A farmland class segmentation method, system and readable storage medium
Patent Information
- Application Number
- CN202610903767.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-23
- Publication Date
- 2026-10-30
AI Technical Summary
然而，受地形起伏、地物光谱混淆及样本不均衡等因素的制约，目前该技术在复杂地形下仍然存在适应性不强的问题
[0029]1）以U-Net网络的Encoder-Decoder结构为骨架，利用CNN提取局部纹理特征，利用Transformer提取全局语义特征，实现了局部细节与全局语境的互补；
Smart Images
Figure CN123016189A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image segmentation technology, and in particular to a method, system, and readable storage medium for segmenting cultivated land types. Background Technology
[0002] With the acceleration of urbanization and the rapid development of the social economy, farmland protection faces many practical challenges. For example, in their study of Changde City, Huang Weiyao and Zou Yu mentioned that the phenomenon of farmland being converted to non-agricultural or non-grain uses occurs frequently in some areas. The pressure on farmland from infrastructure construction and industrial expansion continues to increase. In addition, the impact of natural disasters and ecological protection factors makes it difficult to maintain a stable dynamic balance between the quantity and quality of farmland.
[0003] To accurately grasp the quantity, distribution, and utilization status of arable land for effective protection and scientific management, traditional arable land monitoring methods rely on manual surveys or conventional remote sensing interpretation. While these methods can obtain macroscopic distribution and basic quantity information of arable land within a certain range, they suffer from low efficiency, insufficient accuracy, and delayed updates, making it difficult to meet the needs of refined management. With the widespread application of satellite remote sensing technology and deep learning algorithms, intelligent automatic arable land extraction technology based on deep learning has emerged. This technology can quickly and accurately identify arable land boundaries and distinguish utilization types from high-resolution remote sensing images, providing technical support for timely detection of abandoned land, monitoring arable land transfer, and assessing protection effectiveness. However, due to limitations such as topographic relief, spectral confusion of ground features, and sample imbalance, this technology still faces challenges in adapting to complex terrain. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a method, system and readable storage medium for classifying arable land, in order to overcome the shortcomings of the prior art.
[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A method for classifying arable land types, the method comprising:
[0006] S1. Determine the historical remote sensing images of the target study area and bind the historical remote sensing images to the binary annotation map of cultivated land used to store cultivated land information;
[0007] S2. Cut the bound image pairs according to the preset input image size, and divide the training set based on the resulting cut image pairs;
[0008] S3. Input the training set into the initial TransUnet++ network for model training, and obtain the target TransUnet++ network when the preset training termination condition is reached. The TransUnet++ network is based on the Encoder-Decoder structure of the U-Net network, and embeds a Transformer-Blocks module between the deepest feature map of the Encoder and the shallowest feature map of the Decoder to improve the feature representation capability under complex terrain. The Transformer-Blocks module is composed of multiple Transformer-Blocks stacked together. Each Transformer-Block consists of multiple Transformer layers and ASPP layers connected in series after the multiple Transformer layers.
[0009] S4. Input the acquired real-time remote sensing images into the target TransUnet++ network to obtain a farmland land classification prediction map.
[0010] Furthermore, in step S1, the historical remote sensing images used to determine the target study area include:
[0011] S11. Acquire initial remote sensing images of the target study area and perform geometric transformation enhancement on the initial remote sensing images to simulate the features of remote sensing images from different shooting angles.
[0012] S12. Rotate the transformed and enhanced image at a random angle, and use bilinear interpolation to maintain the image resolution during the rotation, so as to determine the historical remote sensing image of the target study area.
[0013] Furthermore, in step S3, the TransUnet++ network is also connected to a PSE module at the output of the decoder for fine optimization of farmland boundaries and connectivity repair of fragmented plots.
[0014] Furthermore, in step S3, the Transformer-Blocks module includes 24 Transformer layers, and an ASPP layer is inserted every 3 Transformer layers.
[0015] Furthermore, in step S3, during the training process, a target hybrid loss function based on a weighted combination of the Dice Loss function, the Focal Loss function, and the BCE Loss function is adopted.
[0016] Furthermore, the target mixture loss function is calculated in the following form:
[0017] ;
[0018] ;
[0019] ;
[0020] in, Represents the total loss function. Represents the pixel-level loss function. Represents the multi-scale loss function. Indicates the weighting coefficient. This represents the Dice loss function used to handle class imbalance. This represents the Focal loss function used to address the imbalance between easy and difficult samples. This represents the binary cross-entropy loss function. , This represents the weighting coefficient that balances the proportions of the three pixel-level losses mentioned above. This represents the weight coefficient for the k-th scale. This represents the k-th void space pyramid pooling module. This represents the transformed feature map. This indicates the actual label.
[0021] Furthermore, the method also includes: S5, performing spatiotemporal registration and comparative analysis of the predicted farmland classification map and the historical farmland labeling map obtained in step S1, in order to identify the changes in farmland within the target study area and provide data support for the quantitative assessment of subsequent farmland protection measures.
[0022] Secondly, this application discloses a farmland classification system, the system comprising a data acquisition module, a training set partitioning module, a classification model training module, and a farmland classification prediction module, wherein:
[0023] The data acquisition module is used to determine the historical remote sensing images of the target study area and bind the historical remote sensing images to the binary annotation map of cultivated land used to store cultivated land information;
[0024] The training set partitioning module is used to cut the bound image pairs according to the preset input image size, and to partition the training set based on the resulting cut image pairs.
[0025] The segmentation model training module is used to input the training set into the initial TransUnet++ network for model training, and to obtain the target TransUnet++ network when the preset training termination condition is reached. The TransUnet++ network is based on the Encoder-Decoder structure of the U-Net network, and a Transformer-Blocks module for improving the feature representation capability under complex terrain is embedded between the deepest feature map of the Encoder and the shallowest feature map of the Decoder. The Transformer-Blocks module is composed of multiple Transformer-Blocks stacked together. Each Transformer-Block consists of multiple Transformer layers and ASPP layers connected in series after the multiple Transformer layers.
[0026] The cultivated land classification prediction module is used to input the acquired real-time remote sensing images into the target TransUnet++ network to obtain a cultivated land classification prediction map.
[0027] Thirdly, this application discloses a readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for classifying arable land categories as described in any of the preceding claims.
[0028] The beneficial effects of this invention are:
[0029] 1) Using the Encoder-Decoder structure of the U-Net network as the backbone, CNN is used to extract local texture features, and Transformer is used to extract global semantic features, thus achieving complementarity between local details and global context;
[0030] 2) After the ASPP layer is concatenated with multiple Transformer layers, global self-attention is first calculated through the Transformer layer to establish long-range dependencies between pixels; then multi-scale dilated convolution is performed through the ASPP layer to extract local features under different receptive fields. This "global first, multi-scale later" concatenation method enables the network to "see far" and "see in detail" when processing remote sensing images under complex terrain, which greatly improves the feature representation ability.
[0031] 3) The Transformer's self-attention mechanism combined with U-Net's skip connections enables the network to more accurately recover the spatial details of ground features during the decoding stage. In particular, it can generate segmentation masks that are more in line with reality for the irregular edges of farmland commonly found in remote sensing images. Attached Figure Description
[0032] Figure 1This is a flowchart illustrating a method for classifying arable land categories disclosed in this invention.
[0033] Figure 2 This is a schematic diagram illustrating the labeled images and the final recognition results of the model.
[0034] Figure 3 A schematic diagram of the TransUnet++ network structure;
[0035] Figure 4 This is a structural diagram of the Transformer-Blocks module;
[0036] Figure 5 Training results of a TransUnet++ model with ASPP layers inserted at different layer intervals;
[0037] Figure 6 A visual comparison chart of the TransUnet network and the TransUnet++ network proposed in this invention;
[0038] Figure 7 This is a schematic diagram of the structure of a farmland classification system disclosed in this invention;
[0039] Figure 8 This is a schematic diagram of the structure of a readable storage medium disclosed in this invention. Detailed Implementation
[0040] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0041] like Figure 1 As shown, this application discloses a method for classifying arable land categories, the method comprising:
[0042] Step S1: Determine the historical remote sensing images of the target study area and bind the historical remote sensing images to the binary annotation map of cultivated land used to store cultivated land information.
[0043] Step S2: Cut the bound image pairs according to the preset input image size, and divide the training set based on the resulting cut image pairs.
[0044] Please refer to Figure 2Regarding steps S1 and S2, it should be noted that this application uses Hengyang City change survey images as the dataset source. The images contain img files storing Hengyang City's coordinates and map information, as well as shapefile (SHP) annotations for different land parcel categories (e.g., cultivated land, forest land, wetlands) within Hengyang City. Specifically, in implementation, this application uses a raster tool to extract cultivated land location information from the labeled SHP information for different land parcel categories in Hengyang City and converts it into a masked tiff file that only stores cultivated land location information. It should be noted that in this masked tiff file, cultivated land has a value of 1, and non-cultivated land has a value of 0.
[0045] In one embodiment, this application uses the Python batch processing tool in ArcGIS software to align the map information and land use information of Hengyang City, and cut the corresponding Hengyang City map img file and the two-dimensional mask tiff image containing Hengyang cultivated land information to obtain an image of size 256×256 pixels, and input it into the TransUnet++ network for training.
[0046] Step S3: Input the training set into the initial TransUnet++ network for model training, and obtain the target TransUnet++ network when the preset training termination condition is reached. The TransUnet++ network is based on the Encoder-Decoder structure of the U-Net network, and embeds a Transformer-Blocks module between the deepest feature map of the Encoder and the shallowest feature map of the Decoder to improve the feature representation capability under complex terrain. The Transformer-Blocks module is composed of multiple Transformer-Blocks stacked together. Each Transformer-Block consists of multiple Transformer layers and ASPP layers connected in series after the multiple Transformer layers.
[0047] Please refer to Figure 3 The TransUnet++ model proposed in this application is based on the Encoder-Decoder structure of the U-Net network, wherein:
[0048] (1) The encoder in the U-Net network extracts local features through convolutional blocks and downsampling.
[0049] (2) An enhanced Transformer module is introduced between the deepest feature map of the Encoder and the shallowest feature map of the Decoder, and ASPP is inserted intermittently to model the global context relationship, so as to improve the feature representation capability under complex terrain.
[0050] (3) The decoder in the U-Net network fuses multi-scale features through upsampling and skip connections. In one embodiment, a PSE module for fine optimization of farmland boundaries and connectivity repair of fragmented plots is also connected to the output of the decoder.
[0051] Step S4: Input the acquired real-time remote sensing image into the target TransUnet++ network to obtain a farmland land classification prediction map.
[0052] As can be seen from the above, the farmland land classification method disclosed in this application uses the Encoder-Decoder structure of the U-Net network as its backbone, utilizes CNN to extract local texture features, and utilizes Transformer to extract global semantic features, thus achieving complementarity between local details and global context. After concatenating the ASPP layer with multiple Transformer layers, global self-attention calculation is first performed through the Transformer layer to establish long-range dependencies between pixels. Then, multi-scale dilated convolution is performed through the ASPP layer to extract local features under different receptive fields. This "global first, multi-scale later" concatenation method enables the network to "see far" and "see in detail" when processing remote sensing images with complex terrain, greatly improving the feature representation capability. The self-attention mechanism of Transformer combined with the skip connections of U-Net enables the network to more accurately recover the spatial details of land features during the decoding stage, especially for the irregular edges of farmland commonly found in remote sensing images, and can generate a segmentation mask that is more in line with the real situation.
[0053] In one embodiment, according to the method of claim 1, the step S1, determining the historical remote sensing images of the target study area, includes:
[0054] Step S11: Acquire initial remote sensing images of the target study area and perform geometric transformation enhancement on the initial remote sensing images to simulate the features of remote sensing images from different shooting angles.
[0055] Specifically, during the geometric transformation enhancement process, this application will flip the original image horizontally and vertically to change the direction of terrain distribution in order to simulate the characteristics of remote sensing images taken from different angles.
[0056] In one embodiment, this application will also acquire the digital elevation model (DEM) data corresponding to the target study area, and adaptively adjust the intensity of the above-mentioned geometric transformation based on the terrain undulation gradient information in the DEM. For example, in flat terrain areas, a larger geometric transformation is allowed to fully simulate the multi-angle shooting features; in steep terrain areas, the range of values for flipping and rotating angles is adaptively reduced to avoid introducing unreasonable enhancement samples due to terrain structure distortion caused by excessive geometric transformation, thus ensuring the physical rationality of the enhanced image.
[0057] Step S12: Rotate the transformed and enhanced image at a random angle, and use bilinear interpolation to maintain the image resolution while rotating the image, so as to determine the historical remote sensing image of the target study area.
[0058] Specifically, this application will perform random angle rotation within the range of [-5°, 5°]. In addition, this application will also use bilinear interpolation to maintain image resolution when rotating the image, thereby avoiding the model's tendency to make untargeted judgments about edge areas due to excessive rotation angle.
[0059] In one embodiment, before the image is input into the network, the present application will also perform regularization processing to standardize the pixel value distribution in order to accelerate gradient convergence during model training and reduce the interference caused by differences in data distribution on model parameter updates.
[0060] In one embodiment, in step S3, the TransUnet++ network is also connected to a PSE module at the output of the decoder for fine optimization of farmland boundaries and connectivity repair of fragmented plots.
[0061] Specifically, the PSE module, as an efficient post-processing strategy, demonstrates significant advantages in remote sensing segmentation: 1) It can effectively connect small land parcels caused by noise, resolution limitations, or incomplete model segmentation through the progressive expansion of multi-scale kernels, enhancing land parcel connectivity and integrity; 2) Its multi-scale kernel design can accurately capture geometric features, while segmenting from the smallest kernel and utilizing the geometric edge differences between adjacent instances to achieve effective separation, avoiding the erroneous merging of features in dense scenes. Its computational process can be described as follows:
[0062] Step 1: The PSE module will first start from the smallest core ( Begin by extracting the initial connected components.
[0063] Step 2: Based on 4-neighborhood or 8-neighborhood rules, from arrive Expand the connected components sequentially;
[0064] Step 3: Ensure the continuity of expansion through breadth-first search and maintain instance consistency by combining label propagation, and finally output the complete segmentation result.
[0065] This strategy can synergistically optimize the multi-scale features of remote sensing data, and the parameters can be flexibly adjusted according to the scene to adapt to the connected domain merging task under different terrain and lighting conditions.
[0066] In one embodiment, in step S3, the Transformer-Blocks module includes 24 Transformer layers, and an ASPP layer is inserted every 3 Transformer layers.
[0067] For details, please refer to Figure 4 In the diagram, the top gray box, Transformer-Blocks, represents the entire encoder or feature extraction backbone network. It is composed of multiple identical Transformer-Blocks (blue boxes) stacked vertically. This stacking design allows the network to extract semantic features layer by layer from low to high levels. The three blue cuboids arranged on the left within the dashed box represent multiple standard Transformer Encoder layers. Each layer typically contains a multi-head self-attention mechanism and a feedforward neural network, which utilizes self-attention to capture long-range dependencies in the image, crucial for understanding the distribution of large-scale ground features in remote sensing images. The rightmost yellow cuboid within the dashed box represents the ASPP (Atrous Spatial Pyramid Pooling) module, which extracts multi-scale features at different dilation rates.
[0068] In the current embodiment, the ASPP layer is embedded into the Transformer-Block to compensate for the shortcomings of the pure Transformer in capturing local details and multi-scale textures. The synergy and complementarity of the two enhances global semantic understanding while preserving the detailed textures of ground features. This structure is designed specifically for remote sensing segmentation tasks and can effectively balance the global modeling of large-scale ground feature distribution with the need for accurate identification of small-scale ground features.
[0069] In one embodiment, during step S3, a target hybrid loss function based on a weighted combination of the Dice Loss function, the Focal Loss function, and the BCE Loss function is used during the training process.
[0070] In one embodiment, the target mixture loss function is calculated as follows:
[0071] ;
[0072] ;
[0073] ;
[0074] in, Represents the total loss function. Represents the pixel-level loss function. Represents the multi-scale loss function. Indicates the weighting coefficient. This represents the Dice loss function used to handle class imbalance. This represents the Focal loss function used to address the imbalance between easy and difficult samples. This represents the binary cross-entropy loss function. , This represents the weighting coefficient that balances the proportions of the three pixel-level losses mentioned above. This represents the weight coefficient for the k-th scale. This represents the k-th void space pyramid pooling module. This represents the transformed feature map. This indicates the actual label.
[0075] Specifically, the objective hybrid loss function constructs a dual-supervision mechanism combining "pixel-level fine-grained classification" and "multi-scale feature constraints," aiming to optimize network performance from different dimensions. Among these, The function term employs a weighted fusion of the DiceLoss, FocalLoss, and BCELoss functions to provide complementary optimizations for class imbalance, imbalance between easy and difficult samples, and basic pixel classification problems, respectively. Specifically, the DiceLoss function focuses on increasing the overlap rate of ground feature regions to address background-dominated class imbalance; the FocalLoss function focuses on focusing on difficult samples to enhance the learning of boundary details; and the BCELoss function provides stable gradients to ensure the model's basic convergence capability. The term utilizes the ASPP module to extract features under different receptive fields and calculates their L1 distance with the ground truth (GT). This function term aims to bring the supervision signal down to the feature extraction stage, forcing the network to learn feature representations with multi-scale semantic consistency, thereby narrowing the distance between the prediction results and the real distribution at the feature level and improving the model's ability to perceive ground objects at different scales.
[0076] In one embodiment, the method further includes:
[0077] Step S5 involves performing spatiotemporal registration and comparative analysis between the predicted farmland classification map and the historical farmland labeling map obtained in step S1 to identify changes in farmland within the target study area and provide data support for the quantitative assessment of subsequent farmland protection measures.
[0078] Specifically, this application first performs spatiotemporal registration based on SIFT feature point matching and the RANSAC algorithm. By extracting corner features from two images and removing mismatched points, subpixel-level geometric correction is achieved by calculating the affine transformation matrix, eliminating registration errors caused by terrain undulations and shooting angles. Then, adaptive morphological filtering is used to denoise and optimize connected components of changing pixels. Opening operations are used to remove isolated noise points, and closing operations are used to fill internal holes, generating a dynamic change map of cultivated land with clear boundaries and accurate area, providing high-precision data for assessing the effectiveness of cultivated land protection.
[0079] It should be noted that this application also includes multiple sets of experiments for comparison: no ASPP layer, one ASPP layer every 3 layers, one ASPP layer every 4 layers, one ASPP layer every 5 layers, and one ASPP layer every 6 layers. Additionally, this application uses `dept=n` to indicate the interval between ASPP layers when using a PSE. The training results are shown in Table 1 below. It can be seen that as the number of Transformer layers between ASPP layers decreases, the performance of the TransUnet++ network improves.
[0080] Table 1. Test results of TransUnet++ network with different ASPP layer insertion depths.
[0081]
[0082] It should be further noted that the training process of the TransUnet++ model with ASPP inserted between Transformer-Block layers of different depths is shown in the following figure. Figure 5 As shown, the model's accuracy improves as the interval training loss of the interval insertion ASPP decreases.
[0083] Furthermore, this application also tested the farmland segmentation task on the Hengyang City farmland dataset using TransUnet++ networks with and without tail-based PSE modules, respectively. The test results are shown in Table 2 below. The experimental results verify the effectiveness of the PSE module in the upsampling process within the Encoder-Decoder structure.
[0084] Table 2. TransUnet++ network test results with and without PSE upsampling at the tail.
[0085]
[0086] It should be further noted that the training comparison curves of the TransUnet network and the TransUnet++ network on the Hengyang City farmland remote sensing segmentation dataset are as follows: Figure 6 As shown, the loss reduction rate of the TransUnet++ network is much better than that of the TransUnet network, which verifies the effectiveness of the TransUnet++ network in the farmland segmentation task in Hengyang City.
[0087] Please refer to Figure 7 This application discloses a farmland classification system, the system comprising a data acquisition module, a training set partitioning module, a classification model training module, and a farmland classification prediction module, wherein:
[0088] The data acquisition module is used to determine the historical remote sensing images of the target study area and bind the historical remote sensing images to the binary annotation map of cultivated land used to store cultivated land information.
[0089] The training set partitioning module is used to cut the bound image pairs according to a preset input image size, and to partition the training set based on the resulting cut image pairs.
[0090] The segmentation model training module is used to input the training set into the initial TransUnet++ network for model training, and obtain the target TransUnet++ network when the preset training termination condition is reached. The TransUnet++ network is based on the Encoder-Decoder structure of the U-Net network, and a Transformer-Blocks module for improving the feature representation capability under complex terrain is embedded between the deepest feature map of the Encoder and the shallowest feature map of the Decoder. The Transformer-Blocks module is composed of multiple Transformer-Blocks stacked together. Each Transformer-Block consists of multiple Transformer layers and ASPP layers connected in series after the multiple Transformer layers.
[0091] The cultivated land classification prediction module is used to input the acquired real-time remote sensing images into the target TransUnet++ network to obtain a cultivated land classification prediction map.
[0092] In one embodiment, the above modules are also used to implement a farmland classification system as described in any of the preceding claims, and this application does not limit this.
[0093] As can be seen from the above, the farmland land classification system disclosed in this application uses the Encoder-Decoder structure of the U-Net network as its backbone, utilizes CNN to extract local texture features, and utilizes Transformer to extract global semantic features, thus achieving complementarity between local details and global context. After concatenating ASPP layers with multiple Transformer layers, global self-attention calculation is first performed through the Transformer layers to establish long-range dependencies between pixels. Then, multi-scale dilated convolution is performed through the ASPP layers to extract local features under different receptive fields. This "global first, multi-scale later" concatenation method enables the network to "see far" and "see in detail" when processing remote sensing images with complex terrain, greatly improving the feature representation capability. The self-attention mechanism of Transformer combined with the skip connections of U-Net enables the network to more accurately recover the spatial details of land features during the decoding stage, especially for the irregular edges of farmland commonly found in remote sensing images, and can generate a segmentation mask that is more in line with the real situation.
[0094] Please refer to Figure 8 This application discloses a readable storage medium storing a computer program that, when executed by a processor, implements a method for classifying arable land categories as described in any of the foregoing embodiments.
[0095] As can be seen from the above, the readable storage medium disclosed in this application uses the Encoder-Decoder structure of the U-Net network as its backbone, utilizes CNN to extract local texture features, and utilizes Transformer to extract global semantic features, thus achieving complementarity between local details and global context. After concatenating ASPP layers with multiple Transformer layers, global self-attention calculation is first performed through the Transformer layers to establish long-range dependencies between pixels. Then, multi-scale dilated convolution is performed through the ASPP layers to extract local features under different receptive fields. This "global first, multi-scale later" concatenation method enables the network to "see far" and "see in detail" when processing remote sensing images with complex terrain, greatly improving the feature representation capability. The self-attention mechanism of Transformer combined with the skip connections of U-Net enables the network to more accurately recover the spatial details of ground objects during the decoding stage, especially for the irregular edges of farmland commonly found in remote sensing images, and can generate segmentation masks that are more in line with the real situation.
[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for classifying arable land by type, characterized in that, The method includes: S1. Determine the historical remote sensing images of the target study area and bind the historical remote sensing images to the binary annotation map of cultivated land used to store cultivated land information; S2. Cut the bound image pairs according to the preset input image size, and divide the training set based on the resulting cut image pairs; S3. Input the training set into the initial TransUnet++ network for model training, and obtain the target TransUnet++ network when the preset training termination condition is reached. The TransUnet++ network is based on the Encoder-Decoder structure of the U-Net network, and embeds a Transformer-Blocks module between the deepest feature map of the Encoder and the shallowest feature map of the Decoder to improve the feature representation capability under complex terrain. The Transformer-Blocks module is composed of multiple Transformer-Blocks stacked together. Each Transformer-Block consists of multiple Transformer layers and ASPP layers connected in series after the multiple Transformer layers. S4. Input the acquired real-time remote sensing images into the target TransUnet++ network to obtain a farmland land classification prediction map.
2. The method according to claim 1, characterized in that, In step S1, the historical remote sensing images used to determine the target study area include: S11. Acquire initial remote sensing images of the target study area and perform geometric transformation enhancement on the initial remote sensing images to simulate the features of remote sensing images from different shooting angles. S12. Rotate the transformed and enhanced image at a random angle, and use bilinear interpolation to maintain the image resolution during the rotation, so as to determine the historical remote sensing image of the target study area.
3. The method according to claim 1, characterized in that, In step S3, the TransUnet++ network is also connected to a PSE module at the output of the decoder for fine optimization of farmland boundaries and connectivity repair of fragmented plots.
4. The method according to claim 1, characterized in that, In step S3, the Transformer-Blocks module includes 24 Transformer layers, and an ASPP layer is inserted every 3 Transformer layers.
5. The method according to claim 1, characterized in that, In step S3, during the training process, a target hybrid loss function based on a weighted combination of Dice Loss, Focal Loss, and BCE Loss functions is used.
6. The method according to claim 5, characterized in that, The target mixture loss function is calculated in the following form: ； ； ； in, Represents the total loss function. Represents the pixel-level loss function. Represents the multi-scale loss function. Indicates the weighting coefficient. This represents the Dice loss function used to handle class imbalance. This represents the Focal loss function used to address the imbalance between easy and difficult samples. This represents the binary cross-entropy loss function. , This represents the weighting coefficient that balances the proportions of the three pixel-level losses mentioned above. This represents the weight coefficient for the k-th scale. This represents the k-th void space pyramid pooling module. This represents the transformed feature map. This indicates the actual label.
7. The method according to claim 1, characterized in that, The method further includes: S5. Perform spatiotemporal registration and comparative analysis between the predicted farmland classification map and the historical farmland labeling map obtained in step S1 to identify farmland changes within the target study area and provide data support for the quantitative assessment of subsequent farmland protection measures.
8. A system for classifying arable land categories, characterized in that, The system includes a data acquisition module, a training set partitioning module, a segmentation model training module, and a farmland land type segmentation prediction module, wherein: The data acquisition module is used to determine the historical remote sensing images of the target study area and bind the historical remote sensing images to the binary annotation map of cultivated land used to store cultivated land information; The training set partitioning module is used to cut the bound image pairs according to the preset input image size, and to partition the training set based on the resulting cut image pairs. The segmentation model training module is used to input the training set into the initial TransUnet++ network for model training, and to obtain the target TransUnet++ network when the preset training termination condition is reached. The TransUnet++ network is based on the Encoder-Decoder structure of the U-Net network, and a Transformer-Blocks module for improving the feature representation capability under complex terrain is embedded between the deepest feature map of the Encoder and the shallowest feature map of the Decoder. The Transformer-Blocks module is composed of multiple Transformer-Blocks stacked together. Each Transformer-Block consists of multiple Transformer layers and ASPP layers connected in series after the multiple Transformer layers. The cultivated land classification prediction module is used to input the acquired real-time remote sensing images into the target TransUnet++ network to obtain a cultivated land classification prediction map.
9. A readable storage medium, characterized in that, It stores a computer program, characterized in that, when the program is executed by a processor, it implements a method for classifying arable land categories as described in any one of claims 1-7.