CN110866364A - Ground surface temperature downscaling method based on machine learning - Google Patents
Ground surface temperature downscaling method based on machine learning Download PDFInfo
- Publication number
- CN110866364A CN110866364A CN201911141709.2A CN201911141709A CN110866364A CN 110866364 A CN110866364 A CN 110866364A CN 201911141709 A CN201911141709 A CN 201911141709A CN 110866364 A CN110866364 A CN 110866364A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- surface temperature
- spatial resolution
- rows
- data
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Processing (AREA)
Abstract
The invention relates to a machine learning-based earth surface temperature downscaling method, which comprises the following steps: acquiring low spatial resolution remote sensing index data; acquiring low-spatial-resolution earth surface temperature remote sensing data; constructing a training sample set; training a BP neural network model; acquiring high spatial resolution remote sensing index data; and taking the high spatial resolution remote sensing index data as an input independent variable, and obtaining high spatial resolution ground surface temperature downscaling calculation data by using the BP neural network model after input training. The method realizes the high-spatial-resolution ground surface temperature product data simulation, and solves the problem that the traditional downscaling method ignores the high variability of the ground surface covering type internal reflection spectrum and the thermal characteristic in the heterogeneous region.
Description
Technical Field
The invention belongs to the field of thermal infrared remote sensing, and particularly relates to a ground surface temperature downscaling method based on machine learning.
Background
The thermal infrared remote sensing can acquire the earth surface temperature in the global and regional range, has the advantages of wide coverage, relatively low cost, periodic acquisition and the like, and is an important way for acquiring the earth surface temperature at present. At present, main satellite-borne thermal infrared sensors include AVHRR, MODIS, Landsat series, enter, and the like, but due to technical limitations, spatial resolution of an obtained thermal infrared image is generally low, and actual application requirements cannot be met. In view of the close relationship between the surface temperature distribution and the surface coverage change, the method for performing the spatial downscaling of the surface temperature by using the surface coverage information with high spatial resolution becomes an effective means for acquiring the surface temperature with high resolution. The existing ground surface temperature space scale reduction method mainly comprises a statistical-based method and a spectral mixture model-based method.
The statistical-based downscaling algorithm applies a statistical relationship between the low-resolution surface temperature and the spectral index to the high resolution by assuming "the relationship scale is unchanged", which is usually obtained by building a linear regression model. For example, Kustas et al propose the DisTracd algorithm by using the relationship between the surface temperature and the NDVI, and successfully obtain the surface temperature with the resolution of 250m through the MODIS thermal infrared image with the resolution of 1 km; the TsHARP algorithm is provided on the basis of the Distrad algorithm, and the unary linear relation between the vegetation coverage index and the earth surface temperature is proved to have the optimal effect on the size reduction of the earth surface temperature; dominguez et al proposes a HUTS algorithm based on the two previous algorithms, and obtains surface temperature information with high spatial resolution by using scale invariance of a relationship between surface temperature and NDVI and a surface albedo binary quartic polynomial. Meanwhile, on the basis of a traditional linear regression model, in order to ensure the downscaling precision, Zhu et al proposes an improved hierarchical stepwise regression method, applies ETM + images of Shanghai regions to decompose the earth surface temperature, and finds that the method has better precision than DisTrad and TsHAPP algorithms; guo Huizhi et al indicate that the city index (UI) has a high correlation with LST, and that accuracy can be improved by using a UI hierarchical regression method to perform surface temperature downscaling. Although these algorithms consider the physical factors behind the LST variation, and are reliable in precision and easy to operate, this kind of method usually lacks the consideration of regional differences, and the existing models are applied to other regions with larger errors.
The method based on the spectrum mixed model is that the mixed pixel value is assumed to be equal to the abundance weighting of units in the mixed pixel, and high-resolution and low-resolution earth surface temperatures are directly related by obtaining the abundance of different earth surface components, so that the spatial downscaling is carried out, and the earth surface temperature decomposition is realized. Zhukov et al proposed an earth surface temperature decomposition method-MMT method based on linear spectral decomposition, and applied it to TM images and simulated ASTER images, to obtain higher resolution earth surface temperature; deng and Wu synthesize the abundance of different earth surface components and earth surface temperatures of each component, and provide an earth surface temperature downscaling method-SUTM based on mixed pixel decomposition and thermal mixing, which effectively realizes the spatial downscaling of urban earth surface temperatures, and then they provide a method (VHR-SUTM) for estimating high-resolution earth surface temperatures on the basis of the SUTM algorithm, successfully utilize IKONOS high-resolution images and TM thermal infrared images to estimate earth surface temperatures with 4m spatial resolution, and provide a feasible solution for obtaining high-spatial resolution earth surface temperatures. However, the traditional ground surface temperature downscaling method based on the spectrum mixed model assumes the same object and temperature, neglects the variability problem of reflection spectrum and thermal characteristics of the same ground surface coverage type, especially, the intra-class variability is larger under high spatial resolution, which causes a large amount of ground surface temperature changes under the high spatial resolution to be unable to be accurately estimated, and greatly affects the accuracy of high-resolution ground surface temperature.
Disclosure of Invention
The invention aims to solve the technical problem of providing a machine learning-based ground surface temperature downscaling method, which not only solves the problem of instability of a traditional regression model, but also can solve the problem of high variability of high-resolution ground surface covering type internal reflection spectrum and thermal characteristics.
The earth surface temperature downscaling method based on machine learning comprises the following steps:
step 1: obtaining low spatial resolution remote sensing index data which is a three-dimensional floating point type image matrix, namely the height, the width and the number of layers (ro) of the imagews columns bands); height and width (rows columns) represent geographical location, each floor (bands) represents a telemetry index, x(m,n,b)Representing a b-th remote sensing index value of a pixel x of the remote sensing index image at the position of m rows and n columns;
acquiring low spatial resolution earth surface temperature remote sensing data, wherein the low spatial resolution earth surface temperature remote sensing data is a two-dimensional image matrix, namely the height and width (rows and columns) of an image correspond to a remote sensing index, and y is(m,n)Representing the surface temperature value of a pixel y of the surface temperature image at the position of m rows and n columns;
acquiring high spatial resolution remote sensing index data, wherein the high spatial resolution remote sensing index data is a three-dimensional floating point type image matrix, namely height, width, and layers (rows columns) of an image; height and width (rows columns) represent geographical location, each floor (bands) represents a telemetry index, X(m,n,b)Representing the size of a b-th remote sensing index value of a pixel X of the remote sensing index image at the position of m rows and n columns;
step 2: constructing a training sample set, respectively selecting x and y at corresponding positions according to the corresponding relation of height-width (rows) of an image matrix to construct a feature vector, taking the low-spatial-resolution remote sensing index data as an input independent variable, taking the low-spatial-resolution earth surface temperature remote sensing data as an input dependent variable, and obtaining the feature vector
And step 3: training a BP neural network model by adopting the feature vector, wherein the BP neural network model comprises an input layer, a hidden layer and an output layer, and obtaining the trained BP neural network model;
and 4, step 4: and inputting the high spatial resolution remote sensing index data serving as an input independent variable into the trained BP neural network model to obtain high spatial resolution ground surface temperature downscaling calculation data.
Optionally, a parameter customization step is further included between step 1 and step 2, and the parameter customization step includes: and selecting the number of the adopted remote sensing indexes in the low spatial resolution remote sensing index data.
Optionally, the step of customizing parameters further comprises setting the following parameters: the method comprises the steps of learning rate, the number of neurons in a hidden layer, maximum iterative learning times, a convergence threshold value, proportion distribution of training samples and verification samples and a dynamic display period of a learning process.
Optionally, the method further includes an iteration step, which includes calculating an average error after completing one-time learning of all the feature vectors in step 2, determining whether the average error is smaller than a convergence threshold, or whether the iteration number is greater than the maximum iteration learning number, and completing step 2 if the average error is smaller than the convergence threshold or the iteration number is greater than the maximum iteration learning number; otherwise, adjusting the output layer weight and the hidden layer weight by adopting an accumulated error BP algorithm, starting the next iteration step after the weight adjustment is finished, and repeating the iteration step.
Optionally, the method further comprises a parameter adjusting step, wherein in the iteration step, after the set number of times of the learning process dynamic display period is finished each time, the visual scattergram is displayed and output in real time, a user can check the fitting precision of the scattergram and the average loss of the output, and assist in judging whether the parameter measurement is reasonable, if the average loss fluctuates up and down, which indicates that the model learning rate is set too large, the gradient may oscillate back and forth near the minimum value, and the gradient cannot converge, so that the learning rate is adjusted.
Optionally, in the parameter adjusting step, if after the model training in step 3 is finished, the fitting of the scatter diagram and the average error do not meet the requirements, returning to the step of customizing parameters, and performing appropriate adjustment on the model parameters, including increasing the number of neurons in the hidden layer or the number of iterations, until the model training result meets the requirements of the user.
Optionally, the number of remote sensing indexes in the obtained low spatial resolution remote sensing index data is 14.
Optionally, in the step of customizing parameters, the parameters are set as follows: the method comprises the following steps of remote sensing index number 12, learning rate 0.0001, hidden layer neuron number 20, maximum iterative learning number 100000, convergence threshold value 0.001, proportion distribution 1 of training samples and verification samples and dynamic display period 200 of a learning process.
Compared with the prior art, the invention has the beneficial effects that:
according to the invention, visualization of the ground surface temperature downscaling execution process is realized through a packaging machine learning algorithm, interactive operation can be realized, a real-time feedback regulation function is provided, and the influence of regional difference on the result is reduced; the simulation of high-spatial-resolution ground surface temperature product data is realized, the spatial resolution is reduced to 2 m from 30 m, and spatial detail characteristics can be finely expressed; the data input is simple, the optimal remote sensing index can be selected as the input independent variable according to the actual condition of a research area, and the problem of high variability of the surface coverage type internal reflection spectrum and the thermal characteristic neglected in a heterogeneous region in the traditional scale reduction method is solved.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
FIG. 1 is a schematic diagram of the BP neural network model of the present invention.
FIG. 2 is a flow chart of one embodiment of the present invention.
FIG. 3A is a WorldView-2 color image of one embodiment of the present invention.
Fig. 3B is a surface temperature down-scale inversion diagram corresponding to fig. 3A.
FIG. 4 is an interface diagram of a trained BP neural network model according to an embodiment of the present invention.
FIG. 5 is an interface diagram of a model implementation to obtain high spatial resolution surface temperature downscaling calculation data, in accordance with one embodiment of the present invention.
FIG. 6 is an interface diagram of a scatter plot fit and average error that does not meet user requirements, according to an embodiment of the invention.
FIG. 7 is an interface diagram of a scatter plot fit and mean error meeting user requirements, according to an embodiment of the invention.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems consistent with certain aspects of the invention, as detailed in the appended claims.
Principle of BP neural network
Taking a low spatial resolution remote sensing index (independent variable) -low spatial resolution remote sensing earth surface temperature data (dependent variable) as a training sample; a BP (Back Propagation, BP) neural network model is constructed, the BP network is a feedforward network trained according to error inverse Propagation, a large number of mapping relations of input-output modes can be learned and stored, a mathematical equation for describing the mapping relations is not required to be disclosed in advance, the learning rule of the BP neural network model uses a gradient descent algorithm, and the weight and the threshold of the BP neural network are continuously adjusted through inverse Propagation, so that the error square sum of the BP neural network model is minimum.
Fig. 1 is a schematic diagram of a BP neural network model. The BP neural network model topological structure comprises an input layer, a hidden layer and an output layer. The algorithm selects Sigmoid as an activation function of a hidden layer, a linear function as an activation function of an output layer, a jth basic BP neuron (node) is shown in figure 1, and three most basic and most important functions of a biological neuron are simulated: weighting, summing and shifting. Wherein
Represents inputs from neurons 1,2, … …, n, respectively;
respectively representing the connection strength of the neuron 1,2, … …, n and the jth neuron, namely the weight; f (-) is the transfer function, SjIs the net input value, y, of the jth neuronjIs the output of the jth neuron; and adjusting the weight value by a gradient descent algorithm by comparing the difference between the output value and the true value. The specific calculation formula is as follows:
FIG. 2 is a flow chart of one embodiment of the present invention. Fig. 2 shows the method of the invention, comprising the following steps:
a data preparation step: acquiring low spatial resolution remote sensing index data, wherein the low spatial resolution remote sensing index data is a three-dimensional floating point type image matrix, namely height, width and layers (rows columns) of an image; height and width (rows columns) represent geographical location, each floor (bands) represents a telemetry index, x(m,n,b)Representing the size of a b-th remote sensing index value of a pixel x of the remote sensing index image at the position of m rows and n columns; acquiring low spatial resolution earth surface temperature remote sensing data, wherein the low spatial resolution earth surface temperature remote sensing data is a two-dimensional image matrix, namely the height and width (rows and columns) of an image correspond to a remote sensing index, and y is(m,n)Representing the earth surface temperature value of a pixel y of an earth surface temperature image at the position of m rows and n columns, wherein the unit is; acquiring high spatial resolution remote sensing index data, wherein the high spatial resolution remote sensing index data is a three-dimensional floating point type image matrix, namely height, width, and layers (rows columns) of an image; height and width (rows columns) represent geographical location, each floor (bands) represents a telemetry index, X(m,n,b)Representing the size of a b-th remote sensing index value of a pixel X of the remote sensing index image at the position of m rows and n columns;
self-defining parameters: the method comprises the following steps: selecting the number of remote sensing indexes in the low spatial resolution remote sensing index data (wave band selection), and setting the following parameters: the method comprises the steps of learning rate, the number of neurons in a hidden layer, maximum iterative learning times, a convergence threshold value, proportion distribution of training samples and verification samples, and a dynamic display period (breakpoint query) in a learning process.
Constructing a training sample set: according to the corresponding relation of height-width (rows) of an image matrix, x and y on corresponding positions are respectively selected to construct feature vectors, the low-spatial-resolution remote sensing index data is used as an input independent variable, the low-spatial-resolution surface temperature remote sensing data is used as an input dependent variable, and the feature vectors are obtained
Iteration step: after completing one-time learning of all the characteristic vectors in the step of constructing the training sample set, calculating an average error, judging whether the average error is smaller than a convergence threshold or whether the iteration times are larger than the maximum iteration learning times, and completing the step of training the sample set if the average error is smaller than the convergence threshold or the iteration times are larger than the maximum iteration learning times; otherwise, adjusting the output layer weight and the hidden layer weight by adopting an accumulated error BP algorithm, starting the next iteration step after the weight adjustment is finished, and repeating the iteration calculation step.
A parameter adjusting step A: in the iteration step, after the set number of times of the dynamic display period of the learning process is finished every iteration, the visual scatter diagram is displayed and output in real time, a user can check the fitting precision of the scatter diagram and the average loss of the output, whether parameter measurement is reasonable or not is judged in an auxiliary mode, if the average loss fluctuates up and down, the model learning rate is set to be too large, the gradient can oscillate back and forth near the minimum value, convergence cannot be achieved, and the learning rate is adjusted at the moment.
Training a model: training a BP neural network model by adopting the feature vector, wherein the BP neural network model comprises an input layer, a hidden layer and an output layer, and obtaining the trained BP neural network model;
and B, parameter adjustment, namely returning to the step of self-defining parameters to properly adjust the model parameters, including increasing the number of neurons in the hidden layer or the iteration times, until the model training result meets the requirements of the user, if the fitting of the scatter diagram and the average error do not meet the requirements after the step of training the model is finished.
High-resolution surface temperature simulation: and inputting the high spatial resolution remote sensing index data serving as an input independent variable into the trained BP neural network model to obtain high spatial resolution ground surface temperature downscaling calculation data.
And a result evaluation and output step: and evaluating the simulation result, and if obvious abnormal phenomena (the water body temperature is higher than the road and the like) occur, resetting and returning to the step of executing the self-defined parameters. And if the simulation result meets the requirement, outputting the simulation result. And finishing the ground surface temperature downscaling treatment.
Example (b):
(1) data preparation
In this example, a regular area of 870 m × 810 m in Wenchang city is used as a research area; the low spatial resolution remote sensing index data and the low spatial resolution earth surface temperature remote sensing data are derived from Landsat satellite image inversion results (the spatial resolution is 30 meters, wherein the remote sensing index is obtained by correcting an original image into earth surface reflectivity through atmosphere and then performing simple inter-band operation); the high spatial resolution remote sensing index data is derived from a WorldView-2 satellite image inversion result (the spatial resolution is 2 meters and is obtained through simple inter-band operation). FIG. 3A is a WorldView-2 color image of one embodiment of the present invention. Fig. 3B is a surface temperature down-scale inversion diagram corresponding to fig. 3A.
The low spatial resolution remote sensing index data is a three-dimensional floating point type image matrix, namely height, width and layers (rows, columns, banks) of the image; height and width (rows columns) represent geographical location, each floor (bands) represents a telemetry index, x(m,n,b)And (2) representing the value of the b-th remote sensing index DN (digital number) of the pixel x of the remote sensing index image at the position of m rows and n columns.
The low spatial resolution earth surface temperature remote sensing data is a two-dimensional image matrix, namely the height and width (rows and columns) of the image and the remote sensing indexCorresponding to, y(m,n)And the value size of the DN (Digital Number) value of the earth surface temperature of the pixel y of the earth surface temperature image at the position of m rows and n columns is represented, and the unit is C.
The high spatial resolution remote sensing index data is a three-dimensional floating point type image matrix, namely height, width and layers (rows, columns, banks) of the image; height and width (rows columns) represent geographical location, each floor (bands) represents a telemetry index, X(m,n,b)And (2) representing the value of the b-th remote sensing index DN (digital number) of the pixel X of the remote sensing index image at the position of m rows and n columns.
According to the characteristics of vegetation and water body coverage as main parts in a research area and the existing remote sensing data condition, the embodiment utilizes the existing Landsat and Worldview satellite images to calculate 14 remote sensing indexes which are sensitive to vegetation, and the detailed remote sensing index introduction (users can construct different types and different quantities of remote sensing indexes according to the characteristics of the research area and the existing data types) and the calculation formula are as shown in the following table 1.
TABLE 1
ρRED、ρGREEN、ρBLUE、ρNIR、ρCOASTALRespectively representing DN (surface reflectance value) of the corresponding waveband of the Landsat satellite image subjected to the atmospheric correction processing, wherein the DN is high-resolution Worldview satellite image and the low-resolution Landsat satellite image are respectively low-resolution. Such as rhoREDRepresenting the surface reflectance values of the red wave band. By the calculation formula in the table, the high-resolution remote sensing index image and the low-resolution remote sensing index image can be obtained and used as data input.
(2) Data entry
Inputting low-spatial-resolution remote sensing index data as an input independent variable, wherein the size of an input low-spatial-resolution remote sensing index image matrix is (27,29 and 14), the number of pixels is 783, and the number of remote sensing indexes (bands) is 14. The low spatial resolution earth surface temperature data is input as an input dependent variable, and the matrix size of the low spatial resolution earth surface temperature image is (27, 29).
The high spatial resolution remote sensing index image matrix size is (405,435,14), and the difference of the high spatial resolution remote sensing index image matrix size and the low spatial resolution remote sensing index image matrix size is 15 times. The expected result is high resolution surface temperature data for the same area with an image matrix size of (405,435, 14).
(3) Custom parameters
The user can custom select the number and combination of remote sensing indices actually used to construct the feature vector. Similarly, the training result of the BP neural network is also related to the learning rate, the number of neurons in a hidden layer, the maximum iterative learning times, the convergence threshold value, the proportion distribution of training samples and verification samples and the like, and compared with the traditional BP neural network model, the BP neural network model adopts default parameter setting or independently generates a few adjustable parameters. In the embodiment, all the elements are independent, and a user can perform initial setting and subsequent experimental adjustment according to the characteristics of the research area of the user (for example, in an area with strong heterogeneity, namely, a mixed pixel occupies a main part, the number of neurons in a hidden layer and the maximum iterative learning frequency are set to be large). For the convenience of initial setting by the user, table 2 gives explanations of each initial parameter and suggested initial setting values.
TABLE 2
According to the embodiment, after data input is completed, the remote sensing index is customized firstly, the first index (atmospheric impedance vegetation index) and the sixth index (improved nonlinear index) are excluded, the rest 12 indexes are selected to participate in training, and the remote sensing index with the spatial resolution of 30 meters is combined
And 30 m spatial resolutionSurface temperature of
Constructing feature vectors
Meanwhile, user-defined parameters including learning rate, the number of hidden layer network nodes, the maximum learning times, a convergence threshold value, proportion distribution of training samples and verification samples and a dynamic display period of a learning process are input by a user on the interactive interface. The present embodiments are set to 0.0001, 20, 100000, 0.001, 1, and 200, respectively.
(4) Model training
After the user-defined parameters are set, a feature vector training model is adopted, taking the ith input node as an example, and the initial weight matrix from the ith input node to the hidden layer is
Wherein
Length 13 (and constructed feature vector)
Is the number 12 of the remote sensing indexes plus one earth surface temperature value), and the magnitude of the net input value of the hidden layer
(dimension: 20 x 1). The weight matrix from the hidden layer to the output layer is Wi2＝[w1,w2,…w19,w20]，wiIs a scalar quantity. The output layer value is Out-Wi2Sigmoid (hidden). Wherein sigmoid, Hidden are the commonly used gradient descent function and Hidden layer net input values, respectively. Loss of error
FIG. 4 is a model for training a BP neural network according to an embodiment of the present inventionThe interface diagram of (1).
(5) Iterative computation
When one-time learning of all the feature vectors is completed, the average error is calculated
Judging whether the average error is less than 0.0001 or the iteration frequency is greater than 100000, if the average error is less than 0.0001 or the iteration frequency is greater than 100000, finishing the model training, otherwise, adopting the accumulated error BP algorithm to adjust the weight of the output layer
And hidden layer weight
And starting the next iteration after the weight value is adjusted, and repeating the iterative calculation step.
(6) Parameter adjustment
After each iteration is carried out for 200 times, the visual scatter diagram is displayed and output in real time, and a user can check the fitting precision and the output average loss (loss) of the scatter diagram and assist in judging whether the parameter measurement is reasonable or not. If the average loss fluctuates up and down, it means that the model learning rate is set too large, and the gradient may oscillate around the minimum value and cannot converge, and at this time, the learning rate should be adjusted.
And if the fitting of the scatter diagram and the average error do not meet the requirements after the model training is finished, returning to the step of self-defining parameters, and carrying out appropriate adjustment on the model parameters (increasing the number of neurons in the hidden layer or the number of iterations and the like). Until the model training result meets the user requirement.
(7) High resolution surface temperature simulation
After the model training is finished, inputting high-resolution remote sensing index data, starting high-resolution surface temperature simulation, wherein the size of a simulation result image matrix is (405,435,14), the simulation result image matrix is displayed in a drawing area on the right side of the model, and a high-value area is brighter and is usually a road, a bare land and the like; the low value area is darker, and the water body is usually used. FIG. 5 is an interface diagram of a model implementation to obtain high spatial resolution surface temperature downscaling calculation data, in accordance with one embodiment of the present invention.
(8) Result evaluation and output
And evaluating the simulation result, and if obvious abnormal phenomena (the water body temperature is higher than the road and the like) occur, resetting and returning to the step of executing the self-defined parameters. And if the simulation result meets the requirement, outputting the simulation result. And finishing the ground surface temperature downscaling treatment.
The earth surface temperature downscaling method based on machine learning constructs a network structure of a plurality of spectral indexes and earth surface temperature (LST) by applying a wide error back propagation BP neural network model, and trains on a pixel level to obtain a reliable relation between the LST and the spectral indexes by utilizing the characteristics of self-adaption and self-learning of the BP neural network.
The features and benefits of the present invention are illustrated by reference to the examples. Accordingly, the invention is expressly not limited to these exemplary embodiments illustrating some possible non-limiting combination of features which may be present alone or in other combinations of features.
The above-mentioned embodiments are only specific embodiments of the present invention, which are used for illustrating the technical solutions of the present invention and not for limiting the same, and the protection scope of the present invention is not limited thereto, although the present invention is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: any person skilled in the art can modify or easily conceive the technical solutions described in the foregoing embodiments or equivalent substitutes for some technical features within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the embodiments of the present invention, and they should be construed as being included therein. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (8)
1. A ground surface temperature downscaling method based on machine learning comprises the following steps:
step 1: obtaining low spatial resolutionRemote sensing index data, wherein the low spatial resolution remote sensing index data is a three-dimensional floating point type image matrix, namely height, width, and layers (rows, columns, banks) of an image; height and width (rows columns) represent geographical location, each floor (bands) represents a telemetry index, x(m,n,b)Representing a b-th remote sensing index value of a pixel x of the remote sensing index image at the position of m rows and n columns;
acquiring low spatial resolution earth surface temperature remote sensing data, wherein the low spatial resolution earth surface temperature remote sensing data is a two-dimensional image matrix, namely the height and width (rows and columns) of an image correspond to a remote sensing index, and y is(m,n)Representing the surface temperature value of a pixel y of the surface temperature image at the position of m rows and n columns;
acquiring high spatial resolution remote sensing index data, wherein the high spatial resolution remote sensing index data is a three-dimensional floating point type image matrix, namely height, width and layers (rows columns) of an image; height and width (rows columns) represent geographical location, each floor (bands) represents a telemetry index, X(m,n,b)Representing a b-th remote sensing index value of a pixel X of the remote sensing index image at the position of m rows and n columns;
step 2: constructing a training sample set, respectively selecting x and y at corresponding positions according to the corresponding relation of height-width (rows) of an image matrix to construct a feature vector, taking the low-spatial-resolution remote sensing index data as an input independent variable, taking the low-spatial-resolution earth surface temperature remote sensing data as an input dependent variable, and obtaining the feature vector
And step 3: training a BP neural network model by adopting the feature vector, wherein the BP neural network model comprises an input layer, a hidden layer and an output layer, and obtaining the trained BP neural network model;
and 4, step 4: and inputting the high spatial resolution remote sensing index data serving as an input independent variable into the trained BP neural network model to obtain high spatial resolution ground surface temperature downscaling calculation data.
2. The method of claim 1, further comprising a parameter customization step between step 1 and step 2, the parameter customization step comprising: and selecting the number of the adopted remote sensing indexes in the low spatial resolution remote sensing index data.
3. The method of claim 2, wherein the step of customizing parameters further comprises setting the following parameters: the method comprises the steps of learning rate, the number of neurons in a hidden layer, maximum iterative learning times, a convergence threshold value, proportion distribution of training samples and verification samples and a dynamic display period of a learning process.
4. The method according to claim 3, further comprising an iteration step, which includes calculating an average error after completing one-time learning of all the feature vectors in step 2, determining whether the average error is smaller than a convergence threshold, or whether the number of iterations is greater than a maximum number of iterative learning, and completing step 2 if the average error is smaller than the convergence threshold or the number of iterations is greater than the maximum number of iterative learning; otherwise, adjusting the output layer weight and the hidden layer weight by adopting an accumulated error BP algorithm, starting the next iteration step after the weight adjustment is finished, and repeating the iteration step.
5. The method according to claim 4, further comprising a parameter adjusting step, wherein in the iteration step, after the set number of times of the dynamic display period of the learning process is finished each time, the visualized scatter diagram is displayed and output in real time, a user can check the fitting precision of the scatter diagram and the average loss of the output, and assist in judging whether the parameter measurement is reasonable, if the average loss fluctuates up and down, which indicates that the model learning rate is set too large, the gradient may oscillate around the minimum value, and the gradient cannot converge, the learning rate is adjusted.
6. The method according to claim 5, wherein in the parameter adjusting step, if the fitting of the scatter diagram and the average error do not meet the requirements after the model training in step 3, the step of self-defining parameters is returned, and the appropriate adjustment of the model parameters is performed, including increasing the number of neurons in the hidden layer or the number of iterations, until the model training result meets the requirements of the user.
7. The method of claim 1, wherein: the number of said telemetric indices is 14.
8. The method of claim 3, wherein: the parameter settings are as follows: the method comprises the following steps of counting the number of remote sensing indexes by 12, counting the learning rate by 0.0001, counting the hidden layer neurons by 20, counting the maximum iterative learning time by 100000, counting the convergence threshold value by 0.001, distributing the proportion of training samples and verification samples by 1, and dynamically displaying the period by 200 in the learning process.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911141709.2A CN110866364A (en) | 2019-11-20 | 2019-11-20 | Ground surface temperature downscaling method based on machine learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911141709.2A CN110866364A (en) | 2019-11-20 | 2019-11-20 | Ground surface temperature downscaling method based on machine learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110866364A true CN110866364A (en) | 2020-03-06 |
Family
ID=69656050
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911141709.2A Pending CN110866364A (en) | 2019-11-20 | 2019-11-20 | Ground surface temperature downscaling method based on machine learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110866364A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112115646A (en) * | 2020-09-22 | 2020-12-22 | 北京清大华亿科技有限公司 | Oil refining chemical production process simulation method and system |
| CN113486000A (en) * | 2021-09-08 | 2021-10-08 | 中国测绘科学研究院 | Surface evapotranspiration data downscaling method based on multi-source data and deep learning |
| CN114414090A (en) * | 2021-12-14 | 2022-04-29 | 厦门大学 | Surface temperature prediction method and system based on remote sensing image and multilayer sensing |
| CN114863168A (en) * | 2022-04-27 | 2022-08-05 | 国家基础地理信息中心 | Remote sensing image classification method, device and equipment of self-adaptive recursive incremental model |
| CN117148350A (en) * | 2023-10-27 | 2023-12-01 | 航天宏图信息技术股份有限公司 | SAR image-based earth surface temperature space downscaling method, device and equipment |
| CN117172131A (en) * | 2023-10-12 | 2023-12-05 | 中国科学院地理科学与资源研究所 | Method and system for predicting high-resolution underground water level based on artificial neural network |
| CN118094395A (en) * | 2024-04-22 | 2024-05-28 | 安徽农业大学 | Surface temperature estimation method, system and computer equipment |
| WO2024113427A1 (en) * | 2022-12-02 | 2024-06-06 | 深圳先进技术研究院 | Land surface temperature remote sensing product downscaling method and system, device, and medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107423537A (en) * | 2017-01-25 | 2017-12-01 | 河海大学 | A kind of method of the surface temperature NO emissions reduction based on adaptive threshold |
| CN109271605A (en) * | 2018-10-12 | 2019-01-25 | 中国科学院地理科学与资源研究所 | High spatial resolution remote sense surface temperature method for computing data and device |
| CN109472393A (en) * | 2018-09-30 | 2019-03-15 | 广州地理研究所 | Space NO emissions reduction precipitation data detection method, device and electronic equipment |
| CN109635309A (en) * | 2018-10-17 | 2019-04-16 | 广州地理研究所 | A kind of surface temperature space NO emissions reduction method |
| CN109885959A (en) * | 2019-03-05 | 2019-06-14 | 中国科学院地理科学与资源研究所 | A kind of surface temperature robust NO emissions reduction method |
| CN110427836A (en) * | 2019-07-11 | 2019-11-08 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | A kind of high-resolution remote sensing image Clean water withdraw method based on multi_dimension optimization |
-
2019
- 2019-11-20 CN CN201911141709.2A patent/CN110866364A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107423537A (en) * | 2017-01-25 | 2017-12-01 | 河海大学 | A kind of method of the surface temperature NO emissions reduction based on adaptive threshold |
| CN109472393A (en) * | 2018-09-30 | 2019-03-15 | 广州地理研究所 | Space NO emissions reduction precipitation data detection method, device and electronic equipment |
| CN109271605A (en) * | 2018-10-12 | 2019-01-25 | 中国科学院地理科学与资源研究所 | High spatial resolution remote sense surface temperature method for computing data and device |
| CN109635309A (en) * | 2018-10-17 | 2019-04-16 | 广州地理研究所 | A kind of surface temperature space NO emissions reduction method |
| CN109885959A (en) * | 2019-03-05 | 2019-06-14 | 中国科学院地理科学与资源研究所 | A kind of surface temperature robust NO emissions reduction method |
| CN110427836A (en) * | 2019-07-11 | 2019-11-08 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | A kind of high-resolution remote sensing image Clean water withdraw method based on multi_dimension optimization |
Non-Patent Citations (2)
| Title |
|---|
| YANG G 等: ". A novel method to estimate sub pixel temperature by", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》 * |
| 汪子豪 等: "基于BP神经网络的地表温度空间降尺度方法", 《遥感技术与应用》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112115646A (en) * | 2020-09-22 | 2020-12-22 | 北京清大华亿科技有限公司 | Oil refining chemical production process simulation method and system |
| CN113486000A (en) * | 2021-09-08 | 2021-10-08 | 中国测绘科学研究院 | Surface evapotranspiration data downscaling method based on multi-source data and deep learning |
| CN113486000B (en) * | 2021-09-08 | 2021-12-17 | 中国测绘科学研究院 | Surface evapotranspiration data downscaling method based on multi-source data and deep learning |
| CN114414090A (en) * | 2021-12-14 | 2022-04-29 | 厦门大学 | Surface temperature prediction method and system based on remote sensing image and multilayer sensing |
| CN114414090B (en) * | 2021-12-14 | 2023-04-28 | 厦门大学 | Surface temperature prediction method and system based on remote sensing images and multilayer sensing |
| CN114863168A (en) * | 2022-04-27 | 2022-08-05 | 国家基础地理信息中心 | Remote sensing image classification method, device and equipment of self-adaptive recursive incremental model |
| WO2024113427A1 (en) * | 2022-12-02 | 2024-06-06 | 深圳先进技术研究院 | Land surface temperature remote sensing product downscaling method and system, device, and medium |
| CN117172131A (en) * | 2023-10-12 | 2023-12-05 | 中国科学院地理科学与资源研究所 | Method and system for predicting high-resolution underground water level based on artificial neural network |
| CN117172131B (en) * | 2023-10-12 | 2024-05-14 | 中国科学院地理科学与资源研究所 | Method and system for predicting high-resolution underground water level based on artificial neural network |
| CN117148350A (en) * | 2023-10-27 | 2023-12-01 | 航天宏图信息技术股份有限公司 | SAR image-based earth surface temperature space downscaling method, device and equipment |
| CN117148350B (en) * | 2023-10-27 | 2024-01-30 | 航天宏图信息技术股份有限公司 | SAR image-based earth surface temperature space downscaling method, device and equipment |
| CN118094395A (en) * | 2024-04-22 | 2024-05-28 | 安徽农业大学 | Surface temperature estimation method, system and computer equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110866364A (en) | Ground surface temperature downscaling method based on machine learning | |
| CN110738252B (en) | Space autocorrelation machine learning satellite precipitation data downscaling method and system | |
| Camps-Valls et al. | Physics-aware Gaussian processes in remote sensing | |
| CN107909607B (en) | A kind of year regional vegetation coverage computational methods | |
| Roodposhti et al. | Towards automatic calibration of neighbourhood influence in cellular automata land-use models | |
| CN108416805B (en) | Intrinsic image decomposition method and device based on deep learning | |
| CN107748736A (en) | A kind of multiple-factor Remote Sensing temperature space NO emissions reduction method based on random forest | |
| CN107423537B (en) | Surface temperature downscaling method based on self-adaptive threshold | |
| CN109885959B (en) | A kind of surface temperature robust NO emissions reduction method | |
| CN106779067B (en) | Soil moisture method for reconstructing and system based on multi- source Remote Sensing Data data | |
| CN106803256A (en) | A kind of 3D shape based on projection convolutional network is split and semantic marker method | |
| CN109685743A (en) | Image mixed noise removing method based on noise learning neural network model | |
| CN107590515A (en) | The hyperspectral image classification method of self-encoding encoder based on entropy rate super-pixel segmentation | |
| CN101482929B (en) | Remote-sensing image processing method and system | |
| Jiang et al. | Application of back propagation neural network in the classification of high resolution remote sensing image: take remote sensing image of Beijing for instance | |
| CN109033543B (en) | Method, device and equipment for estimating vegetation coverage of heterogeneous land areas | |
| CN104050643A (en) | Remote-sensing image relative correction method and system integrating geometry and radiation | |
| CN110503137B (en) | Determination method of remote sensing image space-time fusion basic image pair based on cross fusion | |
| CN106897986A (en) | A kind of visible images based on multiscale analysis and far infrared image interfusion method | |
| CN109671019A (en) | A kind of remote sensing image sub-pixed mapping drafting method based on multi-objective optimization algorithm and sparse expression | |
| CN113256733B (en) | Camera spectral sensitivity reconstruction method based on confidence voting convolutional neural network | |
| CN106599548A (en) | Spatial and temporal scale matching method and device for land surface evapotranspiration remote sensing estimation | |
| CN110991230B (en) | Method and system for detecting ship by using remote sensing images in any direction based on rotation candidate frame | |
| CN108335263A (en) | A kind of microwave remote sensing image super-resolution rebuilding method based on VDSR | |
| CN112907441A (en) | Space downscaling method based on super-resolution of underground water satellite image |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200306 |
|
| RJ01 | Rejection of invention patent application after publication |