Target-oriented crop field classification method and device for fragmented farmland

Abstract

This application provides a target-oriented crop plot-level classification method and device for fragmented farmland. The method includes: using a U-TAE encoder-decoder backbone network to extract multi-scale features from the temporal data of multi-temporal remote sensing image sequences in the input information; using a parallel spatiotemporal self-attention module to enhance the temporal and spatial features of the features output from the encoding stage of the backbone network; using a parameterless plot aggregation operator to aggregate the fused features into a plot-level representation; performing plot-level classification on the plot-level representation; and back-broadcasting the classification results to the corresponding pixel location of the plot based on the plot ID map. This application captures long-range dependencies through U-TAE and spatiotemporal attention, achieves end-to-end plot-level classification through parameterless plot aggregation and bi-branch prediction, eliminates salt-and-pepper noise while maintaining accurate boundaries, suppresses SAR speckle noise, and achieves efficient and accurate identification of fragmented farmland.

Description

Technical Field

[0001] This application belongs to the technical field of remote sensing image information extraction, and in particular relates to a target-oriented crop plot-level classification method and equipment for fragmented farmland. Background Technology

[0002] Remote sensing-based crop classification plays a crucial role in precision agriculture and food security monitoring. Accurately identifying crop types helps optimize resource allocation, yield estimation, and agricultural policy formulation, thereby promoting sustainable agricultural management.

[0003] In the field of remote sensing crop classification, there is a lack of a technical solution that can achieve accurate and consistent plot-level crop classification in fragmented farmland scenarios, while also taking into account temporal dynamic modeling and spatial consistency maintenance. Therefore, there is an urgent need to provide a target-oriented plot-level crop classification method and device for fragmented farmland. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this application is to provide a target-oriented crop plot classification method and device for fragmented farmland.

[0005] The first aspect of this application proposes a target-oriented crop plot-level classification method for fragmented farmland, comprising the following steps: S101, acquire the multi-temporal remote sensing image sequence of the area to be classified and the pre-delineated field boundary data as input information; the field boundary raster data is a field ID map that is spatiotemporally aligned with the multi-temporal remote sensing image sequence; S102, a backbone network with a U-TAE encoder-decoder structure is used to extract multi-scale features from the temporal data of the multi-temporal remote sensing image sequence in the input information to obtain intermediate feature maps; the U-TAE encoder-decoder structure includes a U-Net multi-scale convolutional framework and a lightweight temporal self-attention module, which is used to capture long-range temporal dependencies while maintaining the original spatial resolution. S103, the parallel spatiotemporal self-attention module is used to enhance the temporal and spatial features of the features output by the encoding stage of the backbone network to generate fused features; S104, using a parameterless plot aggregation operator, the fusion features are aggregated from pixel level to plot level representation based on the plot ID map; S105, using a dual-branch prediction head, perform plot-level classification on the plot-level representation, and refine the pixel-level boundaries based on the intermediate feature map; broadcast the classification results of the plot-level classification back to the pixel position corresponding to the plot according to the field ID map, and output a plot-level crop type classification map with consistent spatial boundaries.

[0006] Preferably, the multi-scale feature extraction includes the following steps: The temporal data is subjected to shallow semantic feature extraction by an input convolutional layer, which includes two 3×3 convolutional units connected in sequence, a normalization layer and a ReLU activation function. The shallow semantic features are progressively downsampled and expanded through three cascaded downsampled residual blocks; each downsampled residual block reduces the spatial resolution of the feature map to 1 / 2×1 / 2, 1 / 4×1 / 4 and 1 / 8×1 / 8 through convolution operations, while expanding the number of feature channels to 64, 64 and 128 respectively. A lightweight temporal self-attention module is applied to the end layer of the encoder to perform temporal weighted aggregation of the downsampled features; The spatial resolution of the temporally weighted aggregated features is restored by using three cascaded upsampling blocks. The upsampling blocks employ transposed convolution and skip connection structures to restore the original spatial resolution to the same as the temporal data and serve as intermediate feature maps.

[0007] Preferably, the U-TAE encoder-decoder structure further includes the following configuration: Replace the zero-padding method of the convolutional layer with reflection padding to suppress edge artifacts; A Dropout2d layer is inserted before the output convolutional layer, and its dropout rate is set to 0.2 to suppress network overfitting. It provides three switchable time series aggregation modes, including attention grouping mode, attention mean mode, and mean mode, and dynamically selects the time series aggregation mode as the time series smoothing strategy according to the data sparsity of the input time series data.

[0008] Preferably, the parallel spatiotemporal self-attention module includes temporal branches and spatial branches; step S103 includes: The temporal branch is used to process the feature tensor output by the encoder, extract vectors for each spatial location, rearrange them, and then perform weighted aggregation through a 16-head self-attention mechanism. The multi-head outputs are then concatenated and mapped through a two-layer multilayer perceptron to obtain temporal features. Using the aforementioned spatial branch, the features are first averaged along the temporal dimension to suppress speckle noise in synthetic aperture radar. Then, the features are projected to 128 dimensions through group normalization and 1×1 convolution, and learnable row and column position codes are injected. After flattening, spatial weights are calculated through a 4-head self-attention mechanism to obtain the spatial feature representation. The temporal feature representation and the spatial feature representation are concatenated along the channel dimension, compressed to 128 channels by 1×1 convolution, processed by the GELU activation function and Dropout2d regularization, and then added to the temporal feature representation by element-wise residual addition to generate fused features.

[0009] Preferably, step S104 includes: The fused features are reshaped into a two-dimensional feature matrix, and the field ID map is flattened into a field index vector; For all pixels in the plot index vector that are marked as valid plot identifiers, a parallel scattering accumulation operation is performed to accumulate the sum of the feature vectors corresponding to each plot identifier and the pixel count, respectively. Based on the sum of the feature vectors and the pixel count, a stabilization term is introduced to calculate the average feature vector of each plot, generating a plot embedding matrix as the plot-level representation.

[0010] Preferably, step S105 includes: The dual-branch prediction head shares the backbone network parameters, including the land parcel branch and the pixel branch; The plot branch receives the plot embedding matrix represented by the plot level and performs nonlinear mapping through a two-layer fully connected network. The first layer maps the plot features to the embedding dimension space and activates them with ReLU. The second layer maps them to the crop category number dimension and generates the crop category probabilities of each plot level through softmax normalization. The pixel branch applies a 1×1 convolution to the intermediate feature map output by the decoder to generate a pixel-level classification score, which is used for boundary detail refinement. The dual-branch prediction head is jointly optimized and trained using a weighted joint loss function, which includes a weighted sum of plot-level focus loss and pixel-level focus loss. The weighted joint loss function employs a balancing strategy of inverse frequency class weights to mitigate classification bias caused by the long-tail distribution of crop categories.

[0011] In a second aspect, this application also provides an electronic device, the electronic device including a memory and at least one processor, the memory storing a computer program, the at least one processor being configured to execute the computer program to perform the following steps: S101, acquire the multi-temporal remote sensing image sequence of the area to be classified and the pre-delineated field boundary data as input information; the field boundary raster data is a field ID map that is spatiotemporally aligned with the multi-temporal remote sensing image sequence; S102, a backbone network with a U-TAE encoder-decoder structure is used to extract multi-scale features from the temporal data of the multi-temporal remote sensing image sequence in the input information to obtain intermediate feature maps; the U-TAE encoder-decoder structure includes a U-Net multi-scale convolutional framework and a lightweight temporal self-attention module, which is used to capture long-range temporal dependencies while maintaining the original spatial resolution. S103, the parallel spatiotemporal self-attention module is used to enhance the temporal and spatial features of the features output by the encoding stage of the backbone network to generate fused features; S104, using a parameterless plot aggregation operator, the fusion features are aggregated from pixel level to plot level representation based on the plot ID map; S105, using a dual-branch prediction head, perform plot-level classification on the plot-level representation, and refine the pixel-level boundaries based on the intermediate feature map; broadcast the classification results of the plot-level classification back to the pixel position corresponding to the plot according to the field ID map, and output a plot-level crop type classification map with consistent spatial boundaries.

[0012] Thirdly, this application also provides a computer-readable storage medium storing a computer program that, when executed by at least one processor, implements the steps of any of the above methods.

[0013] Fourthly, this application also provides a computer program product comprising a computer program that, when executed by at least one processor, implements the steps of any of the methods described above.

[0014] This application provides a target-oriented crop plot-level classification method and device for fragmented farmland. The technical solution provided in this application captures long-range temporal dependencies while maintaining the original spatial resolution through a U-TAE encoder-decoder structure. Combined with a parallel spatiotemporal self-attention module to jointly enhance temporal and spatial features, it effectively solves the difficulty of feature extraction caused by small plot areas and complex temporal data in fragmented farmland. This achieves synchronous and accurate perception of fragmented farmland scenes (e.g., small plots less than 3×3 pixels and large contiguous plots) in highly fragmented landscapes, improving crop classification accuracy. A parameterless plot aggregation operator smoothly transitions pixel-level features to plot-level representations, and a dual-branch prediction head is used for collaborative optimization of plot-level classification and pixel-level boundary refinement. This architecture enforces consistency of all pixel labels within a plot, eliminating the "salt and pepper" noise generated by traditional pixel-level methods. Simultaneously, a reverse broadcast mechanism ensures consistent output. The precise alignment between the generated classification map and the pre-defined field boundaries ensures that the classification results not only conform to the characteristics of plot units in agricultural management practice but also maintain fine spatial boundary details. The entire framework, through a fully differentiable end-to-end design, unifies multi-scale feature extraction, spatiotemporal attention enhancement, plot-level aggregation, and dual-classification branching into a joint optimization system. This avoids the error propagation problem caused by the separation of segmentation and classification in traditional object-oriented analysis methods, enabling the network to automatically learn the optimal feature representation to adapt to complex fragmented scenarios. Addressing the speckle noise characteristics of SAR data, the scheme constructs a multi-level noise suppression mechanism by combining temporal averaging suppression of spatial branches with statistical denoising of plot-level mean aggregation, significantly improving the model's robustness under complex environmental conditions. Furthermore, the computational design of lightweight temporal self-attention at compressed resolution and the linear complexity of the plot aggregation operator enable efficient, consistent, and end-to-end generation from raw images to plot-level classification results. Attached Figure Description

[0015] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Throughout the drawings, the same reference numerals denote the same components. Obviously, the drawings described below are merely some embodiments described in this application, and those skilled in the art can obtain other drawings based on these drawings.

[0016] Figure 1 This is a flowchart illustrating the overall structure provided in this application.

[0017] Figure 2 This is a location diagram of the research area provided in this application.

[0018] Figure 3 This is a schematic diagram of the structure of the parallel spatiotemporal self-attention module provided in this application.

[0019] Figure 4 This is a schematic diagram of the workflow of the land aggregation operator provided in this application.

[0020] Figure 5 This is a crop classification result diagram using the method of the embodiments provided in this application.

[0021] Figure 6 This is a flowchart illustrating the target-oriented crop plot-level classification method for fragmented farmland provided in this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions in the embodiments of this application, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. It should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application. Furthermore, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily obscuring the concepts disclosed in this application.

[0023] The technical solutions of the embodiments of this application and how the technical solutions of the embodiments of this application solve the above-mentioned technical problems will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the various embodiments or technical features described below can be arbitrarily combined to form new embodiments. The order of description of the embodiments below is not intended to limit the preferred order of embodiments. The same or similar concepts or processes may not be described again in some embodiments. Obviously, the described embodiments are some embodiments of the embodiments of this application, but not all embodiments.

[0024] To facilitate understanding, let's first take one application scenario involved in this application as an example for explanation: In many regions, such as southern China, farmland is highly fragmented, with small plots typically less than 0.1 hectares, corresponding to less than 3×3 pixels at a 10-meter spatial resolution. This fragmented landscape poses a significant challenge to accurate crop classification, primarily in the following aspects: (1) Limitations of pixel-level methods: Traditional pixel-level classification methods often produce 'salt and pepper' noise in fragmented farmland scenarios, making it difficult to maintain spatial consistency within plots, and are sensitive to speckle noise in synthetic aperture radar images. In addition, pixel-level methods cannot naturally correspond to agricultural management practices based on plots, and cannot make full use of spatial consistency within plots to improve classification robustness.

[0025] (2) Shortcomings of traditional goal-oriented methods: Traditional goal-oriented image analysis (OBIA) methods rely on the quality of pre-segmentation, making it difficult to extract temporal features when applied to long-term time-series data. Moreover, most of them require manual feature design and lack the ability to learn representations end-to-end from the original images. The temporal modeling of these methods usually relies on simple statistics or shallow networks, making it difficult to capture the long-term dynamic features of crop growth, especially in highly fragmented landscapes with complex phenological patterns.

[0026] (3) Deficiencies of existing attention mechanism methods: Although attention mechanism-based methods are good at modeling long-term temporal dependencies and global spatial relationships, they are usually pixel-level operations and cannot explicitly enforce plot-level consistency. Moreover, the computational complexity increases significantly with sequence length and spatial resolution, which hinders their deployment in large-scale agricultural monitoring scenarios.

[0027] In other words, how to achieve accurate and consistent plot-level crop classification in fragmented farmland scenarios, while taking into account both temporal dynamic modeling and spatial consistency, is a technical problem that urgently needs to be solved in the field of remote sensing crop classification.

[0028] This application provides a target-oriented crop plot-level classification method and system for fragmented farmland, which solves the above-mentioned technical problems in the following ways: (a) Parallel Spatiotemporal Self-Attention Mechanism for Joint Modeling: The parallel spatiotemporal self-attention module (STSA-LTAE module) captures both the temporal dynamics of crop growth and the spatial relationships between pixels through the parallel design of temporal and spatial branches. The temporal branch utilizes a multi-head attention mechanism to weighted aggregate data from multi-temporal remote sensing image sequences, extracting temporally consistent and speckle-resistant embedding features; the spatial branch explicitly models pixel relationships within the same temporal phase through learnable location encoding and spatial self-attention, improving the ability to distinguish crops with similar temporal features but different spatial distributions.

[0029] (b) Achieving a smooth transition from pixel to plot through the parameterless ParcelPooling operator: The parameterless ParcelPooling operator automatically aggregates pixel features based on the plot ID, suppressing speckle noise and spectral heterogeneity within plots, allowing the classifier to focus on discriminative differences between plots. Compared to traditional object-oriented methods, this operator is fully differentiable, supports end-to-end training, and avoids the impact of segmentation errors on classification accuracy.

[0030] (c) Dual-branch prediction head balances accuracy and boundary: The parcel branch uses the macroscopic representation aggregated by ParcelPooling for classification to ensure spatial consistency within the parcel; the pixel branch captures fine-grained edge structure and provides complementary gradients through joint optimization, improving boundary accuracy while maintaining parcel-level accuracy.

[0031] (d) Weighted Focal Loss for class imbalance: Crop categories in agricultural areas are usually distributed in a long tail. Weighted Focal Loss effectively alleviates the class imbalance problem by using inverse frequency weights and focusing parameters, thereby improving the identification accuracy of rare crops.

[0032] The following text first explains the target-oriented crop plot-level classification method, and then explains the system, etc.

[0033] Example 1.

[0034] See Figure 6 This embodiment provides a target-oriented crop plot-level classification method for fragmented farmland, including S101 to S105: S101, acquire the multi-temporal remote sensing image sequence of the area to be classified and the pre-delineated field boundary data as input information; the field boundary raster data is a field ID map that is spatiotemporally aligned with the multi-temporal remote sensing image sequence; S102, a backbone network with a U-TAE encoder-decoder structure is used to extract multi-scale features from the temporal data of the multi-temporal remote sensing image sequence in the input information to obtain intermediate feature maps; the U-TAE encoder-decoder structure includes a U-Net multi-scale convolutional framework and a lightweight temporal self-attention module LTAE, which is used to capture long-range temporal dependencies while maintaining the original spatial resolution. The multi-scale feature extraction includes the following steps: S201, shallow semantic feature extraction is performed on the time series data through the input convolutional layer, the input convolutional layer includes two 3×3 convolutional units connected in sequence, a normalization layer and a ReLU activation function; S202, the shallow semantic features are progressively downsampled and expanded through three cascaded downsampling residual blocks; wherein each downsampling residual block reduces the spatial resolution of the feature map to 1 / 2×1 / 2, 1 / 4×1 / 4 and 1 / 8×1 / 8 in sequence through convolution operation, while expanding the number of feature channels to 64, 64 and 128 in sequence. S203, a lightweight temporal self-attention module LTAE is applied to the end layer of the encoder to perform temporal weighted aggregation of the downsampled features; S204 involves restoring the spatial resolution of the temporally weighted aggregated features through three cascaded upsampling blocks. These upsampling blocks employ transposed convolution and skip connection structures to restore the original spatial resolution to the same level as the temporal data (SAR temporal image) and serve as intermediate feature maps. SAR is an abbreviation for Synthetic Aperture Radar.

[0035] The technical solution provided by the above steps first compresses the temporal-spatial information to extract deep features, and then restores the spatial resolution to maintain positioning accuracy. While maintaining the original spatial resolution, the original SAR temporal data (T×C×H×W) is converted into a feature representation with preliminary semantic expression, laying the foundation for subsequent multi-scale processing.

[0036] For example, traditional pixel-level CNNs use a single-scale receptive field, which cannot simultaneously adapt to the scale differences between fragmented small plots and contiguous large plots. This solution addresses this limitation by constructing a multi-scale pyramid through multi-level downsampling. Standard U-Net lacks temporal modeling capabilities; this solution introduces LTAE to implement temporal attention. Although 3D-CNNs can process temporal data, their computational complexity is high and they struggle with long temporal sequences. This solution performs temporal attention at compressed resolution, improving efficiency.

[0037] The U-TAE encoder-decoder structure also includes the following settings: Replace the zero-padding method of the convolutional layer with reflection padding to suppress edge artifacts; A Dropout2d layer is inserted before the output convolutional layer, and its dropout rate is set to 0.2 to suppress network overfitting. It provides three switchable time series aggregation modes, including attention grouping mode, attention mean mode, and mean mode, and dynamically selects the time series aggregation mode as the time series smoothing strategy according to the data sparsity of the input time series data.

[0038] S103, the parallel spatiotemporal self-attention module is used to enhance the temporal and spatial features of the features output by the encoding stage of the backbone network to generate fused features; Specifically, the parallel spatiotemporal self-attention module includes temporal branches and spatial branches; step S103 includes: S301, the temporal branch is used to process the feature tensor output by the encoder, extract vectors for each spatial location, rearrange them and then perform weighted aggregation through a 16-head self-attention mechanism, and the multi-head outputs are spliced ​​together and mapped through two layers of multilayer perceptron to obtain temporal features; S302, using the spatial branch, the features are first averaged along the temporal dimension to suppress speckle noise of synthetic aperture radar. Then, the features are projected to 128 dimensions through group normalization and 1×1 convolution, and learnable row direction position codes and column direction position codes are injected. After flattening, spatial weights are calculated through a 4-head self-attention mechanism to obtain spatial feature representation. S303, the temporal feature representation and the spatial feature representation are concatenated along the channel dimension, compressed to 128 channels by 1×1 convolution, processed by the GELU activation function and Dropout2d regularization, and then added to the temporal feature representation by element-wise residual addition to generate fused features.

[0039] The STSA-LTAE module breaks away from the limitations of traditional attention mechanisms that focus on only a single dimension. Through the parallel design of temporal and spatial branches, it achieves joint modeling of crop growth temporal dynamics and spatial relationships between pixels. The two branches independently process different dimensions of features, and finally generate fused features that combine temporal discriminative power and spatial consistency through feature fusion.

[0040] Unlike typical approaches that separate temporal modeling from spatial processing or apply attention mechanisms only in a single dimension, STSA-LTAE's parallel architecture enhances attention in both dimensions simultaneously, creating a synergistic effect. The temporal branch captures the annual dynamics of crop growth, distinguishing phenological differences among different crops; the spatial branch establishes relationships between pixels, improving the ability to differentiate crops with similar temporal features but different spatial distributions. The fusion of these two branches ensures that the final representation of each pixel includes both its temporal evolution characteristics and its neighborhood spatial context, resulting in a more discriminative feature representation.

[0041] S104, using the parameterless parcel pooling operator, the fusion features are aggregated from pixel level to parcel level representation based on the field ID map; Specifically, S104 includes: S401, the fused features are reshaped into a two-dimensional feature matrix, and the field ID map is flattened into a field index vector; S402, for all pixels in the parcel index vector marked as valid parcel identifiers, perform a parallel scattering accumulation operation to accumulate the sum of the feature vectors corresponding to each parcel identifier and the pixel count respectively; S403, based on the sum of the feature vectors and the pixel count, a stabilization term is introduced to calculate the average feature vector of each plot, and a plot embedding matrix is ​​generated as the plot-level representation.

[0042] The ParcelPooling operator contains no learnable parameters, making it a pure feature transformation operation that does not increase model complexity and avoids the risk of overfitting. Both scattering accumulation and mean calculation are differentiable operations, supporting gradient backpropagation, enabling the entire framework to be optimized from plot-level classification errors down to the pixel-level feature extraction network.

[0043] S105, using a dual-branch prediction head, perform plot-level classification on the plot-level representation, and refine the pixel-level boundaries based on the intermediate feature map; broadcast the classification results of the plot-level classification back to the pixel position corresponding to the plot according to the field ID map, and output a plot-level crop type classification map with consistent spatial boundaries.

[0044] Specifically, the dual-branch prediction head shares the backbone network parameters, including the land parcel branch and the pixel branch; step S105 includes: S501, the plot branch receives the plot embedding matrix represented by the plot level, and performs nonlinear mapping through a two-layer fully connected network. The first layer maps the plot features to the embedding dimension space and activates them through ReLU. The second layer maps them to the crop category number dimension and generates the crop category probabilities of each plot level through softmax normalization. S502, the pixel branch applies a 1×1 convolution on the intermediate feature map output by the decoder to generate a pixel-level classification score for boundary detail refinement; S503, the dual-branch prediction head is jointly optimized and trained using a weighted joint loss function. The weighted joint loss function includes a weighted sum of plot-level focus loss and pixel-level focus loss. The weighted joint loss function adopts a balancing strategy of inverse frequency class weights to mitigate classification bias caused by the long-tail distribution of crop categories.

[0045] In summary, this technical solution captures long-range temporal dependencies while maintaining the original spatial resolution through a U-TAE encoder-decoder structure. Combined with a parallel spatiotemporal self-attention module to jointly enhance temporal and spatial features, it effectively solves the difficulty of feature extraction caused by small plot areas and complex temporal data in fragmented farmland. This achieves synchronous and accurate perception of fragmented farmland scenes (e.g., small plots less than 3×3 pixels and large contiguous plots) in highly fragmented landscapes, improving crop classification accuracy. Furthermore, a parameter-free plot aggregation operator smoothly transitions pixel-level features to plot-level representations, and a dual-branch prediction head is used for collaborative optimization of plot-level classification and pixel-level boundary refinement. This architecture enforces consistency of all pixel labels within a plot, eliminating the "salt and pepper" noise generated by traditional pixel-level methods. Simultaneously, a reverse broadcast mechanism ensures precise alignment between the output classification map and the pre-defined plot boundaries. This approach ensures that the classification results not only conform to the characteristics of plot units in agricultural management practice but also maintain fine spatial boundary details. The entire framework, through a fully differentiable end-to-end design, unifies multi-scale feature extraction, spatiotemporal attention enhancement, plot-level aggregation, and dual-classification branching into a joint optimization system. This avoids the error propagation problem caused by the separation of segmentation and classification in traditional object-oriented analysis methods, enabling the network to automatically learn the optimal feature representation to adapt to complex fragmented scenarios. Addressing the speckle noise characteristics of SAR data, the scheme constructs a multi-level noise suppression mechanism by combining temporal averaging suppression of spatial branches with statistical denoising of plot-level mean aggregation, significantly improving the model's robustness under complex environmental conditions. Furthermore, the computational design of lightweight temporal self-attention at compressed resolution and the linear complexity of the plot aggregation operator enable efficient, consistent, and end-to-end generation from raw images to plot-level classification results.

[0046] As an example, a target-oriented plot-level crop classification method for fragmented farmland is provided, which includes the following steps: Step 1: Data Preparation and Preprocessing In the region to be classified, a multi-temporal Sentinel-1 synthetic aperture radar image sequence (multi-temporal remote sensing image sequence) and pre-delineated field boundary data were acquired. The SAR images in the multi-temporal remote sensing image sequence are dual-polarization GRD images, containing two polarization channels, VV and VH, with a spatial resolution of 10 meters and a temporal resolution of 12-24 days. The SAR temporal images are organized into a four-dimensional tensor X∈R^(T×C×H×W), where T is the number of timestamps (temporal phases), C is the number of polarization channels, and H×W is the spatial resolution. Field boundary data were obtained through automatic extraction from high-resolution images combined with manual annotation and correction. After rasterization, the data was aligned with the SAR temporal grid to form a field ID map P∈N^(H×W), which was used for subsequent plot-level feature aggregation.

[0047] Step 2: Multi-scale feature extraction of backbone network A U-TAE encoder-decoder backbone network is employed to extract multi-scale features from SAR temporal data in multi-temporal remote sensing image sequences. This network combines the multi-scale convolutional framework of U-Net with the lightweight temporal self-attention module LTAE, capturing long-range temporal dependencies while maintaining the original spatial resolution.

[0048] The specific processing procedure is as follows: S2.1: The input convolutional layer (inConv) consists of two 3×3 convolutions, normalization, and ReLU activation to extract shallow semantic features; S2.2: The three downsampled residual blocks gradually reduce the spatial size to 1 / 2×1 / 2, 1 / 4×1 / 4 and 1 / 8×1 / 8 through convolution, while expanding the number of channels to 64, 64 and 128; S2.3: Apply LTAE for temporal weighted aggregation in the final coding layer; S2.4: The three upsampling blocks use transposed convolution and skip connections to restore spatial resolution.

[0049] The backbone network employs the following improvements to adapt to the characteristics of radar data: replacing zero padding with reflection padding to mitigate edge artifacts; inserting a Dropout2d(0.2) layer before the output convolution to suppress overfitting; and providing three switchable temporal aggregation modes (att_group, att_mean, and mean, i.e., attention grouping mode, attention mean mode, and mean mode) to dynamically smooth features based on data sparsity.

[0050] Step 3: Parallel Spatiotemporal Self-Attention Feature Enhancement In the final stage of the encoder, a parallel spatiotemporal self-attention module (STSA-LTAE) is introduced. This module contains a temporal branch LTAE and a spatial branch STSA, which enhances both temporal and spatial feature representations.

[0051] (a) Timing branch (LTAE): For the encoder output tensor F∈R^(B×T×C×H×W), a vector of length T is extracted for each spatial location (h,w). After rearrangement, a temporally weighted aggregation is performed using a 16-head attention mechanism. The formula for calculating the attention weights is:

[0052] (1) Where q is the learnable query vector, and k t and v t d represents the key and value of frame t. k The key vector dimension is used. The multi-head outputs are concatenated and passed through a two-layer MLP to obtain the temporal feature Z.time ∈R^(B×C'×H×W).

[0053] (b) Spatial Branching (STSA): First, average along the time dimension to suppress speckle noise, and obtain F. ∈R^(B×C×H×W). Then, GroupNorm and 1×1 convolutional projection features are applied to d_model=128 dimensions to inject learnable row and column position encodings: (2) Where p row ∈R^(1×d / 2×H×1) and p col ∈R^(1×d / 2×1×W) represents the learnable row and column position encoding. After flattening, spatial weights are calculated using 4-head multi-head self-attention to obtain the spatial feature Z. space ∈R^(B×128×H×W).

[0054] (c) Feature fusion: The two branch outputs are concatenated along the channel dimension and compressed to 128 channels through a 1×1 convolution. After GELU activation and Dropout2d(0.2) processing, they are added element-wise with the temporal output to obtain the residual, generating the fused feature Z. fused This is for use in subsequent land parcel-level aggregation.

[0055] Step 4: Parcel-level feature aggregation (ParcelPooling) The parameterless ParcelPooling operator is used to aggregate pixel-level features into plot-level representations based on a pre-rasterized plot ID map. This operator has no additional learnable parameters, is fully differentiable, and supports end-to-end training.

[0056] The specific operating procedure is as follows: S4.1: Reshape the feature map F into X∈R (N×C) Where N = H × W, and the flattened field index vector p ∈ N^N; S4.2: For all p i For valid pixels i with a value greater than 0, perform a parallel scatter-add operation to accumulate the feature vector and pixel count within each field: (3) S4.3: Calculate the eigenvector of the land parcel mean with stability term: (4) Where P represents the total number of land parcels, and ε is the stability term. The final land parcel embedding matrix V∈R is obtained. (P×C) .

[0057] The entire process runs on the GPU with O(N) computational complexity, and memory usage is linearly related to the number of pixels, supporting large-scale scene processing with hundreds of millions of pixels. Unlike ROI-Pooling, ParcelPooling accurately averages irregular plot shapes, preserving the true geometry of farmland.

[0058] Step 5: Bi-branch prediction and joint optimization A dual-branch prediction head is used for parcel-level classification and pixel-level boundary refinement. The two branches share backbone network parameters but focus on information at different spatial scales.

[0059] (a) Parcel Branches: Let the land parcel feature matrix V∈R (P×C) Send to a two-layer fully connected network: (5) Where C emb =256 is the embedding dimension, N cls Let be the number of crop categories, and σ be the ReLU activation function. Apply softmax to each plot k to generate plot-level category probabilities: (6) (b) Pixel branch: The feature map F∈R is directly recovered in the decoder (C_feat×H×W) Applying a single 1×1 convolution to produce pixel-level logits G∈R (B×N_cls×H×W) It is used to assist in boundary refinement.

[0060] (c) Joint optimization: The two branches are jointly optimized and trained using weighted Focal Loss: (7) Where λ pix The pixel loss weights are used. Weighted Focal Loss uses inverse frequency-class weights α. c With the focusing parameter γ, the problem of long-tailed distribution of crop categories can be effectively addressed.

[0061] Step 6: Output classification results The classification results of the plot branches are back-broadcast to the corresponding pixel positions based on the plot ID map, outputting a plot-level crop type classification map with consistent spatial boundaries. All pixels within each plot in the classification map have the same crop type label, ensuring spatial consistency and boundary clarity.

[0062] As another example, to further clarify the content of this invention, the following is combined with... Figure 1The structural process and specific examples of the invention further illustrate the detailed description of the invention. It should be understood that not all features of all actual embodiments are the same as those in this example, and the implementation details of the invention may vary in actual engineering projects depending on specific conditions and objectives.

[0063] In this example, crop classification in Fuyang District, Hangzhou City, Zhejiang Province is selected to further describe the technical solution of the present invention in detail: Step 1: Data Preparation The study used Sentinel-1 synthetic aperture radar time-series data from April 2024 to April 2025 for Fuyang District, Hangzhou City, Zhejiang Province. The study area is located at 29°45′N~30°10′N, 119°25′E~120°15′E. Figure 2 As shown, it is located in the hilly area of ​​northwestern Zhejiang in the middle and upper reaches of the Qiantang River. It has a subtropical monsoon climate, which is warm and humid with abundant heat and rainfall, forming a typical Jiangnan rice-growing system. Rice-rapeseed (or rice-wheat) double-cropping is carried out throughout the year.

[0064] Acquire dual-polarized Sentinel-1A / B GRD images (VV+VH). After removing severely polluted time points, the temporal resolution is 12-24 days, and the spatial resolution is 10 meters. All images are processed into a 4D SAR temporal tensor X∈R. (T×C×H×W) T≈20, C=2, H=W=128. Field boundaries were obtained using a workflow of 'automatic extraction from high-resolution images + manual annotation and correction': first, the EDTER model was used to extract field boundaries from Google Earth 0.5m resolution images; then, MMSegmentation was used to generate pixel-level semantic segmentation and vectorization; finally, crop type was assigned to each vectorized plot based on field survey notes from 2024-2025 and agricultural census data.

[0065] The dataset contains five crop categories: double-cropping rice (2159 plots), rice-wheat rotation (9725 plots), rice-rapeseed rotation (8919 plots), tea (13250 plots), and others (31250 plots). Image patches of size 128×128 are divided into training, validation, and test sets in a 3:1:1 ratio. Fragmented plots (regions occupying less than 4×4 pixels in Sentinel-1 imagery) comprise approximately 70% of the Fuyang dataset.

[0066] Step 2: Backbone Network Feature Extraction U-TAE was used as the backbone network. First, shallow semantic features were extracted using inConv (composed of two 3×3 convolutions, GroupNorm normalization, and ReLU). Then, the spatial size was progressively reduced to 1 / 2×1 / 2, 1 / 4×1 / 4, and 1 / 8×1 / 8 using convolutions through three downsampled residual blocks, while simultaneously expanding the number of channels to 64, 64, and 128. Reflection padding was used instead of zero padding in the backbone network, and a Dropout2d(0.2) layer was inserted before the output convolution. The temporal aggregation mode was selected based on the data sparsity: att_group mode was used when the temporal coverage exceeded 75%; att_mean mode was used when it was between 50% and 75%; and mean mode was used when it was below 50%.

[0067] Step 3: Parallel Spatiotemporal Self-Attention Enhancement like Figure 3 As shown, a parallel STSA-LTAE module is introduced in the final stage of the encoder. The temporal branch extracts a vector of length T for each spatial location, rearranges it, and then calculates the temporal attention weights using a 16-head LTAE and aggregates them with weights. The spatial branch first averages along the temporal dimension to suppress speckle noise, then applies four sets of GroupNorm and 1×1 convolutional projection features to d_model=128 dimensions, injects learnable row and column position encodings, flattens it, and applies 4-head multi-head self-attention to calculate spatial weights. The outputs of the two branches are concatenated along the channel dimension and compressed by a 1×1 convolution. After GELU activation and Dropout2d(0.2) processing, they are added element-wise with the temporal output as the residual.

[0068] Step 4: Parcel Pooling like Figure 4 As shown, the parameterless ParcelPooling operator is used for pixel-to-plot feature aggregation. The feature map F is reshaped into X∈R^(N×C), and the corresponding plot index vector p∈N^N is flattened. For all p i For valid pixels i with values ​​greater than 0, a parallel scatter-add operation is performed to accumulate the feature vector and pixel count, and then the average feature vector of the plot is calculated. The entire process runs on the GPU with a computational complexity of O(N), supporting large-scale scene processing containing hundreds of millions of pixels.

[0069] Step 5: Two-branch prediction and training The parcel branch will divide the parcel feature matrix V∈R (P×C)The data is fed into a two-layer fully connected network (embedding dimension 256), where softmax is applied to each plot to generate plot-level class probabilities. The pixel branch directly applies a single 1×1 convolution to the feature map recovered by the decoder to generate pixel-level logits. The two branches are jointly optimized using weighted Focal Loss, where the class weights are constructed using inverse frequency and a smoothing exponent β=0.25 is set, the focusing parameter γ=2, and the pixel loss weight λ... pix =0.3.

[0070] The AdamW optimizer was used with an initial learning rate of 2 × 10^(-4), which was decayed to 1 × 10^(-6) by cosine annealing, and training was performed for 100 epochs. To prevent gradient explosion, the gradient norm of all trainable parameters was clipped to 5. Data augmentation included random horizontal and vertical flips, 90° rotations, and affine perturbations with an amplitude of 0.1. During training, the plot-level F1 score on the validation set was monitored, and training was stopped early if there was no improvement for 10 consecutive epochs.

[0071] Step 6: Results Output and Evaluation Output a plot-level crop type classification map. To ensure objectivity, only test set data is used for prediction. The model prediction results are as follows: Figure 5 As shown, Figure 5 (a) shows the classification results of fragmented land parcels. Figure 5 (b) shows the classification results for regularized plots. The method of this invention achieves an overall accuracy (OA) of 81.7%, a mean F1 score (mF1) of 65.1%, and a mean intersection-union ratio (mIoU) of 51.3% on the Fuyang dataset. Compared to the best baseline method, the overall accuracy is improved by 4.2%, and the mean F1 score is improved by 2.5%.

[0072] The F1 scores for each category are as follows: double-cropping rice 0.635, rice-wheat rotation 0.778, rice-rapeseed rotation 0.778, tea 0.638, and others 0.666. Ablation experiments show that removing the parallel STSA module decreased OA by 0.8%, removing ParcelPooling decreased OA by 2.4%, removing pixel-assisted branches decreased OA by 1.1%, and removing the category-weighted strategy decreased OA by 0.5%, demonstrating the effectiveness of each component.

[0073] The technical solution provided in this embodiment has the following advantages and beneficial effects: (a) High classification accuracy: Achieves an overall accuracy (OA) of 81.7%, a mean F1 score (mF1) of 65.1%, and a mean crossover ratio (mIoU) of 51.3% in highly fragmented agricultural landscapes, outperforming existing methods.

[0074] (b) Good spatial consistency: Through ParcelPooling and plot-level supervision, the 'salt and pepper' noise generated by traditional pixel-level methods is effectively eliminated, ensuring the consistency of classification results within plots.

[0075] (c) Accurate boundary details: The joint optimization design of the dual-branch prediction head improves the accuracy of boundary details while ensuring the accuracy at the plot level.

[0076] (d) High computational efficiency: The ParcelPooling operator operates with O(N) computational complexity, and memory usage is linearly controllable, supporting efficient processing of large-scale agricultural monitoring scenarios.

[0077] (e) End-to-end training: The entire framework is fully differentiable, supporting end-to-end learning from raw SAR images to parcel-level classification, avoiding error propagation caused by segmentation and classification separation in traditional methods.

[0078] (f) Strong noise resistance: The combination of parallel spatiotemporal self-attention and ParcelPooling effectively suppresses SAR speckle noise and improves robustness in complex environments.

[0079] Example 2.

[0080] This embodiment provides an electronic device, which includes a memory and at least one processor. The memory stores a computer program, and the at least one processor is configured to execute the computer program to perform the following steps: S101, acquire the multi-temporal remote sensing image sequence of the area to be classified and the pre-delineated field boundary data as input information; the field boundary raster data is a field ID map that is spatiotemporally aligned with the multi-temporal remote sensing image sequence; S102, a backbone network with a U-TAE encoder-decoder structure is used to extract multi-scale features from the temporal data of the multi-temporal remote sensing image sequence in the input information to obtain intermediate feature maps; the U-TAE encoder-decoder structure includes a U-Net multi-scale convolutional framework and a lightweight temporal self-attention module LTAE, which is used to capture long-range temporal dependencies while maintaining the original spatial resolution. S103, the parallel spatiotemporal self-attention module is used to enhance the temporal and spatial features of the features output by the encoding stage of the backbone network to generate fused features; S104, using a parameterless plot aggregation operator, the fusion features are aggregated from pixel level to plot level representation based on the plot ID map; S105, using a dual-branch prediction head, perform plot-level classification on the plot-level representation, and refine the pixel-level boundaries based on the intermediate feature map; broadcast the classification results of the plot-level classification back to the pixel position corresponding to the plot according to the field ID map, and output a plot-level crop type classification map with consistent spatial boundaries.

[0081] In one exemplary embodiment, the multi-scale feature extraction includes the following steps: The temporal data is subjected to shallow semantic feature extraction by an input convolutional layer, which includes two 3×3 convolutional units connected in sequence, a normalization layer and a ReLU activation function. The shallow semantic features are progressively downsampled and expanded through three cascaded downsampled residual blocks; each downsampled residual block reduces the spatial resolution of the feature map to 1 / 2×1 / 2, 1 / 4×1 / 4 and 1 / 8×1 / 8 through convolution operations, while expanding the number of feature channels to 64, 64 and 128 respectively. Lightweight temporal self-attention module LTAE is applied to the end layer of the encoder to perform temporal weighted aggregation of the downsampled features; Spatial resolution restoration is performed on the temporally weighted aggregated features by three cascaded upsampling blocks. The upsampling blocks employ transposed convolution and skip connection structures to restore the original spatial resolution to the same as the temporal data (SAR temporal image) and serve as intermediate feature maps.

[0082] In one exemplary embodiment, the U-TAE encoder-decoder structure further includes the following configuration: Replace the zero-padding method of the convolutional layer with reflection padding to suppress edge artifacts; A Dropout2d layer is inserted before the output convolutional layer, and its dropout rate is set to 0.2 to suppress network overfitting. It provides three switchable time series aggregation modes, including attention grouping mode, attention mean mode, and mean mode, and dynamically selects the time series aggregation mode as the time series smoothing strategy according to the data sparsity of the input time series data.

[0083] In an exemplary embodiment, the parallel spatiotemporal self-attention module includes a temporal branch and a spatial branch; step S103 includes: The temporal branch is used to process the feature tensor output by the encoder, extract vectors for each spatial location, rearrange them, and then perform weighted aggregation through a 16-head self-attention mechanism. The multi-head outputs are then concatenated and mapped through a two-layer multilayer perceptron to obtain temporal features. Using the aforementioned spatial branch, the features are first averaged along the temporal dimension to suppress speckle noise in synthetic aperture radar. Then, the features are projected to 128 dimensions through group normalization and 1×1 convolution, and learnable row and column position codes are injected. After flattening, spatial weights are calculated through a 4-head self-attention mechanism to obtain the spatial feature representation. The temporal feature representation and the spatial feature representation are concatenated along the channel dimension, compressed to 128 channels by 1×1 convolution, processed by the GELU activation function and Dropout2d regularization, and then added to the temporal feature representation by element-wise residual addition to generate fused features.

[0084] In an exemplary embodiment, step S104 includes: The fused features are reshaped into a two-dimensional feature matrix, and the field ID map is flattened into a field index vector; For all pixels in the plot index vector that are marked as valid plot identifiers, a parallel scattering accumulation operation is performed to accumulate the sum of the feature vectors corresponding to each plot identifier and the pixel count, respectively. Based on the sum of the feature vectors and the pixel count, a stabilization term is introduced to calculate the average feature vector of each plot, generating a plot embedding matrix as the plot-level representation.

[0085] In an exemplary embodiment, step S105 includes: The dual-branch prediction head shares the backbone network parameters, including the land parcel branch and the pixel branch; The plot branch receives the plot embedding matrix represented by the plot level and performs nonlinear mapping through a two-layer fully connected network. The first layer maps the plot features to the embedding dimension space and activates them with ReLU. The second layer maps them to the crop category number dimension and generates the crop category probabilities of each plot level through softmax normalization. The pixel branch applies a 1×1 convolution to the intermediate feature map output by the decoder to generate a pixel-level classification score, which is used for boundary detail refinement. The dual-branch prediction head is jointly optimized and trained using a weighted joint loss function, which includes a weighted sum of plot-level focus loss and pixel-level focus loss. The weighted joint loss function employs a balancing strategy of inverse frequency class weights to mitigate classification bias caused by the long-tail distribution of crop categories.

[0086] Example 3.

[0087] This embodiment provides a computer-readable storage medium, the specific embodiments of which are consistent with the embodiments described in the above method embodiments and the technical effects achieved, and some contents will not be repeated.

[0088] The computer-readable storage medium stores a computer program that, when executed by at least one processor, implements the steps of any of the above methods.

[0089] A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. In embodiments of this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0090] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable storage medium may also be any computer-readable medium capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, or any suitable combination thereof. Program code for performing operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar programming languages. The program code may be executed entirely on a user computing device, partially on a user device, as a standalone software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to user computing devices via any type of network, including local area networks (LANs) or wide area networks (WANs), or they can be connected to external computing devices (e.g., via the Internet using an Internet service provider).

[0091] Example 4.

[0092] This application also provides a computer program product, the specific embodiments of which are consistent with the embodiments and technical effects achieved in the above method embodiments, and some contents will not be repeated.

[0093] This application provides a computer program product comprising a computer program that, when executed by at least one processor, implements the steps of any of the above methods.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A target-oriented crop plot-level classification method for fragmented farmland, characterized in that, Includes the following steps: S101, acquire the multi-temporal remote sensing image sequence of the area to be classified and the pre-delineated field boundary data as input information; the field boundary raster data is a field ID map that is spatiotemporally aligned with the multi-temporal remote sensing image sequence; S102, a backbone network with a U-TAE encoder-decoder structure is used to extract multi-scale features from the temporal data of the multi-temporal remote sensing image sequence in the input information to obtain intermediate feature maps; the U-TAE encoder-decoder structure includes a U-Net multi-scale convolutional framework and a lightweight temporal self-attention module, which is used to capture long-range temporal dependencies while maintaining the original spatial resolution. S103, the parallel spatiotemporal self-attention module is used to enhance the temporal and spatial features of the features output by the encoding stage of the backbone network to generate fused features; S104, using a parameterless plot aggregation operator, the fusion features are aggregated from pixel level to plot level representation based on the plot ID map; S105, using a dual-branch prediction head, perform plot-level classification on the plot-level representation, and refine the pixel-level boundaries based on the intermediate feature map; broadcast the classification results of the plot-level classification back to the pixel position corresponding to the plot according to the field ID map, and output a plot-level crop type classification map with consistent spatial boundaries.

2. The target-oriented crop plot-level classification method according to claim 1, characterized in that, The multi-scale feature extraction includes the following steps: The temporal data is subjected to shallow semantic feature extraction by an input convolutional layer, which includes two 3×3 convolutional units connected in sequence, a normalization layer and a ReLU activation function. The shallow semantic features are progressively downsampled and expanded through three cascaded downsampled residual blocks; each downsampled residual block reduces the spatial resolution of the feature map to 1 / 2×1 / 2, 1 / 4×1 / 4 and 1 / 8×1 / 8 through convolution operations, while expanding the number of feature channels to 64, 64 and 128 respectively. A lightweight temporal self-attention module is applied to the end layer of the encoder to perform temporal weighted aggregation of the downsampled features; The spatial resolution of the temporally weighted aggregated features is restored by using three cascaded upsampling blocks. The upsampling blocks employ transposed convolution and skip connection structures to restore the original spatial resolution to the same as the temporal data and serve as intermediate feature maps.

3. The target-oriented crop plot-level classification method according to claim 2, characterized in that, The U-TAE encoder-decoder structure also includes the following settings: Replace the zero-padding method of the convolutional layer with reflection padding to suppress edge artifacts; A Dropout2d layer is inserted before the output convolutional layer, and its dropout rate is set to 0.2 to suppress network overfitting. It provides three switchable time series aggregation modes, including attention grouping mode, attention mean mode, and mean mode, and dynamically selects the time series aggregation mode as the time series smoothing strategy according to the data sparsity of the input time series data.

4. The target-oriented crop plot-level classification method according to claim 1, characterized in that, The parallel spatiotemporal self-attention module includes temporal branches and spatial branches; step S103 includes: The temporal branch is used to process the feature tensor output by the encoder, extract vectors for each spatial location, rearrange them, and then perform weighted aggregation through a 16-head self-attention mechanism. The multi-head outputs are then concatenated and mapped through a two-layer multilayer perceptron to obtain temporal features. Using the aforementioned spatial branch, the features are first averaged along the temporal dimension to suppress speckle noise in synthetic aperture radar. Then, the features are projected to 128 dimensions through group normalization and 1×1 convolution, and learnable row and column position codes are injected. After flattening, spatial weights are calculated through a 4-head self-attention mechanism to obtain the spatial feature representation. The temporal feature representation and the spatial feature representation are concatenated along the channel dimension, compressed to 128 channels by 1×1 convolution, processed by the GELU activation function and Dropout2d regularization, and then added to the temporal feature representation by element-wise residual addition to generate fused features.

5. The target-oriented crop plot-level classification method according to claim 1, characterized in that, Step S104 includes: The fused features are reshaped into a two-dimensional feature matrix, and the field ID map is flattened into a field index vector; For all pixels in the plot index vector that are marked as valid plot identifiers, a parallel scattering accumulation operation is performed to accumulate the sum of the feature vectors corresponding to each plot identifier and the pixel count, respectively. Based on the sum of the feature vectors and the pixel count, a stabilization term is introduced to calculate the average feature vector of each plot, generating a plot embedding matrix as the plot-level representation.

6. The target-oriented crop plot-level classification method according to claim 5, characterized in that, Step S105 includes: The dual-branch prediction head shares the backbone network parameters, including the land parcel branch and the pixel branch; The plot branch receives the plot embedding matrix represented by the plot level and performs nonlinear mapping through a two-layer fully connected network. The first layer maps the plot features to the embedding dimension space and activates them with ReLU. The second layer maps them to the crop category number dimension and generates the crop category probabilities of each plot level through softmax normalization. The pixel branch applies a 1×1 convolution to the intermediate feature map output by the decoder to generate a pixel-level classification score, which is used for boundary detail refinement. The dual-branch prediction head is jointly optimized and trained using a weighted joint loss function, which includes a weighted sum of plot-level focus loss and pixel-level focus loss. The weighted joint loss function employs a balancing strategy of inverse frequency class weights to mitigate classification bias caused by the long-tail distribution of crop categories.

7. An electronic device, characterized in that, The electronic device includes a memory and at least one processor, the memory storing a computer program, and the at least one processor being configured to execute the computer program to perform the following steps: S101, acquire the multi-temporal remote sensing image sequence of the area to be classified and the pre-delineated field boundary data as input information; the field boundary raster data is a field ID map that is spatiotemporally aligned with the multi-temporal remote sensing image sequence; S102, a backbone network with a U-TAE encoder-decoder structure is used to extract multi-scale features from the temporal data of the multi-temporal remote sensing image sequence in the input information to obtain intermediate feature maps; the U-TAE encoder-decoder structure includes a U-Net multi-scale convolutional framework and a lightweight temporal self-attention module, which is used to capture long-range temporal dependencies while maintaining the original spatial resolution. S103, the parallel spatiotemporal self-attention module is used to enhance the temporal and spatial features of the features output by the encoding stage of the backbone network to generate fused features; S104, using a parameterless plot aggregation operator, the fusion features are aggregated from pixel level to plot level representation based on the plot ID map; S105, using a dual-branch prediction head, perform plot-level classification on the plot-level representation, and refine the pixel-level boundaries based on the intermediate feature map; broadcast the classification results of the plot-level classification back to the pixel position corresponding to the plot according to the field ID map, and output a plot-level crop type classification map with consistent spatial boundaries.

8. The electronic device according to claim 7, characterized in that, The multi-scale feature extraction includes the following steps: The temporal data is subjected to shallow semantic feature extraction by an input convolutional layer, which includes two 3×3 convolutional units connected in sequence, a normalization layer and a ReLU activation function. The shallow semantic features are progressively downsampled and expanded through three cascaded downsampled residual blocks; each downsampled residual block reduces the spatial resolution of the feature map to 1 / 2×1 / 2, 1 / 4×1 / 4 and 1 / 8×1 / 8 through convolution operations, while expanding the number of feature channels to 64, 64 and 128 respectively. A lightweight temporal self-attention module is applied to the end layer of the encoder to perform temporal weighted aggregation of the downsampled features; The spatial resolution of the temporally weighted aggregated features is restored by using three cascaded upsampling blocks. The upsampling blocks employ transposed convolution and skip connection structures to restore the original spatial resolution to the same as the temporal data and serve as intermediate feature maps.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by at least one processor, implements the steps of the method according to any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by at least one processor, implements the steps of the method according to any one of claims 1-6.

Images