Two-channel CNN crop classification system, methods, storage media and electronic devices

Abstract

This invention discloses a dual-channel CNN method, system, storage medium, and electronic device for crop classification. The method includes preprocessing Sentinel-1 images to obtain backscattering maps and a polarization matrix C2 in VV / VH dual-polarization mode; converting the C2 matrix into Stokes vectors and performing polarization decomposition to generate maps of different polarization components; preprocessing Sentinel-2 images to obtain multispectral images; creating training, testing, and label sets; constructing a dual-channel convolutional neural network to extract deep features and fusing features from multi-source data to obtain a new feature set with richer information; inputting the created sample set into the constructed network model and outputting and saving the optimal trained model. This invention can accurately classify crops, providing basic data for crop growth monitoring, yield assessment, and other related work. Furthermore, this invention has good performance and is easy to implement in engineering.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, and in particular to a dual-channel CNN crop classification system, method, storage medium, and electronic device. Background Technology

[0002] Currently, remote sensing technology, with its advantages of macroscopic, real-time, dynamic, and large-scale observation of the Earth, is widely used in crop identification and monitoring. The fusion of information from different sensors can reduce or eliminate potential ambiguity, incompleteness, uncertainty, and variability in target features; therefore, the potential of multi-source satellite imagery is attracting increasing attention. Combining two or more satellite image data to achieve crop identification and classification is becoming a new development trend.

[0003] Optical images with rich spectral information and microwave remote sensing images with scattering information are important data sources for remote sensing crop classification. The Sentinel series satellites, in particular, offer short revisit times, high resolution, and their data is freely available worldwide. Sentinel-1 and...

[0004] The fusion of Sentinel-2 data has become a research hotspot for crop classification tasks. However, existing methods do not fully utilize the polarization information of SAR dual-polarization data. Furthermore, due to the different modes of data from different sensors, existing methods that directly combine spectral bands with SAR backscattering coefficients and then use SVM, RF, decision trees, or neural networks to achieve crop classification have limited ability to fuse heterogeneous source data, leaving room for further improvement in the accuracy of crop classification.

[0005] Therefore, how to introduce SAR polarization information into crop classification tasks and how to effectively integrate heterogeneous source data to improve crop classification accuracy and achieve refined crop classification mapping are urgent problems to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a crop classification method, system, storage medium, and electronic device based on a dual-channel CNN, which can extract deep features of data through convolutional neural networks and deeply integrate features from heterogeneous source data to form a new data set with richer information, effectively integrating multi-source information to improve the classification accuracy of crops.

[0007] The technical solution adopted in this invention is as follows:

[0008] A crop classification method using a two-channel CNN, including

[0009] Step S101: Preprocess the Sentinel-1 image to obtain the backscattering map of the VV / VH dual polarization mode;

[0010] Step S102: Preprocess the Sentinel-1 image (generate the polarization matrix C2 after Deburst in step S101) to obtain the C2 matrix;

[0011] Step S103: Perform polarization decomposition on the C2 matrix based on the physical model to generate different polarization component maps;

[0012] Step S104: Preprocess the Sentinel-2 image to obtain a multispectral image;

[0013] Step S105: Divide the real crop samples obtained from field investigation and visual interpretation into training set samples and test set samples in a 7:3 ratio, and label each crop type to generate a label set; label different crop types to form a label set;

[0014] Step S106: Construct a dual-channel convolutional neural network, which consists of an encoder, a feature fusion and a decoder module to extract deep features from the input data and fuse multi-source features. The fused new feature set is then used by a Softmax classifier to classify crops.

[0015] Step S107: Model training. Input the constructed training set, test set, and label set into the dual-channel convolutional neural network model for training. Judge the model training effect through accuracy evaluation index, adjust the parameters in the network to obtain the optimal crop classification model, and save the optimal network model. Step S108: Load the saved optimal model. Input the crop image to be classified into the network for classification, and output and save the final classification result.

[0016] In step S101, the SAR image is preprocessed, including orbit correction, thermal noise removal, radiometric calibration, deburst, multi-look, filtering, terrain correction, and decibel conversion. Through the preprocessing operations, each pixel of the image represents the true radar backscattering coefficient, thereby forming backscattering maps with different polarization modes.

[0017] In step S102, the polarization matrix C2 is generated as follows: each pixel in the image is represented by a 2*2 covariance matrix, as shown in Formula 1:

[0018]

[0019] Where * denotes conjugate transpose, c 11 c 22 All are real numbers, c 12 The C2 matrix is ​​a complex number; it provides the basic data for subsequent polarization decomposition based on the physical model.

[0020] In step S103, the C2 matrix is ​​subjected to polarization decomposition based on a physical model to extract SAR dual-polarization scattering information, specifically as follows:

[0021] The C2 matrix can represent the polarization scattering information of each pixel in the image. By converting the C2 matrix into a Stokes vector and performing polarization decomposition, we can obtain the different polarization scattering components of each pixel, thereby obtaining different polarization scattering component maps, i.e., SAR dual polarization scattering information.

[0022] The preprocessing in step S104 includes atmospheric correction, resampling, and band synthesis. Specifically, during the Sentinel-2 preprocessing, three bands that are not strongly correlated with vegetation are removed, and the other ten spectral bands are uniformly resampled to 10m resolution and synthesized into a multispectral map with 10 bands.

[0023] In step S106, the dual-channel convolutional neural network extracts deep features from multi-source data and effectively fuses heterogeneous source feature maps to obtain a feature set containing more information, thereby improving the model's classification accuracy for crops.

[0024] In step S108, the optimal model parameters are loaded, and the backscattering map and polarization component feature map obtained by Sentinel-1 and the multispectral map obtained by Sentinel-2 are used as input for classification. The classification results are output and saved.

[0025] A dual-channel CNN crop classification system includes:

[0026] The preprocessing unit is configured to preprocess the Sentinel-1 image to obtain a VV / VH backscatter map; perform orbit correction, thermal noise removal, radiometric calibration, deburst, polarization matrix C2 generation, multi-view, filtering, terrain correction, and decibel conversion operations on the Sentinel-1 image to obtain backscatter images and polarization matrices C2 with different polarization modes; and preprocess the Sentinel-2 image to obtain a multispectral map.

[0027] The polarization decomposition unit is configured to perform polarization decomposition on the polarization matrix C2 to obtain different polarization component maps.

[0028] The dual-channel convolutional neural network unit is configured as an encoder, feature fusion, decoder, fully connected layer and classifier. It performs feature extraction, feature fusion and other operations on different dataset inputs in the two branches to obtain the optimal network model and parameters.

[0029] The test unit is configured to classify crops based on the input Sentinel-1 and Sentinel-2 images using the optimal network model and parameters, and to evaluate the model's classification performance through accuracy assessment.

[0030] A computer-readable storage medium having a computer program stored thereon, the program being a crop classification method using a dual-channel CNN as described in any of the above descriptions.

[0031] An electronic device includes: a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the dual-channel CNN crop classification method.

[0032] This invention accurately classifies crops using Sentinel-1 / 2 image data and generates reliable crop distribution maps, which is of great significance for guiding agricultural production, rationally allocating resources, and solving food security issues. This invention effectively incorporates the polarization information of dual-polarization data extracted from Sentinel-1 images into crop classification, fully utilizing the powerful feature extraction capabilities of convolutional neural networks while effectively fusing rich information from SAR and optics. This improves classification accuracy, achieves fine-grained crop classification, meets the requirements for crop classification mapping, and provides a research foundation for subsequent crop growth monitoring, yield prediction, and other related work. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic diagram of the process of this invention;

[0035] Figure 2 The image shows the farmland area being classified in an embodiment of the present invention; Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] like Figure 1 As shown, the present invention includes the following steps:

[0038] Step S101: Preprocess the Sentinel-1 image (orbit correction, thermal noise removal, radiometric calibration, deburst, multi-look, filtering, terrain correction, decibel conversion) to obtain backscattered images with different polarization modes; Step S102: Preprocess the Sentinel-1 image (adding a polarization matrix C2 generation operation after the Deburst preprocessing step above) to obtain the polarization matrix C2; Step S103: Perform physical model-based polarization decomposition on the polarization matrix C2 to generate images with different polarization components; Step S104: Preprocess the Sentinel-2 image (atmospheric correction, resampling, band synthesis). Obtain multispectral images; Step S105: Divide the samples from field investigation and visual interpretation into training and test sets, and label them to obtain a label set; Step S106: Construct a dual-channel convolutional neural network structure, extract deep features from the two input datasets through branch networks, fuse them, and achieve the final crop classification, outputting the classification result; Step S107: Train the network model, input the training set, test set, and label set into the network model for training, and save the optimal network model; Step S108: Load the saved model to classify the target, save and output the result, and output and save the final result of crop classification according to different needs.

[0039] Preferably, in steps S101 and S102, the filtering operation is selected as Refined Lee filtering, and the window size is 7*7.

[0040] Preferably, in step S103, a dual-polarization SAR data polarization decomposition method based on a physical model is selected to decompose the polarization matrix C2, thereby obtaining the volume scattering polarization component m. v and the remaining polarization components m s This forms images with different polarization components.

[0041] Preferably, in step S104, the coastal / aerosol band, water vapor band, and cirrus band, which are not strongly correlated with vegetation, are removed, and the final synthesized multispectral image contains the other 10 spectral bands of Sentinel-2.

[0042] Preferably, in step S106, the feature fusion process is implemented by three modules in the dual-channel convolutional neural network: the encoder, the feature fusion module, and the decoder module. The encoder and feature fusion module are used to extract deep features and perform deep fusion of heterogeneous sources, respectively. The decoder module maps the fused features back to the spatial information of the original image. This module is then connected to a fully connected layer to use a Softmax classifier for classification.

[0043] Preferably, in step S109, the following formula is used as the evaluation index for the classification results:

[0044]

[0045] In the formula, TP represents the number of pixels that are actually and predicted to be the same crop, FP represents the number of pixels that are actually crop A but the network predicts to be crop B or C, and FN represents the number of pixels that are actually crop B or C but the network predicts to be crop A. By evaluating the learning level of the training samples and labels, the accuracy of the test set samples is assessed, the evaluation metrics for each model are calculated, and the model with the highest accuracy is saved as the optimal training model.

[0046] To achieve the above objectives, the method of the present invention is implemented as follows:

[0047] Sentinel-1 images were preprocessed (orbit correction, thermal noise removal, radiometric calibration, deburst, multi-view, filtering, terrain correction, decibel conversion) to obtain backscattered images with different polarizations.

[0048] The Sentinel-1 image is preprocessed (the polarization matrix C2 generation operation is added after the Deburst preprocessing step in step 1) to obtain the polarization matrix C2;

[0049] The polarization matrix C2 is subjected to polarization decomposition based on a physical model to generate polarization component feature images;

[0050] Sentinel-2 images were preprocessed (atmospheric correction, resampling, and band synthesis) to obtain multispectral images. During preprocessing, bands 1 (coastal / aerosol band), 9 (water vapor band), and 10 (cirrus band), which were not strongly correlated with vegetation, were removed. The final multispectral image contained the remaining 10 bands of Sentinel-2.

[0051] The samples from field investigations and visual interpretations were divided into training and testing sets in a 7:3 ratio, and labels were obtained for various crops.

[0052] A dual-channel convolutional neural network is constructed, including encoder, feature fusion, and decoder modules. This network performs feature extraction and deep fusion of heterogeneous source features from the input data, and utilizes the fused feature set for crop classification. Each module contains convolutional layers, Batch Normalization (BN) layers, max pooling layers, and activation layers.

[0053] The processed training and test sets are input into the constructed dual-channel convolutional neural network model for training. The classification model is obtained by learning the features of the training set. During the training process, the classification effect of the model is adjusted by changing the parameters. The classification accuracy of the model is judged according to the evaluation index. The network model with the best classification effect is output and saved.

[0054] Load the pre-trained optimal classification model, input the remote sensing image of farmland to be classified, perform the classification work, and output and save the classification results.

[0055] This invention utilizes Sentinel-1 / 2 data images to accurately classify crops and draw reliable crop distribution maps, which is of great significance for guiding agricultural production, rationally allocating resources, and solving food security issues.

[0056] The applicant's research on existing technologies revealed that current methods for classifying crops using multi-source remote sensing images can be mainly divided into traditional classification methods based on SVM (Support Vector Machine), RF (Random Forest), decision trees, etc., and deep learning classification methods based on CNN (Convolutional Neural Network), FCN (Fully Convolutional Network), etc.

[0057] Traditional classification methods such as SVM, RF, and decision trees tend to focus more on spectral features in images, neglecting the crucial role of spatial structure in classification tasks. These methods require manual parameter setting before feature extraction; however, features extracted through manual parameter setting are only specific to the target object, lacking flexibility. Furthermore, some features, such as texture and shape, are highly variable, making it impossible to describe all types of features through empirical parameter settings, resulting in low robustness. Moreover, when using multi-source data for classification, these traditional methods often rely on shallow features to categorize targets, failing to extract essential, abstract, and distinguishable features.

[0058] Deep learning-based classification methods, such as CNNs and FCNs, require little to no human intervention in classification tasks and can extract deeper, more abstract features from data. They also demonstrate the powerful computational capabilities of deep learning networks when handling massive datasets. Compared to traditional classification methods, deep learning-based methods are more advantageous when dealing with high-dimensional data. However, existing deep learning-based methods for multi-source crop classification do not fully utilize the polarization information of SAR images, and when using multi-source data, they directly resample heterogeneous source data to the same resolution for combination, limiting their ability to fuse multi-source features.

[0059] It is evident that existing techniques for crop classification using multi-source remote sensing images do not utilize the rich polarization information of SAR images and lack more advanced strategies to eliminate the differences between heterogeneous source features, enabling them to be fully integrated to form a new dataset with richer information. Therefore, the applicant proposes a dual-channel CNN method for crop classification. Sentinel-1 images are preprocessed (orbit correction, thermal noise removal, radiometric calibration, deburst, multi-view, terrain correction, decibel conversion) to obtain backscattered images with different polarization modes. During the Sentinel-1 image preprocessing (adding a polarization matrix C2 generation operation after the Deburst step), polarization matrix C2 is obtained. Polarization decomposition is performed on polarization matrix C2 to generate images with different polarization components. Sentinel-2 images are preprocessed (atmospheric correction, resampling, band synthesis) to obtain multispectral images. Field survey and visual interpretation samples are divided into training and testing sets in a 7:3 ratio and labeled. A dual-channel convolutional neural network structure is constructed, using the dataset combining backscattering coefficients and polarization component features obtained from Sentinel-1 as one input to the branch network, and the 10-band multispectral images obtained from Sentinel-2 as the other input. Deep features from the two input datasets are extracted and fused. A network model is trained using the training sets of the Sentinel-1 and Sentinel-2 input data. Relevant parameters are adjusted based on the accuracy evaluation metrics used in the test set to obtain the network model with the best classification accuracy, and its output is saved. In step S108, the saved optimal model is loaded, and the crops to be classified are classified. The results are then saved and output. This method effectively incorporates the polarization information of the bipolar data extracted from Sentinel-1 images into crop classification. It fully utilizes the powerful feature extraction capabilities of convolutional neural networks while effectively fusing rich information from SAR and optics, improving classification accuracy and achieving fine-grained crop classification. This meets the requirements for crop classification mapping and provides a research foundation for subsequent crop growth monitoring, yield prediction, and other related work.

[0060] The following examples illustrate the method:

[0061] Figure 1 This is a flowchart of a dual-channel CNN crop classification method according to the present invention. Figure 1 As shown, the method includes:

[0062] Step S101: Preprocess the Sentinel-1 image (orbit correction, radiometric calibration, deburst, multi-view, filtering, terrain correction, decibel conversion) to obtain the backscattering coefficient image of VV / VH polarization mode.

[0063] Orbit correction of SAR images yields more accurate orbit files, enabling more precise subsequent processing. Radiometric calibration ensures that each pixel value represents the true radar backscatter value, resulting in an accurate backscatter image. Deburst effectively removes signal-free portions from Sentinel-1IW SLC format images. Multi-look processing effectively eliminates speckle noise in SAR images, reducing its impact on subsequent classification and significantly decreasing data volume. Filtering further eliminates speckle noise in SAR images, minimizing its influence on subsequent processing. Terrain correction performs topographic radiometric correction on SAR images, making the preprocessed image more consistent with real-world terrain features. Preprocessing yields backscatter images with both VV and VH polarizations.

[0064] Step S102: Preprocess the Sentinel-1 image (orbit correction, radiometric calibration, Deburst, generation of polarization matrix C2, multi-view, filtering, terrain correction) to obtain polarization matrix C2.

[0065] In step S102, a polarization matrix C2 generation operation is added to the SAR image preprocessing. After this preprocessing step, each pixel in the image is represented by a 2*2 covariance matrix C2, as shown in Formula 1:

[0066]

[0067] Where * denotes conjugate transpose, c 11 c 22 All are real numbers, c 12 It is a complex number.

[0068] Step S103: Perform polarization decomposition on the C2 matrix based on the physical model, and then extract SAR polarization scattering information.

[0069] Convert the C2 matrix into a Stokes vector, as shown in formula (2):

[0070]

[0071] In the formula, Re(c 12 ) represents c 12 The real part, Im(c 12 ) represents c 12 The imaginary part.

[0072] According to the model, the Stokes vector can be converted into the sum of three polarization components, as shown in formula (3):

[0073] S =m v sv +m s s p +n s n (3)

[0074] In the formula, n is the noise term. s n Given a stokes vector with random polarization, the product of these two terms can be removed by filtering during preprocessing, so this term can be ignored; the main focus is on m. v Volume scattering polarization component, m s The remaining polarization components can be obtained from these two items.

[0075] By polarization decomposing the polarization matrix C2 of the SAR image, maps of different polarization components are obtained.

[0076] Step S104: Preprocess the Sentinel-2 image (atmospheric correction, resampling, band synthesis) to obtain a multispectral image.

[0077] Atmospheric correction can eliminate the influence of atmospheric and light factors on ground reflection; except for three bands that are not strongly correlated with vegetation (coastal / aerosol band, water vapor band, and cirrus band), the other 10 bands of Sentinel-2 data are uniformly resampled to 10m resolution; the 10 resampled bands are combined and used as one of the inputs to the network.

[0078] Step S105: Create training and test set samples for the training network.

[0079] Based on field surveys and visual interpretation, some real-world land cover samples were obtained. All samples were divided into training and testing sets in a 7:3 ratio. Labels were then added according to the actual crop types. Step S106: Construct a dual-channel convolutional neural network.

[0080] The network consists of encoder, feature fusion, and decoder modules to extract high-level features from the input data and perform deep fusion of heterogeneous source features in branch channels. A Softmax classifier is then connected to achieve crop classification. The encoder module includes convolutional layers, batch normalization (BN) layers, max pooling layers, and activation layers. The convolutional kernels are 3x3 with 16 kernels; BN layers accelerate the training process; 2x2 max pooling layers are used for downsampling to reduce data variance and computational complexity; and the ReLU function is chosen as the activation function for the activation layers. The feature fusion module includes convolutional layers, BN layers, max pooling layers, and activation layers. First, the deep features extracted from the two inputs by the encoder module are concatenated, and then they are deeply fused through convolution and other operations to obtain a feature map with richer information. The decoder module includes upsampling layers and activation layers to map the fused features back to the spatial information of the original image size. A fully connected layer is then connected to a Softmax classifier to perform crop classification.

[0081] Step S107: Train a dual-channel convolutional neural network and save the optimal model.

[0082] The relevant parameters in the network are set as follows: bathh_size is 32, the maximum number of iterations is 200, and the learning rate is 0.001. Then, the training dataset and labels are input into the constructed network model to train the network. Every 50 iterations, the network performance is tested using the validation dataset, and the model with the highest evaluation obtained from the training is saved.

[0083] Step S108: Load the saved optimal network model and perform classification on the images to be classified. Load the optimal model parameters saved in step S107, input the remote sensing image dataset of the target to be classified into the trained network model, and the output layer is the classification result for each crop in the image. Save the classification results to create a crop distribution map.

[0084] Figure 1 This is a flowchart illustrating the crop classification process using Sentinel-1 / 2 images according to the present invention. Figure 2 This is a map showing the distribution of farmland in Tongxiang City, Zhejiang Province. This invention utilizes Sentinel-1 / 2 images and a dual-channel convolutional neural network to classify crops into three categories: rapeseed, wheat, and bare land. The overall accuracy reaches 98.4%, which, compared to other commonly used neural network classification methods (2D-CNN 94.6%, FCN 96.3%, SegNet 96.9%), produces a more reliable crop distribution map and therefore has significant practical value.

[0085] In the description of this invention, it should be noted that the directional terms such as "center", "lateral", "longitudinal", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", and "counterclockwise" indicate the orientation and positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. They should not be construed as limiting the specific protection scope of this invention.

[0086] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and claims of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0087] Note that the above description is merely a preferred embodiment and application of the technical principles of the present invention. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the specific embodiments described herein, and may include many other effective embodiments without departing from the concept of the present invention. The scope of the present invention is determined by the scope of the appended claims.

Claims

1. A crop classification method using a dual-channel CNN, characterized in that: include Step S101: Preprocess the Sentinel-1 image to obtain a backscattering map of VV / VH dual polarization mode. In step S101, the SAR image preprocessing includes orbit correction, thermal noise removal, radiometric calibration, deburst, multi-look, filtering, terrain correction, and decibel conversion. Through the preprocessing operation, each pixel of the image represents the true radar backscattering coefficient, thereby constructing backscattering maps of different polarization modes. Step S102: After Deburst in step S101, the polarization matrix C2 is generated, and the C2 matrix is ​​obtained; Step S103: Perform polarization decomposition on the C2 matrix based on the physical model to generate different polarization component maps; In step S103, the C2 matrix is ​​subjected to polarization decomposition based on a physical model to extract SAR dual-polarization scattering information, specifically as follows: The C2 matrix represents the polarization scattering information of each pixel in the image. The C2 matrix is ​​converted into a Stokes vector for polarization decomposition to obtain the different polarization scattering components of each pixel, thereby obtaining a map of different polarization scattering components, i.e., SAR dual polarization scattering information. Step S104: Preprocess the Sentinel-2 image to obtain a multispectral image; the preprocessing in step S104 includes atmospheric correction, resampling and band synthesis; during the Sentinel-2 preprocessing, three bands that are not strongly correlated with vegetation are removed, and the other ten spectral bands are uniformly resampled to 10m resolution and synthesized into a multispectral image with 10 bands. Step S105: Divide the real crop samples obtained from field investigation and visual interpretation into training set samples and test set samples in a 7:3 ratio, and label each crop type to generate a label set; label different crop types to form a label set; Step S106: Construct a dual-channel convolutional neural network, which consists of an encoder, a feature fusion and a decoder module to extract deep features from the input data and fuse multi-source features. The fused new feature set is then used by a Softmax classifier to classify crops. Step S107: Model training. Input the constructed training set, test set, and label set into the dual-channel convolutional neural network model for training. Judge the model training effect through accuracy evaluation index, adjust the parameters in the network to obtain the optimal crop classification model, and save the optimal network model. Step S108: Load the saved optimal model, input the crop map to be classified into the network for classification, and output and save the final classification result.

2. The crop classification method using a dual-channel CNN according to claim 1, characterized in that, In step S102, the polarization matrix C2 is generated as follows: each pixel in the image is represented by a 2*2 covariance matrix, as shown in Formula 1: Where * denotes conjugate transpose, c 11 c 22 All are real numbers, c 12 The C2 matrix is ​​a complex number; it provides the basic data for subsequent polarization decomposition based on the physical model.

3. The crop classification method using a dual-channel CNN according to claim 1, characterized in that, In step S106, the dual-channel convolutional neural network extracts deep features from multi-source data and effectively fuses heterogeneous source feature maps to obtain a feature set containing more information.

4. The crop classification method using a dual-channel CNN according to claim 1, characterized in that, In step S108, the optimal model parameters are loaded, and the backscattering map and polarization component feature map obtained by Sentinel-1 and the multispectral map obtained by Sentinel-2 are used as input for classification. The classification results are output and saved.

5. A crop classification system based on a dual-channel CNN, characterized in that, The crop classification method using a dual-channel CNN as described in claim 1 includes: The preprocessing unit is configured to preprocess the Sentinel-1 image to obtain a VV / VH backscatter map; perform orbit correction, thermal noise removal, radiometric calibration, deburst, polarization matrix C2 generation, multi-view, filtering, terrain correction, and decibel conversion operations on the Sentinel-1 image to obtain backscatter images and polarization matrices C2 with different polarization modes; and preprocess the Sentinel-2 image to obtain a multispectral map. The polarization decomposition unit is configured to perform polarization decomposition on the polarization matrix C2 to obtain different polarization component maps; the dual-channel convolutional neural network unit is configured as an encoder, feature fusion, decoder, fully connected layer and classifier, which performs feature extraction on different dataset inputs of the two branches, and the feature fusion operation obtains the optimal network model and parameters. The test unit is configured to classify crops based on the input Sentinel-1 and Sentinel-2 images using the optimal network model and parameters, and to evaluate the model's classification performance through accuracy assessment.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is a crop classification method using a dual-channel CNN as described in any one of claims 1-4.

7. An electronic device, characterized in that, include: The memory, the processor, and the program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the crop classification method of the dual-channel CNN as described in any one of claims 1-4.

Images