.
Optionally, the etching operation comprises the steps of:
s403 a: and (3) corroding the suspected water body area obtained in the step (S3) through a corrosion window with the size of 3 x 3 to expand the range of the suspected water body, and containing the peripheral area of the suspected water body area into the range of the suspected water body area, so that the range of the suspected water body area is expanded.
S403 b: and (4) performing the etching operation with the window of 3 × 3 again on the etched result of the step S403a, obtaining the water body peripheral area, and taking the partial area as the non-water body sample extraction area.
It should be noted that, in order to avoid a threshold segmentation error, the range of the suspected water body area is further expanded, and it is ensured that the non-water body sample extraction area does not include a water body, in this embodiment, the suspected water body area is subjected to twice iterative etching operations with 3 × 3 pixel points as the etching window size, so that areas within a range of approximately 5 meters around the water body are all drawn into the suspected water body area, and the partial areas are used as the non-water body sample extraction area.
S404: and acquiring an original spectrum corresponding to the non-water body sample extraction area from an original hyperspectral image according to the coordinates of the non-water body sample extraction area, wherein the original spectrum is the negative sample.
It should be noted that, in order to reduce the computational complexity, part of non-water body sample extraction areas may be randomly selected, and the full-band original spectrum corresponding to the randomly selected part of non-water body sample extraction areas in the original hyperspectral data is used as a negative sample.
Therefore, in the step, the suspected water body area is screened based on the spatial characteristics of the prior urban water body area, the area which can be determined as the water body and the water body edge part which is difficult to distinguish by only using near infrared spectrum band data at the periphery of the water body are obtained, and water body spectrum sample data (positive sample) and non-water body spectrum sample data (negative sample) are automatically extracted on the basis.
S5: and finishing the training of the spectral feature classifier according to the spectral features of the positive sample and the spectral features of the negative sample by adopting a machine learning method.
Optionally, the method for acquiring the spectral characteristics of the positive sample and the spectral characteristics of the negative sample comprises:
randomly selecting part of positive samples and part of negative samples to train a spectral feature extraction network model by adopting a deep learning method, and obtaining the spectral feature extraction network model;
and extracting a network model according to the spectral characteristics, and extracting the spectral characteristics of the remaining positive samples and the remaining negative samples.
Referring to fig. 4, it is a flowchart of the spectral feature extraction and classifier training provided in this embodiment.
The method for extracting spectral features is described as follows, comprising the following steps:
firstly, carrying out nonlinear mapping on an original spectrum data set T of a positive sample and a negative sample through a random projection strategy so as to obtain hidden node output h (T) of a network model of a feature extraction part, wherein the expression of h (T) is as follows:
h(T)＝[g1(T),....,gN(T)]
gi(T)＝g(ωiT+bi),i＝1,2,……N
wherein, giAnd (T) is the output of the ith hidden layer node. The hidden node weight omega and the bias b are both randomly generated through a random projection strategy, and the activation function g (x) of the hidden node is selected as a sigmoid function.
When the total number of the positive samples and the negative samples of the training set is N, the whole process of spectral feature extraction can be expressed as follows:
Hβ＝T
wherein H ═ H1 T(T),....hT N(T)]TThe network output weight β can be obtained by solving the above equation. Beta is the characteristic characterization base of the input spectrum data and the parameter to be solved of the characteristic extraction network.
The process of solving the network weight beta by the feature extraction network model can optimize the following function to solve the problem:
Figure BDA0001663327060000111
wherein λ is the training error and l1Balance weights between regularization terms. By traversing the limited candidate values, parameters which can enable the performance of the model to reach the best can be obtained, and the parameters at the moment are selected as the optimal values of the feature extraction network. The spectral characteristics of the input spectral data are characteristic base beta of the input spectral data in solutionAEThe calculation formula of the projection is as follows:
Figure BDA0001663327060000112
Tproji.e. as an output of the spectral feature extraction network part, may be an input to the classifier.
Based on the above method for extracting spectral features, the following introduces a training process of a classifier, including the following steps:
firstly, randomly initializing hidden node weight omega and bias b; then, a hidden node output matrix H is calculated0(ii) a And finally, optimizing output weight and training error:
Figure BDA0001663327060000113
in the above formula, T0＝[t1,t2,...,tN]TAre labels of positive and negative spectral samples,
Figure BDA0001663327060000114
β0i.e. the offline decision parameters of the classifier. And (5) finishing the classifier training of the positive and negative spectrum samples by solving the expression.
S6: and (4) rechecking the original spectrum of the suspected water body area obtained in the step (S3) in the original hyperspectral image by adopting the spectral feature classifier, so as to obtain the urban water body area.
Specifically, the spectral data of the suspected water body area acquired in step S3 is transmitted as input data to the spectral feature extraction network model acquired in step S5, so as to realize spectral feature extraction; and then, the spectral characteristics are input to the classifier obtained in the step S5, and final urban water body detection is realized through characteristic classification, so that the interference of factors such as asphalt and shadow is removed through the original spectrum in the full-waveband range, namely all the spectral information, and the detection false alarm rate is reduced.
The review process can be described as: and (3) if the original spectrum of the suspected water body area in the original hyperspectral image is a water body spectrum sample to be detected, extracting a network model from the trained features of the water body spectrum sample to be detected T', and obtaining a response value f (T) of a final output node of the deep learning network according to the node parameters of the network model. The response value f (t) is calculated as follows:
f(T')＝π(T')βdecision
wherein pi (T') represents the spectral characteristics of the extracted spectral sample to be measured, betadecisionNetwork parameters of the decision classification layer. When the response value is higher than the fixed threshold value, the rechecking of the suspected water body area obtained in the step S2 can be completed, so that the elimination of interference ground objects such as shadow, asphalt and the like in the suspected water body area is realized, and the urban water body area is obtained.
Optionally, the method for acquiring the original spectrum of the suspected water body region corresponding to the original hyperspectral image includes:
acquiring coordinates of pixel points where the suspected water body area is located;
and in the full-wave band range of the original hyperspectral image, taking the pixel value corresponding to the pixel point with the same coordinate as the suspected water body area as the original spectrum corresponding to the suspected water body area in the original hyperspectral image. That is to say, the pixel value of the pixel point in the full-wave band range is obtained, and the curve formed by the pixel value with the wavelength as the independent variable is the original spectrum corresponding to the pixel point in the original hyperspectral image.
Therefore, the hyperspectral urban water body self-adaptive detection based on the automatic sample selection is completed.
The present invention may be embodied in other specific forms without departing from the spirit or essential attributes thereof, and it will be understood by those skilled in the art that various changes and modifications may be made herein without departing from the spirit and scope of the invention as defined in the appended claims.

Claims (3)

1. A hyperspectral urban water body automatic detection method is characterized by comprising the following steps:
s1: acquiring a mean value image of all near infrared spectrum band images in an original hyperspectral image, and taking the mean value image as a first mean value image;
s2: respectively acquiring SSIM values of the first mean image and all near infrared spectrum band images, eliminating near infrared spectrum band images corresponding to the SSIM value smaller than a preset threshold value T1, and acquiring mean images of the rest near infrared spectrum band images again, wherein the mean images are used as second mean images;
s3: performing threshold segmentation on the second mean value image to obtain a suspected water body area;
s4: performing expansion operation on the suspected water body area, and removing the suspected water body area with the area smaller than a preset threshold value T2, so that the remaining suspected water body area is used as a water body sample extraction area;
acquiring an original spectrum corresponding to the water body central area from an original hyperspectral image according to the coordinates of the water body central area in the water body sample extraction area, wherein the original spectrum is a positive sample;
carrying out corrosion operation on the suspected water body area, expanding the suspected water body area, wherein the expanded part is a peripheral area of the suspected water body area, and taking the peripheral area as a non-water body sample extraction area;
acquiring an original spectrum corresponding to the non-water body sample extraction area from an original hyperspectral image according to the coordinates of the non-water body sample extraction area, wherein the original spectrum is a negative sample;
the original spectrum is a spectrum formed in a full-wave band range by taking the wavelength as a variable according to the pixel value of each pixel point in the original hyperspectral image;
s5: training a spectral feature classifier according to the spectral features of the positive sample and the spectral features of the negative sample by adopting a machine learning method;
s6: and (4) rechecking the original spectrum of the suspected water body area obtained in the step (S3) in the original hyperspectral image by adopting the spectral feature classifier, so as to obtain the urban water body area.
2. The method for automatically detecting the water body in the hyperspectral urban area according to claim 1, wherein the method for acquiring the spectral characteristics of the positive sample and the spectral characteristics of the negative sample comprises the following steps:
randomly selecting a part of positive samples and a part of negative samples to train a spectral feature extraction network model by adopting a deep learning method, and acquiring the spectral feature extraction network model;
and extracting a network model according to the spectral characteristics, and extracting the spectral characteristics of the remaining positive samples and the remaining negative samples.
3. The method for automatically detecting the water body in the hyperspectral urban area according to claim 1, wherein the method for acquiring the original spectrum of the suspected water body area corresponding to the original hyperspectral image comprises the following steps:
acquiring coordinates of pixel points where the suspected water body area is located;
and in the full-wave band range of the original hyperspectral image, taking the pixel value corresponding to the pixel point with the same coordinate as the suspected water body area as the original spectrum corresponding to the suspected water body area in the original hyperspectral image.
CN201810471321.8A 2018-05-17 2018-05-17 Hyperspectral urban water body detection method based on self-adaptive sample selection Active CN108734122B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810471321.8A CN108734122B (en) 2018-05-17 2018-05-17 Hyperspectral urban water body detection method based on self-adaptive sample selection

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810471321.8A CN108734122B (en) 2018-05-17 2018-05-17 Hyperspectral urban water body detection method based on self-adaptive sample selection

Publications (2)

Publication Number Publication Date
CN108734122A CN108734122A (en) 2018-11-02
CN108734122B true CN108734122B (en) 2022-05-20

Family

ID=63937530

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810471321.8A Active CN108734122B (en) 2018-05-17 2018-05-17 Hyperspectral urban water body detection method based on self-adaptive sample selection

Country Status (1)

Country Link
CN (1) CN108734122B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110287843B (en) * 2019-06-18 2022-01-25 清华大学深圳研究生院 Cultural relic fingerprint area selection method
CN111968101A (en) * 2020-08-26 2020-11-20 成都唐源电气股份有限公司 Tunnel water seepage detection method and device
CN114049571A (en) * 2022-01-13 2022-02-15 西南石油大学 Method and device for extracting water body area of hyperspectral image and electronic equipment
CN117409203B (en) * 2023-11-14 2024-04-02 自然资源部国土卫星遥感应用中心 Shallow lake area extraction method

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104463203A (en) * 2014-12-03 2015-03-25 复旦大学 Hyper-spectral remote sensing image semi-supervised classification method based on ground object class membership grading
CN106503727A (en) * 2016-09-30 2017-03-15 西安电子科技大学 A kind of method and device of classification hyperspectral imagery
CN106886760A (en) * 2017-01-24 2017-06-23 北京理工大学 A kind of EO-1 hyperion Ship Detection combined based on empty spectrum information
US9785851B1 (en) * 2016-06-30 2017-10-10 Huron Valley Steel Corporation Scrap sorting system

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010057081A1 (en) * 2008-11-14 2010-05-20 The Scripps Research Institute Image analysis platform for identifying artifacts in samples and laboratory consumables

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104463203A (en) * 2014-12-03 2015-03-25 复旦大学 Hyper-spectral remote sensing image semi-supervised classification method based on ground object class membership grading
US9785851B1 (en) * 2016-06-30 2017-10-10 Huron Valley Steel Corporation Scrap sorting system
CN106503727A (en) * 2016-09-30 2017-03-15 西安电子科技大学 A kind of method and device of classification hyperspectral imagery
CN106886760A (en) * 2017-01-24 2017-06-23 北京理工大学 A kind of EO-1 hyperion Ship Detection combined based on empty spectrum information

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
高光谱图像海上舰船目标检测方法研究;代威等;《重庆理工大学学报（自然科学）》;20151115;第29卷(第11期);正文第1节 *

Also Published As

Publication number Publication date
CN108734122A (en) 2018-11-02

Similar Documents

Publication Publication Date Title
CN108734122B (en) Hyperspectral urban water body detection method based on self-adaptive sample selection
CN109871902B (en) SAR small sample identification method based on super-resolution countermeasure generation cascade network
CN110309781B (en) House damage remote sensing identification method based on multi-scale spectrum texture self-adaptive fusion
CN107330875B (en) Water body surrounding environment change detection method based on forward and reverse heterogeneity of remote sensing image
CN110060273B (en) Remote sensing image landslide mapping method based on deep neural network
CN109002848B (en) Weak and small target detection method based on feature mapping neural network
CN106023257A (en) Target tracking method based on rotor UAV platform
CN106557740B (en) The recognition methods of oil depot target in a kind of remote sensing images
CN109886146B (en) Flood information remote sensing intelligent acquisition method and device based on machine vision detection
CN107992856B (en) High-resolution remote sensing building shadow detection method under urban scene
CN106485239A (en) One kind is using one-class support vector machines detection river mesh calibration method
CN105447488B (en) SAR image target detection method based on sketch line segment topological structure
CN111274964B (en) Detection method for analyzing water surface pollutants based on visual saliency of unmanned aerial vehicle
CN104680151A (en) High-resolution panchromatic remote-sensing image change detection method considering snow covering effect
CN104866852B (en) Extract the method and device of windy and sandy soil information in remote sensing image
CN114764801A (en) Weak and small ship target fusion detection method and device based on multi-vision significant features
CN114022782B (en) Sea fog detection method based on MODIS satellite data
CN111310566A (en) Static and dynamic multi-feature fusion mountain fire detection method and system
CN112241956B (en) PolSAR image ridge line extraction method based on region growing method and variation function
CN111624606B (en) Radar image rainfall identification method
CN116385866B (en) SAR image-based railway line color steel house change detection method and device
CN110211106B (en) Mean shift SAR image coastline detection method based on segmented Sigmoid bandwidth
CN108764016B (en) Polarimetric SAR image ship detection method based on rotation domain characteristics and CNN
CN115861669A (en) Infrared dim target detection method based on clustering idea
CN112686222B (en) Method and system for detecting ship target by satellite-borne visible light detector

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant