Time sequence SAR image rice extraction method based on boundary constraint

Abstract

The invention discloses a time sequence SAR image rice extraction method based on boundary constraint, and relates to the technical field of rice planting area extraction based on time sequence SAR images, and the method comprises the steps: carrying out the rice region label and rice boundary label labeling of a high-resolution optical remote sensing image in a region; acquiring a multi-period time sequence SAR image in the same-region rice growth cycle; processing the time sequence SAR image and the high-resolution optical image to form a data set; inputting the data of the data set into the network model, and performing supervised training on the network model based on the rice region label and the rice boundary label; and predicting a time sequence SAR image by using the trained network model, and obtaining rice distribution information. The method comprises the steps of preprocessing, feature extraction, loss function design and post-processing, and the accuracy and fineness of rice extraction are effectively improved.

Description

Technical Field

[0001] The invention relates to the technical field of rice planting area extraction based on time-series SAR images, and in particular to a rice extraction method based on time-series SAR images based on boundary constraints. Background Art

[0002] Domestic grain production functional areas and important agricultural product production protection areas include rice, wheat, and corn three major grain production functional areas and soybean, cotton, rapeseed, sugar cane, natural rubber and other five important agricultural product production protection areas. As one of the important food crops, rice is mainly to ensure "basic self-sufficiency in grain and absolute food security". my country is one of the countries with the largest rice planting area in the world, and rice occupies a pivotal position in my country's agricultural production. However, due to the complexity and variability of the rice planting environment, traditional manual survey methods can no longer meet the needs of rice planting area extraction.

[0003] At present, remote sensing is the most convenient, economical and timely means of monitoring agricultural information. Remote sensing technology is objective and timely, and can obtain a large amount of ground information over a wide range in a short period of time. Remote sensing technology can monitor the status of ground objects over a large area and is widely used in remote sensing crop classification and growth monitoring. Accurate monitoring of crop planting areas has always been the core issue of agricultural remote sensing. Traditional optical satellite remote sensing monitoring mainly extracts crop phenological characteristics through time series, or realizes crop classification monitoring through spectral differences between different crops. At present, rice planting monitoring mainly adopts the time series phenological feature extraction and analysis method.

[0004] The application of deep learning technology in the field of remote sensing has become an important research direction. It processes and analyzes a large amount of remote sensing data by simulating the neural network of the human brain, thereby realizing the automatic recognition and classification of ground objects. In recent years, with the development of deep learning algorithms, its application in remote sensing image processing has made significant progress.

[0005] Deep learning technology has shown great potential in extracting objects from remote sensing images. For example, by using deep learning models such as convolutional neural networks (CNN) and recurrent neural networks (RNN), target detection and object extraction in remote sensing images can be effectively performed. The CNN-LSTM combined model has also been used for time series remote sensing data analysis, such as NDVI prediction and feature importance assessment, which shows that deep learning technology can not only process static image data, but also dynamic time series data.

[0006] For example: the invention application with application number 202210017568.9 proposes a SAR identification method for single-season rice in small and medium-sized fields under complex terrain environment, performs average synthesis and time series filtering on the backscatter intensity time series of the SAR image of the study area throughout the year, and uses a decision tree algorithm model to identify single-season rice planting areas. The invention application with application number 202211689677.1 proposes a rice planting area extraction method and device based on multi-temporal SAR images, uses a preset classification method to classify the ground object samples to obtain the initial rice planting area identification result; the initial rice planting area identification result is optimized to obtain the optimized rice planting area identification result.

[0007] The above scheme also has the following problems: when using deep learning time-series SAR images to extract rice planting area and distribution, there is a lack of application of boundary constraints, resulting in the extracted rice area patches and planting boundary patches being insufficiently refined. Summary of the invention

[0008] In view of the above-mentioned problems, the object of the present invention is to provide a rice extraction method for time-series SAR images based on boundary constraints, so as to improve the accuracy and precision of rice planting area extraction based on boundary constraints.

[0009] The embodiment of the present invention provides a rice extraction method for time series SAR images based on boundary constraints.

[0010] A first aspect: A rice extraction method for time-series SAR images based on boundary constraints, comprising the steps of:

[0011] S1. Label the rice area and rice boundary labels on the high-resolution optical remote sensing images in the area;

[0012] S2, obtain time-series SAR images of multiple periods in the rice growth cycle of the same area;

[0013] S3, processing the time series SAR images and optical images to form a data set;

[0014] S4, inputting the data set into the network model, and conducting supervised training on the network model based on the rice region label and the rice boundary label;

[0015] S5. Use the trained network model to predict time-series SAR images and obtain rice distribution information.

[0016] Furthermore, the S1 includes: selecting a certain number of rice heading period rate optical remote sensing images as auxiliary annotation images, manually marking the rice in the optical remote sensing images, and forming rice area labels and rice boundary label annotations.

[0017] Furthermore, the step of forming a data set in S3 includes:

[0018] S31. According to the rice phenological characteristics in the region, a certain number of SAR images in different growth periods are selected for preprocessing;

[0019] S32, performing pixel registration between the preprocessed time series SAR images, and performing registration and alignment between the time series SAR images and the optical images;

[0020] S32, after converting the time series SAR images according to the decibel (dB) value, stacking them in time order in the channel dimension to form a data set with the annotated optical remote sensing images;

[0021] The preprocessing in S31 includes radiation calibration, atmospheric correction, orthorectification and geometric alignment.

[0022] Furthermore, the network model in S4 adopts the DeepLabV3+ network as the framework, and the ResNet-50 network with the hybrid attention mechanism module CBAM and the spatial pyramid pooling module ASPP as the backbone network, which is divided into a rice region feature extraction branch and a rice boundary feature extraction branch.

[0023] Furthermore, the rice region feature extraction branch includes the following processing steps:

[0024] The time series SAR images output shallow features of rice semantic segmentation through STAGE0 and STAGE1; the deep feature map of rice semantic segmentation is output through STAGE0-STAGE4, ​​and then the deep features are obtained through ASPP and upsampling. After the shallow features and deep features are fused, the size of the feature map is restored, and then the sigmoid classifier is used to predict the rice regional features.

[0025] Furthermore, the rice boundary feature branch processing steps include:

[0026] The time series SAR image is convolved and upsampled by STAGE0 to obtain Feature edge1; the output features of STAGE0 are convolved and upsampled after STAGE1 to obtain Feature edge2; STAGE1 is successively passed through STAGE2, STAGE3 and STAGE4 to obtain output features, and the output features of each stage are convolved and upsampled to obtain Feature edge3, Featureedge4 and Feature edge5 respectively; the multi-layer feature maps are fused and superimposed, and then the rice boundary features are predicted through the sigmoid classifier.

[0027] Furthermore, when the network model is supervisedly trained in S4, a joint cost loss function is used, including:

[0028] The binary cross entropy loss function is used to calculate the rice area loss, and the RCF loss function is used to calculate the rice boundary loss. The weighted sum of the two losses is used to iteratively train the network model weight parameters. The formula is expressed as:

[0029] L=L1+L2

[0030] Among them, L1 is the binary cross entropy loss function and L2 is the RCF loss function.

[0031] Furthermore, the S5 also includes: post-processing the rice distribution information, removing small patches and misclassifications, and obtaining the final rice spatial distribution result.

[0032] A second aspect: An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method provided in the first aspect when executing the program.

[0033] A third aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0034] Beneficial effects of the present invention:

[0035] 1. The extraction method of the present invention comprises a feature extraction module, a hybrid attention mechanism module CBAM, a spatial pyramid pooling module ASPP and a rice boundary constraint module, etc., which uses high-resolution optical remote sensing images to perform auxiliary annotation of rice distribution patches during the rice heading period, and improves the accuracy and refinement of extracting rice areas and planting boundaries from time-series SAR images based on deep learning.

[0036] 2. The network model used in the extraction method of the present invention adopts a binary cross entropy loss function and an RCF loss function for joint training, and the weighted sum of the two losses is used to iterate the model weight parameters. The network model has high modeling accuracy and improves the accuracy of recognition and extraction of rice-growing areas.

[0037] 3. The network model used in the method of the present invention introduces the CBAM module and the spatial pyramid pooling module ASPP, which enhances the rice feature extraction capability. Through the network model prediction, more detailed rice patches and planting boundaries can be obtained.

[0038] 4. The method of the present invention performs post-processing on the acquired distribution information to optimize the extraction results, thereby effectively improving the accuracy and precision of rice extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1It is a flow chart of the rice extraction method of time-series SAR images based on boundary constraints of the present invention;

[0040] Figure 2 A schematic diagram of the network model structure applied by the method of the present invention;

[0041] Figure 3 This is a schematic diagram of the network model stage structure of the present invention;

[0042] Figure 4 It is a schematic diagram of the internal structure of the STAGE 1 stage of the network model of the present invention;

[0043] Figure 5 This is a schematic diagram of the internal structure of the network model ASPP of the present invention;

[0044] Figure 6 A schematic diagram of a time-series SAR image sample with boundary constraints of the present invention;

[0045] Figure 7 Another schematic diagram of a time-series SAR image sample with boundary constraints of the present invention;

[0046] Figure 8 A schematic diagram of the Jilin-1 high-resolution optical remote sensing image of the research area of ​​the present invention;

[0047] Fig. 9 A schematic diagram of extracting local rice patches from time-series SAR images according to the present invention;

[0048] Fig.10 It is a comparison diagram of the local rice patch extraction result of the time-series SAR image of the present invention and the optical remote sensing image;

[0049] Fig.11 It is a schematic diagram of the structure of the electronic device of the present invention. DETAILED DESCRIPTION

[0050] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar symbols throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0051] Currently, there is a lack of application of deep learning time-series SAR images with boundary constraints to quickly extract rice planting area and distribution.

[0052] In view of the above problems, the present invention provides a rice extraction method based on boundary constraints in time series SAR images. Figure 1 A schematic flow chart of an extraction method provided by an embodiment of the present invention, the method comprising the steps of:

[0053] S1. Label the rice area and rice boundary labels on the high-resolution optical remote sensing images in the area.

[0054] A certain number of high-resolution optical remote sensing images of rice heading period were selected as auxiliary annotation images, and the rice distribution patches in the high-resolution optical remote sensing images were manually annotated. The annotation contents included rice area patches and rice boundary patches, forming rice area labels and rice boundary label annotations.

[0055] S2. Obtain time-series SAR images of multiple periods within the rice growth cycle in the same area.

[0056] Acquire SAR images of multiple periods within the rice growth cycle in the same area as the high-resolution optical remote sensing images to form time-series SAR images. Preprocess the time-series SAR images, including radiation calibration, atmospheric correction, orthorectification and geometric registration, to obtain standard time-series SAR images.

[0057] S3. Process the time-series SAR images and high-resolution optical images to form a data set.

[0058] Firstly, the time series SAR images and the high-resolution optical images are aligned, specifically including pixel registration between the time series SAR images and registration and alignment between the time series SAR images and the high-resolution optical images.

[0059] The time-series SAR images are expressed in decibels (dB), and then all values ​​are converted to 8 bits (0-255), stacked in time order in the channel dimension, and form a data set with the annotated optical remote sensing images.

[0060] S4. Input the data set into the network model and perform supervised training on the network model based on the rice region label and rice boundary label.

[0061] The network model used in the present invention is as follows Figure 2-5 As shown in the figure, the DeepLabV3+ network is used as the framework, and the ResNet-50 network with the spatial and channel mixed attention mechanism module CBAM (Convolutional Block Attention Module) and the spatial pyramid pooling module ASPP (Atrous spatial pyramid pooling) is used as the backbone network. The entire network is divided into a rice region feature extraction branch and a rice boundary feature extraction branch.

[0062] The dataset is input into the encoding layer. When extracting rice features, the CBAM module is integrated to capture the rice characteristics in spatial and channel dimensions, and rice feature maps of different depths are extracted respectively. The rice region feature extraction branch and the rice boundary feature extraction branch are used, combined with the spatial pyramid pooling module ASPP, to obtain the fusion of rice feature maps of different depths; the decoding layer outputs the rice region feature map, and then based on the rice boundary feature map constraints, the rice planting area and distribution are quickly extracted.

[0063] like Figure 2 As shown in the figure, the rice area feature extraction branch has the following processing steps: the time series SAR image outputs the shallow features of rice semantic segmentation through stage 0 (STAGE0) and stage 1 (STAGE1); the deep feature map of rice semantic segmentation is output through STAGE0-STAGE4, ​​and then the deep features are obtained through ASPP and upsampling; after the shallow features and the deep features are fused, the size of the feature map is restored, and then the sigmoid classifier is used to predict the rice area features.

[0064] The rice boundary feature branch includes the following processing steps: the time series SAR image is convolved and upsampled by STAGE0 to obtain rice boundary feature 1; the output feature of STAGE0 is convolved and upsampled by rice boundary feature 1 (Feature edge1) to obtain Feature edge2; the output feature of STAGE1 is successively passed through STAGE2, STAGE3 and STAGE4 to obtain the output feature, and the output features of each stage are convolved and upsampled to obtain rice boundary features Feature edge3, Feature edge4 and Feature edge5; the multi-layer feature maps are fused and superimposed, and finally the rice boundary feature is predicted by the sigmoid classifier.

[0065] When training the network model, the binary cross entropy loss function is used to calculate the rice area loss, and the RCF loss function is used to calculate the rice boundary loss. The weighted sum of the two losses is used to iteratively train the network model weight parameters. The formula is expressed as:

[0066] L=L1+L2

[0067] Among them, L1 is the binary cross entropy loss function and L2 is the RCF loss function.

[0068] An example of network model training is given as follows: a certain number of high-resolution optical remote sensing images of rice heading period are selected as auxiliary annotations. In this embodiment, the Jilin-1 satellite remote sensing image with a spatial resolution of 0.7m of the southern Anhui region on August 19, 2022 is selected. The rice areas and their boundaries are annotated with artificial rice distribution patches, including rice area patches and rice boundary patches, and rice area and boundary vectors are produced.

[0069] According to the phenological characteristics of rice in the study area, Sentinel-1SAR images of six phases from May to November 2022 during the phenological period were selected, and the SAR images were preprocessed by radiation calibration, atmospheric correction, orthorectification and geometric registration to obtain standard time-series SAR images.

[0070] The time-series SAR images and rice area and boundary vectors are aligned, and the vectors are converted into raster; finally, the time-series SAR images and optical remote sensing images are uniformly divided into sample data sets of size 512×512.

[0071] The final sample data set contains a total of 63,556 groups of samples, such as Figure 6 and Figure 7 As shown in the figure, each group of samples contains 3 images, which are time-series SAR images with a size of 512×512×6 (6 is the number of multi-phases in SAR images); rice area patch label 1 with a size of 512×512, pixel values ​​0 and 1, corresponding to non-rice areas and rice areas respectively; rice boundary constraint is label 2 with a size of 512×512, pixel values ​​0 and 1, corresponding to non-rice area boundaries and rice area boundaries respectively; the sample data set is divided into training set, validation set and test set according to 7:2:1.

[0072] The network model training process includes: inputting the data set into the encoding layer, and each stage of the encoding layer contains a CBAM module, which is divided into a rice area feature extraction branch and a rice boundary feature extraction branch.

[0073] Rice regional feature extraction branch Figure 2 and 3As shown in the figure, the feature map extracted from the time series SAR image is interpolated after ASPP. Specifically, the time series SAR image first passes through stage 0, i.e. STAGE0 (STAGE0 contains a 7×7 convolution layer (Conv7×7) and a maximum pooling layer (Maxpool)) and stage 1, i.e. STAGE1 (STAGE1 consists of skeleton network 1 (BONE1), skeleton network 2 (BONE2) and a hybrid attention mechanism module (CBAM)) to output the shallow features of rice semantic segmentation; then the deep feature map is extracted through STAGE0-STAGE4, ​​and then the deep features are obtained through ASPP and upsampling, and the shallow features and deep features are fused to achieve the purpose of combining high-level semantic information with low-level high-resolution features, and the feature map size is restored to 512×512, and then the sigmoid classifier is used for prediction, and the cross entropy loss function value L1 is calculated with the rice distribution sample.

[0074] Rice boundary feature extraction branch Figure 2 As shown in the figure, after the time series SAR image passes through STAGE0, the output feature map is subjected to 2 layers of 1×1 convolution and upsampling to obtain rice boundary feature 1 (Feature edge1); the feature map output by STAGE0 passes through STAGE1 to obtain the feature map, which is then subjected to 2 layers of 1×1 convolution and upsampling to obtain rice boundary feature 2 (Feature edge2); the output feature of STAGE1 passes through STAGE2, STAGE3 and STAGE4 in turn to obtain the boundary feature, and the boundary feature is subjected to 2 layers of 1×1 convolution and upsampling to obtain rice boundary features Feature edge3, Feature edge4 and Feature edge5. The size of the multi-layer feature map is restored to 512×512, and then superimposed and fused in the channel dimension. Finally, the boundary is predicted by the sigmoid classifier, and the loss function value L2 of RCF is calculated with the rice boundary sample.

[0075] The joint cost loss function is L=L1+L2, and then the adaptive matrix estimation Adam optimizer is used to perform back propagation and update the model weights.

[0076] S5. Use the trained network model to predict the time-series SAR images to obtain the rice distribution information, post-process the rice distribution information, remove small patches and misclassifications to obtain the final rice spatial distribution results.

[0077] For example, a part of Xuancheng City in Anhui Province was selected as the research area, and 10-meter resolution Sentinel-1SAR images and 0.7-meter resolution Jilin-1 satellite optical images (such as Figure 8The data is shown in the figure, where the Sentinel-1SAR image range is from June to November 2022, with a total of 6 scenes; the Jilin-1 satellite image time is August 19, 2022, with a total of 1 scene. After data preprocessing, it is synthesized into multi-channel data with a size of 512×512×6. The time series SAR images are input into the network model, and the predicted rice area distribution probability map and rice boundary probability map can be obtained respectively. The rice distribution area probability map and the rice boundary probability are combined to obtain more detailed and accurate rice spatial distribution information.

[0078] The rice distribution information obtained in the above steps is post-processed to remove woodlands, water bodies, buildings and other crops to obtain the final rice spatial distribution results. Fig.10 The subsequent processing steps belong to the prior art and will not be described in detail here.

[0079] Verified by the interpretation of the 0.7-meter-resolution Jilin-1 image, the average overall accuracy (OA) of rice extraction results ranged from 94.18% to 0.92, with a Kappa coefficient of 0.92, indicating a high overall accuracy.

[0080] The rice extraction method based on time series SAR image based on boundary constraint disclosed by the present invention comprises a feature extraction module, a hybrid attention mechanism module, a spatial pyramid pooling module ASPP, and a rice boundary constraint module; firstly, high-resolution optical remote sensing images are used to assist in marking rice distribution spots during the rice heading period, and multi-period SAR image data within the rice growth cycle in the same region are obtained, and the pixels are aligned with the optical images after preprocessing. Then, through the designed model, the time series SAR image data set is input to extract rice and its boundary features, and the binary cross entropy loss function and the RCF loss function are used for joint training. The CBAM module and the spatial pyramid pooling module ASPP are also specially introduced in the steps to enhance the feature extraction capability. Finally, more refined rice spots and planting boundaries are obtained through model prediction, and post-processing is performed to optimize the extraction results. The whole process includes preprocessing, feature extraction, loss function design and post-processing, which effectively improves the accuracy and precision of rice extraction.

[0081] The present invention also provides an electronic device, Fig.11 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Fig.11 As shown, the electronic device may include: a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may call the logic instructions in the memory, for example, to execute the following method:

[0082] S1. Label the rice area and rice boundary labels on the high-resolution optical remote sensing images in the area;

[0083] S2, obtain time-series SAR images of multiple periods in the rice growth cycle of the same area;

[0084] S3, processing the time series SAR images and the high-resolution optical images to form a data set;

[0085] S4, inputting the data set into the network model, and conducting supervised training on the network model based on the rice region label and the rice boundary label;

[0086] S5. Use the trained network model to predict time-series SAR images and obtain rice distribution information.

[0087] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0088] An embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method provided in each of the above embodiments is implemented, for example, including:

[0089] S1. Label the rice area and rice boundary labels on the high-resolution optical remote sensing images in the area;

[0090] S2, obtain time-series SAR images of multiple periods in the rice growth cycle of the same area;

[0091] S3, processing the time series SAR images and the high-resolution optical images to form a data set;

[0092] S4, inputting the data set into the network model, and conducting supervised training on the network model based on the rice region label and the rice boundary label;

[0093] S5. Use the trained network model to predict time-series SAR images and obtain rice distribution information.

[0094] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0095] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rice extraction method based on boundary constraints in time series SAR images, characterized in that: include: S1. Label the rice area and rice boundary labels on the high-resolution optical remote sensing images in the area; S2, obtain time-series SAR images of multiple periods in the rice growth cycle of the same area; S3, processing the time series SAR images and optical images to form a data set; S4, inputting the data set into the network model, and conducting supervised training on the network model based on the rice region label and the rice boundary label; S5. Use the trained network model to predict time-series SAR images and obtain rice distribution information.

2. The extraction method according to claim 1, characterized in that The step S1 includes: selecting a certain number of rice heading period optical remote sensing images as auxiliary annotation images, manually marking the rice in the optical remote sensing images, and forming rice area labels and rice boundary label annotations.

3. The extraction method according to claim 1, characterized in that The steps of forming a data set in S3 include: S31. According to the rice phenological characteristics in the region, a certain number of SAR images in different growth periods are selected for preprocessing; S32, performing pixel registration between the preprocessed time series SAR images, and performing registration and alignment between the time series SAR images and the optical images; S32, after converting the time series SAR images according to the decibel (dB) value, stacking them in time order in the channel dimension to form a data set with the annotated optical remote sensing images; The preprocessing in S31 includes radiation calibration, atmospheric correction, orthorectification and geometric alignment.

4. The extraction method according to claim 1, characterized in that The network model in S4 adopts the DeepLabV3+ network as the framework, and the ResNet-50 network with the hybrid attention mechanism module CBAM and the spatial pyramid pooling module ASPP as the backbone network, which is divided into a rice region feature extraction branch and a rice boundary feature extraction branch.

5. The extraction method according to claim 4, characterized in that The rice region feature extraction branch includes the following processing steps: The time series SAR images output shallow features of rice semantic segmentation through STAGE0 and STAGE1; the deep feature map of rice semantic segmentation is output through STAGE0-STAGE4, ​​and then the deep features are obtained through ASPP and upsampling; after the shallow features and deep features are fused, the size of the feature map is restored, and then the sigmoid classifier is used to predict the rice area features.

6. The extraction method according to claim 4, characterized in that The rice boundary feature branch processing steps include: The time series SAR image is convolved and upsampled by STAGE0 to obtain Feature edge1; the output features of STAGE0 are convolved and upsampled after STAGE1 to obtain Feature edge2; STAGE1 is successively passed through STAGE2, STAGE3 and STAGE4 to obtain output features, and the output features of each stage are convolved and upsampled to obtain Feature edge3, Feature edge4 and Feature edge5 respectively; the multi-layer feature maps are fused and superimposed, and then the rice boundary features are predicted through the sigmoid classifier.

7. The extraction method according to claim 1, characterized in that When the network model is supervised in S4, a joint cost loss function is used, including: The binary cross entropy loss function is used to calculate the rice area loss, and the RCF loss function is used to calculate the rice boundary loss. The weighted sum of the two losses is used to iteratively train the network model weight parameters. The formula is expressed as: L=L1+L2 Among them, L1 is the binary cross entropy loss function and L2 is the RCF loss function.

8. The extraction method according to claim 1, characterized in that The step S5 further includes: post-processing the rice distribution information, removing small patches and misclassifications, and obtaining the final rice spatial distribution result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the extraction method according to any one of claims 1 to 8 are implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the extraction method according to any one of claims 1 to 8 are implemented.

Images