Crop classification method of time series polarimetric SAR based on temporal convolutional neural network

Abstract

The present invention belongs to the technical field of radar remote sensing image processing and crop classification, and provides a time series polarimetric SAR crop classification method based on a time convolutional neural network, including obtaining a time series polarimetric SAR data set of a target area; preprocessing of time series polarimetric SAR data and crop growth monitoring and analysis; construction of a time convolutional neural network for time series polarimetric SAR data, extracting time context information of time series polarimetric SAR data through one-dimensional convolution, and generating high-order time series features; hyperparameter adjustment in a TCN network and crop classification result mapping, adjusting the hyperparameters in the TCN network through a control variable method; geocoding the crop classification results in a SAR coordinate system, and outputting the final classification results in a geographic coordinate system through mapping. The present invention combines time series information and polarimetric features, gives full play to the time information and polarimetric feature information of polarimetric SAR data, and greatly improves the classification accuracy of crops.

Description

Technical Field

[0001] The invention belongs to the technical field of radar remote sensing image processing and crop classification, and in particular relates to a temporal polarimetric SAR crop classification method based on a temporal convolutional neural network. Background Art

[0002] Crop classification and mapping products provide important data support for the digital and intelligent development of agriculture. Remote sensing technology has become the mainstream technical means to achieve large-scale crop classification and mapping. The microwave signals emitted by the synthetic aperture radar (SAR) platform have strong penetration capabilities and can monitor crop growth all day and all the time. The introduction of multi-polarization information further increases the signal's sensitivity to crop morphological structure, orientation distribution, and dielectric properties, making polarimetric SAR an effective means of crop classification and monitoring.

[0003] There are two main existing methods for crop classification based on polarimetric SAR:

[0004] 1. Polarimetric SAR classification method based on shallow machine learning algorithm.

[0005] Classic polarimetric SAR classification stacks scattering features into multi-dimensional feature vectors and uses shallow machine learning algorithms for clustering. For example, the H / A / α-Wishart and Freeman-Wishart maximum likelihood classifiers based on complex Wishart distribution use the classification plane for initial division and then use the Wishart distribution distance for iterative clustering. Later, relevant scholars introduced support vector machine (SVM) and random forest (RF) algorithms to construct nonlinear separation planes between different categories and improve the adaptive ability of classification for complex scenes and multiple categories. For example, Shuai et al. used SVM classifiers to identify corn based on time-series Radarsat-2 (referring to the observation satellite Radarsat-2) polarimetric SAR data, and the classification effect was significant; Waske used random forest classifiers to classify multi-phase and multi-frequency SAR data, using the traditional Gini coefficient as the basis for feature importance judgment. The results showed that the classification effect of random forests was efficient. Although the application of the above shallow machine learning algorithms is simple and efficient, it is difficult to autonomously learn spatial structure information and overcome the interference of speckle noise on crop classification.

[0006] 2. Polarimetric SAR classification method based on deep learning.

[0007] The development of deep learning models provides ideas for mining spatial structural information. Convolutional neural networks (CNNs) can autonomously learn spatial structures through convolution operations and condense them into high-order features, which promotes the clustering ability of polarimetric SAR classifiers for homogeneous pixels within a class. For example, Zhou et al. built a polarimetric SAR classifier based on the CNN model and used AIRSAR PolSAR data (data collected by an airborne polarimetric synthetic aperture radar system) over Flevoland Province in the Netherlands to classify buildings, with an overall classification accuracy of 93.38%. Faced with the demand for high-precision classification and massive polarimetric SAR data and samples, CNN models are constantly optimizing model structures and parameters from network deepening and model lightweighting. Improved network models such as VGGNet (Visual Geometry Group Network), ResNet (Residual Network), U-Net (U-Net) and LinkNet (Link Network) and lightweight models such as SqueezeNet model (Squeeze Network) and MobileNet (Mobile Network) are widely used in polarimetric SAR classification. For example, Bem et al. used two deep learning models, U-net and LinkNet, to extract rice features from multi-temporal SAR data. The results showed that they can better capture the spatiotemporal information of rice-growing areas.

[0008] The above classification methods have achieved good performance in specific fields, but these methods rarely explore the temporal context information of polarimetric SAR data and do not explore the phenological evolution laws of crops. Therefore, there is still much room for improvement in the identification of crop types. Summary of the invention

[0009] The present invention provides a time series polarimetric SAR crop classification method based on a time convolutional neural network. The method combines time series information and polarimetric features, gives full play to the time information and polarimetric feature information of polarimetric SAR data, and improves the accuracy of crop classification. The method specifically includes the following steps:

[0010] S1, obtain the time series polarimetric SAR data set of the target area;

[0011] S2. Time-series polarimetric SAR data preprocessing and crop growth monitoring analysis: Generate a time-series polarimetric coherence matrix based on the time-series polarimetric SAR data set T 3 , using refined Lee filtering, and using H / A / α decomposition and Freeman decomposition methods to extract a time series polarization feature set, analyzing the correlation between the time-varying characteristic curves of crops and the changes in biophysical parameters based on the time series polarization feature set, and studying the ability of time series polarization features to distinguish different crop types;

[0012] S3. Construction of a temporal convolutional neural network for time-series polarimetric SAR data: Using the temporal polarimetric feature set extracted in step S2 as training samples, set the basic structure of the temporal convolutional neural network, extract the temporal context information of the time-series polarimetric SAR data through one-dimensional convolution, and generate high-order temporal features; mine and characterize the temporal scattering laws of different types of crops, construct a TCN (temporal convolutional network) model, and use multiple residual modules to connect and build a TCN network structure;

[0013] S4, hyperparameter adjustment in TCN network and crop classification result mapping: Based on the TCN network structure built in step S3, the hyperparameters in the TCN network are adjusted by the control variable method to obtain the optimal model classification parameters and complete the crop classification;

[0014] S5. Geocoding the crop classification results in the SAR coordinate system, and outputting the final classification results in the geographic coordinate system through mapping.

[0015] In a specific implementation manner, in step S2, the time series polarization SAR data preprocessing and crop growth monitoring analysis are specifically as follows: k p After multiplying the conjugate transpose of T 3 matrix:

[0016] ;

[0017] In formula 1), represents conjugate transpose, and < > represents multi-view averaging.

[0018] In a specific implementation manner, in step S2, T 3 The specific steps of the refined Lee filter processing of the matrix are as follows: based on the filter windows in nine different directions, the mean of the temporal polarization coherence matrix in the window is calculated; paired with the gradient edge operators in the horizontal, vertical, 45° and 135° directions, the operator with the largest absolute value is taken as the edge intensity, and the corresponding gradient direction is the edge direction; then, for each element in the filter window, the total intensity map is used to calculate the weight, and the weighted average of the pixels in the neighborhood is calculated to replace the original value to complete the filtering; the temporal polarization coherence matrix after filtering is:

[0019] ;

[0020] In formula 2), represents the filtered time-series polarization coherence matrix, the matrix represents the local averaged temporal polarimetric coherence matrix, T(i) Indicates the original T3 matrix, b Represents the weight factor.

[0021] In a specific implementation, in step S2, the specific operation of extracting the time series polarization feature set using the H / A / α decomposition and Freeman decomposition method is as follows: from the filtered time series polarization coherence matrix Extract intensity features T 11 , strength characteristics T 22 , strength characteristics T 33 The filtered time-series polarization coherence matrix is ​​processed using H / A / α decomposition and Freeman decomposition to obtain the scattering entropy H, anisotropy A, scattering angle α, and volume scattering intensity P vol , secondary scattering intensity P dbl and surface scattering intensity P odd Features, and then normalize these features to obtain the normalized time series feature set; perform linear interpolation of the time series feature set at intervals of 1 day, and use the SG filter (a smoothing filter based on local polynomial fitting) to perform time series filtering to construct a time-varying scattering characteristic curve, where the basic formula of the SG filter is:

[0022] ;

[0023] In formula 3), Y is the original eigenvalue, Y * is the eigenvalue after filtering, i Indicates the phase number within the time window, m represents the time window radius, C i is the filter window i The coefficients corresponding to the eigenvalues, index j Represents the running index of the ordinate data in the original data table, N is the number of elements in the filtering window; based on the constructed time series scattering characteristic curve, the JM distance (Jeffrey-Matsuhita distance) is used to calculate the discrimination between different time series features for crop categories, and obtain the separation ability of time series polarization features for different crop types.

[0024] In a specific embodiment, in step S2, the intensity feature T 11 , strength characteristics T 22 , strength characteristics T 33, scattering entropy H, anisotropy A, scattering angle α, volume scattering intensity P vol , secondary scattering intensity P dbl and surface scattering intensity P odd The range of feature normalization is (0, 1].

[0025] In a specific implementation, in step S3, the construction of the temporal convolutional neural network for time-series polarimetric SAR data is specifically as follows: defining the core part of the TCN layer, the causal hole convolution, and using residual connection to solve the problem of gradient disappearance and explosion. The specific formula is:

[0026] ;

[0027] In formula 4), x represents the original input of this layer, f(x) Represents the convolution result of this layer, M Represents the output after residual connection; based on the result of residual connection, the structure of TCN layer is defined as a multi-layer residual structure. The residual module in each TCN layer includes a hole causal convolution layer, a SDP (Spatial Dropout 1D) layer and a one-dimensional convolution layer.

[0028] In a specific implementation, in step S3, based on the defined TCN layer, a network model consisting of an input layer, a TCN layer, a flattening layer and a fully connected layer is constructed; wherein the input layer defines the size of the input data, the TCN layer is the core of the entire network, the flattening layer expands the output of the TCN layer, and the fully connected layer uses the output of the flattening layer as input and maps it to a vector by using a softmax activation function to calculate the category membership of the extracted features.

[0029] In a specific implementation, in step S4, the hyperparameters in the TCN network are adjusted by the control variable method, and the method for obtaining the optimal model classification parameters is specifically as follows: through the control variable method, the convolution kernel window size, the number of convolution kernels, the number of feature types, the number of time phases, the number of TCN layers, the learning rate, the optimization algorithm, the loss function and the activation function hyperparameter setting control group are experimentally adjusted, and the overall accuracy of the experimental results under different control groups of the hyperparameter settings is calculated to obtain the optimal model classification parameters, and the classification parameters are finally applied to the constructed TCN network.

[0030] In a specific implementation, in step S5, the crop classification results in the SAR coordinate system are geocoded, and the specific steps of outputting the final classification results in the geographic coordinate system through mapping are: based on the length of the radar echo time and the echo Doppler characteristics, the initial position of the image point is calculated, and the initial lookup table of slant range pixels and geographic pixels is generated; based on the digital elevation model, the simulated SAR intensity image is calculated and matched with the real SAR intensity image to obtain the matching polynomial and generate a fine lookup table; image positioning is completed, and geographic coordinates are assigned to the pixels of each SAR crop classification result image, and finally the mapping and release of the crop classification results are realized.

[0031] The present invention has at least the following beneficial effects:

[0032] 1. By acquiring a time-series polarimetric SAR dataset of the target area, not only does it increase the amount of data available for analysis, it also enables the model to identify the unique scattering characteristics of crops at different growth stages.

[0033] 2. By generating a time-series polarization coherence matrix T 3 , refined Lee filtering processing, H / A / α decomposition and Freeman decomposition technology, deeply explore the rich information in the time series polarimetric SAR data set. The construction of the time series polarimetric feature set not only considers the information in the spatial and polarization dimensions, but also combines the change pattern in the time dimension. This fusion of multi-dimensional features helps to better characterize the unique properties of different types of crops and further improves the classification effect.

[0034] 3. The TCN network effectively extracts the temporal context information of the time-series polarimetric SAR data through one-dimensional convolution operations and generates high-order time-series features. This process not only captures the dependency between adjacent time points, but also can dig out deeper time-series scattering laws. Moreover, the use of the residual module enhances the learning ability of the network, enabling it to handle complex nonlinear relationships, thereby more accurately distinguishing different types of crops.

[0035] 4. By using the control variable method to fine-tune the key parameters in the TCN network, the best model configuration can be found to ensure the optimal classification performance, thereby ensuring the stability and reliability of the classification results.

[0036] In summary, this method combines time series information and polarization features to give full play to the time information and polarization feature information of time series polarimetric SAR data, greatly improving the classification accuracy of crops. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The present invention is a flowchart of a method for crop classification using time-series polarimetric SAR based on a temporal convolutional neural network.

[0038] Figure 2 This is the classification result diagram of Example 1. DETAILED DESCRIPTION

[0039] The drawings are only used for illustrative purposes and are not to be construed as limiting the present application.

[0040] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0041] Embodiment 1:

[0042] like Figure 1-2 As shown, taking the classification of time series Radarsat-2 data in the Flevoland experimental area of ​​the Netherlands as an example, the specific operation steps of the present invention are as follows:

[0043] S1. Obtain the time-series polarimetric SAR dataset of the target area.

[0044] S2. Time series polarimetric SAR data preprocessing and crop growth monitoring analysis:

[0045] By the scattering vector k p After multiplying the conjugate transpose of T 3 matrix:

[0046] ;

[0047] In formula 1), represents conjugate transpose, and <> represents multi-view averaging.

[0048] right T 3 The matrix is ​​refined by Lee filtering: Based on the filter windows in nine different directions, the mean of the temporal polarization coherence matrix in the window is calculated. Paired with the gradient edge operators in the horizontal, vertical, 45° and 135° directions, the operator with the largest absolute value is taken as the edge intensity, and the corresponding gradient direction is the edge direction. Then, for each element in the filter window, the total intensity map is used to calculate the weight, and the weighted average of the pixels in the neighborhood is calculated to replace the original value to complete the filtering. The filtered temporal polarization coherence matrix is:

[0049] ;

[0050] In formula 2), represents the time-series polarization coherence matrix after filtering, the matrix represents the local averaged temporal polarimetric coherence matrix, T(i) Indicates the original T 3 matrix, bRepresents the weight factor.

[0051] Extracting time-series polarimetric feature sets: From the filtered time-series polarimetric coherence matrix Extract intensity features T 11 , strength characteristics T 22 , strength characteristics T 33 , use H / A / α decomposition to process the filtered time-series polarization coherence matrix to obtain the scattering entropy H, anisotropy A and scattering angle α, and use Freeman decomposition to process the filtered time-series polarization coherence matrix to obtain the volume scattering intensity P vol , secondary scattering intensity P dbl and surface scattering intensity P odd Features, and then normalize these features to the range of (0, 1] to obtain the normalized time series feature set. Linear interpolation of the time series feature set is performed at intervals of 1 day, and the SG filter is used for time series filtering to construct the time-varying scattering characteristic curve, where the basic formula of the SG filter is:

[0052] ;

[0053] In formula 3), Y is the original eigenvalue, Y * is the eigenvalue after filtering, i Indicates the phase number within the time window, m represents the time window radius, C i is the filter window i The coefficients corresponding to the eigenvalues, index j Represents the running index of the ordinate data in the original data table, N is the number of elements in the filter window. Based on the constructed time series scattering characteristic curve, the JM distance is used to calculate the discrimination between different time series features for crop categories, and the separation ability of time series polarization features for different crop types is obtained.

[0054] The correlation between the time-varying characteristic curves of crops and the changes in biophysical parameters is analyzed through the time-series polarization feature set, so as to study the discriminative ability of time-series polarization features for different crop types.

[0055] S3. Construction of a temporal convolutional neural network for time-series polarimetric SAR data: Using the time-series polarimetric feature set extracted in step S2 as training samples, set the basic structure of the temporal convolutional neural network, extract the temporal context information of the time-series polarimetric SAR data through one-dimensional convolution, and generate high-order time-series features. Mining and characterizing the temporal scattering laws of different types of crops, constructing a TCN model, and using multiple residual modules to connect and build a TCN network structure. Specifically: define the core part of the TCN layer, the causal hole convolution, and use residual connections to solve the problems of gradient vanishing and explosion. The specific formula is:

[0056] ;

[0057] In formula 4), x represents the original input of this layer, f(x) Represents the convolution result of this layer, M Represents the output after residual connection. Based on the result of residual connection, the structure of TCN layer is defined as a multi-layer residual structure. The residual module in each TCN layer includes a hole causal convolution layer, an SDP layer and a one-dimensional convolution layer.

[0058] Based on the defined TCN layer, a network model consisting of an input layer, a TCN layer, a flattening layer, and a fully connected layer is constructed. The input layer defines the size of the input data, the TCN layer is the core of the entire network, the flattening layer expands the output of the TCN layer, and the fully connected layer uses the output of the flattening layer as input and maps it to a vector by using the softmax activation function to calculate the category membership of the extracted features.

[0059] S4. Hyperparameter adjustment in TCN network and crop classification result mapping: Based on the TCN network structure built in step S3, the convolution kernel window size (3×3, 5×5, 7×7), number of convolution kernels (32, 64, 128, 256), number of feature types (1, 3, 6, 9), number of phases (1, 3, 5, 7), number of TCN layers (1, 2, 3), learning rate (0.1, 0.01, 0.001, 0.0001, 0.00001), optimization algorithm (AdaDelta algorithm, Adagrad algorithm, Adam algorithm), loss function (cross entropy loss function, mean square error loss function) and so on were adjusted. The number of hyperparameters, sparse classification cross entropy loss function) and activation functions (Sigmoid activation function, Softmax activation function, ReLU activation function) were used to experimentally adjust the control groups with different hyperparameter settings through the control variable method. The overall accuracy of the experimental results under different control groups with hyperparameter settings was calculated, and the optimal model classification parameters were as follows: the convolution kernel window size was 3×3, the number of convolution kernels was 64, the number of feature types was 9, the number of time phases was 7, the number of TCN layers was 2, the learning rate was 0.001, the optimization algorithm was the Adam algorithm, the loss function was the sparse classification cross entropy loss function, and the activation function was the ReLU activation function.

[0060] These optimal classification parameters are finally applied to the constructed TCN network to complete crop classification.

[0061] S5. Geocode the crop classification results in the SAR coordinate system, and output the final classification results in the geographic coordinate system through mapping. The specific steps are: based on the length of the radar echo time and the echo Doppler characteristics, calculate the initial position of the image point, and generate the initial lookup table of slant range pixels and geographic pixels. Calculate the simulated SAR intensity image based on the digital elevation model (DEM), match it with the real SAR intensity image, obtain the matching polynomial and generate a fine lookup table. Complete image positioning, assign geographic coordinates to the pixels of each SAR crop classification result image, and finally achieve the following Figure 2 Mapping and publication of crop classification results shown.

[0062] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions and substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for crop classification based on temporal convolutional neural network using polarimetric SAR, characterized in that: The steps include: S1, obtain the time series polarimetric SAR data set of the target area; S2. Time-series polarimetric SAR data preprocessing and crop growth monitoring analysis: Generate a time-series polarimetric coherence matrix based on the time-series polarimetric SAR data set T 3 , using refined Lee filtering, and using H / A / α decomposition and Freeman decomposition methods to extract a time series polarization feature set, analyzing the correlation between the time-varying characteristic curves of crops and the changes in biophysical parameters based on the time series polarization feature set, and studying the ability of time series polarization features to distinguish different crop types; S3, constructing a temporal convolutional neural network for time series polarimetric SAR data: using the temporal polarimetric feature set extracted in step S2 as training samples, setting the basic structure of the temporal convolutional neural network, extracting the temporal context information of the time series polarimetric SAR data through one-dimensional convolution, and generating high-order temporal features; Mining and characterizing the temporal scattering laws of different types of crops, constructing a TCN model, and using multiple residual modules to connect and build a TCN network structure; S4, hyperparameter adjustment in TCN network and crop classification result mapping: Based on the TCN network structure built in step S3, the hyperparameters in the TCN network are adjusted by the control variable method to obtain the optimal model classification parameters and complete the crop classification; S5. Geocoding the crop classification results in the SAR coordinate system, and outputting the final classification results in the geographic coordinate system through mapping.

2. The method for crop classification based on time series polarimetric SAR using a temporal convolutional neural network according to claim 1, characterized in that: In step S2, the time series polarimetric SAR data preprocessing and crop growth monitoring analysis are specifically as follows: k p After multiplying the conjugate transpose of T 3 matrix: ； In formula 1), represents conjugate transpose, and < > represents multi-view averaging.

3. The method for crop classification based on time series polarimetric SAR using a temporal convolutional neural network according to claim 1, characterized in that: In step S2, T 3 The specific steps of the refined Lee filter processing of the matrix are as follows: based on the filter windows in nine different directions, the mean of the temporal polarization coherence matrix in the window is calculated; paired with the gradient edge operators in the horizontal, vertical, 45° and 135° directions, the operator with the largest absolute value is taken as the edge intensity, and the corresponding gradient direction is the edge direction; then, for each element in the filter window, the total intensity map is used to calculate the weight, and the weighted average of the pixels in the neighborhood is calculated to replace the original value to complete the filtering; the temporal polarization coherence matrix after filtering is: ； In formula 2), represents the filtered time-series polarization coherence matrix, the matrix represents the local averaged temporal polarimetric coherence matrix, T(i) Indicates the original T 3 matrix, b Represents the weight factor.

4. The method for crop classification based on time series polarimetric SAR using a temporal convolutional neural network according to claim 3 is characterized in that: In step S2, the specific operation of extracting the time series polarization feature set using the H / A / α decomposition and Freeman decomposition method is as follows: Extract intensity features T 11 , strength characteristics T 22 , strength characteristics T 33 The filtered time-series polarization coherence matrix is ​​processed using H / A / α decomposition and Freeman decomposition to obtain the scattering entropy H, anisotropy A, scattering angle α, and volume scattering intensity P vol , secondary scattering intensity P dbl and surface scattering intensity P odd Features, and then normalize these features to obtain the normalized time series feature set; perform linear interpolation of the time series feature set at intervals of 1 day, and use the SG filter to perform time series filtering to construct a time-varying scattering characteristic curve, where the basic formula of the SG filter is: ； In formula 3), Y is the original eigenvalue, Y * is the eigenvalue after filtering, i Indicates the phase number within the time window, m represents the time window radius, C i is the filter window i The coefficients corresponding to the eigenvalues, index j Represents the running index of the ordinate data in the original data table, N is the number of elements in the filter window; Based on the constructed time series scattering characteristic curve, the JM distance is used to calculate the discrimination between crop categories with different time series features, and the separation ability of time series polarization features for different crop types is obtained.

5. The method for crop classification based on time series polarimetric SAR using a temporal convolutional neural network according to claim 4 is characterized in that: In step S2, the intensity feature T 11 , strength characteristics T 22 , strength characteristics T 33 , scattering entropy H, anisotropy A, scattering angle α, volume scattering intensity P vol , secondary scattering intensity P dbl and surface scattering intensity P odd The range of feature normalization is (0, 1].

6. The method for crop classification based on time series polarimetric SAR using a temporal convolutional neural network according to claim 1, characterized in that: In step S3, the construction of the temporal convolutional neural network for time-series polarimetric SAR data is as follows: the core part of the TCN layer, the causal hole convolution, is defined, and the residual connection is used to solve the problem of gradient disappearance and explosion. The specific formula is: ； In formula 4), x represents the original input of this layer, f(x) Represents the convolution result of this layer, M Represents the output after residual connection; based on the result of residual connection, the structure of TCN layer is defined as a multi-layer residual structure, and the residual module in each TCN layer includes a hole causal convolution layer, an SDP layer and a one-dimensional convolution layer.

7. The method for crop classification based on time series polarimetric SAR using a temporal convolutional neural network according to claim 6, characterized in that: In step S3, based on the defined TCN layer, a network model consisting of an input layer, a TCN layer, a flattening layer and a fully connected layer is constructed; wherein the input layer defines the size of the input data, the TCN layer is the core of the entire network, the flattening layer expands the output of the TCN layer, and the fully connected layer takes the output of the flattening layer as input and maps it to a vector by using a softmax activation function to calculate the category membership of the extracted features.

8. The method for crop classification based on time series polarimetric SAR using a temporal convolutional neural network according to claim 1, characterized in that: In step S4, the hyperparameters in the TCN network are adjusted by the control variable method to obtain the optimal model classification parameters. Specifically, the method is as follows: through the control variable method, the convolution kernel window size, the number of convolution kernels, the number of feature types, the number of time phases, the number of TCN layers, the learning rate, the optimization algorithm, the loss function and the activation function hyperparameter setting control group are experimentally adjusted, and the overall accuracy of the experimental results under different control groups of the hyperparameter settings is calculated to obtain the optimal model classification parameters, and the classification parameters are finally applied to the constructed TCN network.

9. The method for crop classification using time series polarimetric SAR based on a temporal convolutional neural network according to any one of claims 1 to 8, characterized in that: In step S5, the crop classification results in the SAR coordinate system are geocoded, and the specific steps of outputting the final classification results in the geographic coordinate system through mapping are: based on the radar echo time and echo Doppler characteristics, the initial position of the image point is calculated, and the initial lookup table of slant range pixels and geographic pixels is generated; Based on the digital elevation model, simulated SAR intensity images are calculated and matched with real SAR intensity images to obtain matching polynomials and generate a refined lookup table; Complete image positioning, assign geographic coordinates to each pixel of the SAR crop classification result image, and ultimately realize the mapping and publication of crop classification results.

Images