A self-adaptive crop recognition method and system for difference in phenological phase across regions

Abstract

The application discloses a kind of self-adapting phenological period difference's cross-region crop identification method and system, it is related to remote sensing image processing technical field, self-adapting phenological period difference's cross-region crop identification method mainly includes: according to low-resolution crop label, vegetation index time series data calculates the time offset between source area and target area;Low-resolution crop label is super-resolution and is handled to obtain high-resolution pseudo label;Multi-temporal high-resolution remote sensing image sequence data of source area is handled to phenology alignment, and high-resolution pseudo label is combined to neural network model is trained to obtain crop identification model, and the image sequence of target area is predicted and segmented using crop identification model to generate high-resolution crop distribution map.The self-adapting phenological period difference's cross-region crop identification method and system provided by the application can improve the accuracy and robustness of crop identification in the unlabeled, cross-region scene.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and more specifically, to a method and system for cross-regional crop identification based on adaptive phenological differences. Background Technology

[0002] Refined crop identification and area monitoring are key technological links in ensuring national food security and guiding precision agricultural production. With the development of remote sensing technology, high-resolution multi-temporal remote sensing images provide a rich data foundation for the dynamic monitoring of crop growth processes.

[0003] Deep learning methods have demonstrated superior performance in crop identification tasks. However, these models typically rely on a large number of pixel-level precisely labeled training samples. In practical applications, creating a high-quality labeled dataset covering the entire growing season across vast agricultural areas is costly, time-consuming, and labor-intensive, hindering the widespread adoption of this technology.

[0004] To address the label scarcity problem, a common approach is to train a model using data from a region with existing high-quality labels (hereinafter referred to as the source domain) and then apply it to a region lacking labels (hereinafter referred to as the target domain). However, cross-regional application in crop recognition scenarios faces the following challenges: 1. Crop planting patterns, soil background, field morphology, and imaging lighting conditions vary across different regions, resulting in inconsistent visual features in remote sensing images. When a model trained in the source domain is directly applied to the target domain, its performance will significantly decrease due to these spatial feature differences.

[0005] 2. The remote sensing image features of crops change dynamically throughout their growth cycle, including sowing, emergence, vigorous growth, maturity, and harvest. Due to differences in latitude, climate, variety, and management practices, key phenological nodes for the same crop may occur on different dates in different geographical regions. This asynchrony in phenology causes the time-feature mapping relationship learned by the model in the source domain to fail in the target domain, resulting in poor generalization ability of cross-regional crop identification models.

[0006] Therefore, designing a deep learning method that can overcome the dual challenges of spatial visual feature differences and temporal phenological rhythm differences in the target area when there are no local labels, and achieving high-precision cross-regional crop identification, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to provide a cross-regional crop identification method and system that adapts to phenological differences, which can improve the accuracy and robustness of crop identification in label-free, cross-regional scenarios.

[0008] This invention provides a cross-regional crop identification method based on adaptive phenological differences, comprising the following steps: S1: Acquire low-resolution crop tags for the source region, multi-temporal high-resolution remote sensing image sequence data of the source and target regions, and vegetation index time series data; S2: Calculate the time offset between the source region and the target region based on the low-resolution crop tags and vegetation index time series data; S3: Use the first neural network model to perform super-resolution processing on the low-resolution crop tags to obtain high-resolution pseudo-tags that match the multi-temporal high-resolution remote sensing image sequence data of the source region. S4: Based on the time offset between the source region and the target region, perform phenological alignment processing on the multi-temporal high-resolution remote sensing image sequence data of the source region to obtain the phenologically aligned source region image sequence; use the high-resolution pseudo-label and the phenologically aligned source region image sequence to train the second neural network model to obtain the crop recognition model; S5: Based on the time offset between the source region and the target region, select an image sequence from the multi-temporal high-resolution remote sensing image sequence data of the target region that matches the phenological period of the source region image sequence after phenological alignment, obtain the selected image sequence, and use the crop recognition model to predict and segment the selected image sequence to generate a high-resolution crop distribution map of the target region.

[0009] The present invention also provides a cross-regional crop identification system that adapts to phenological differences, the system comprising the following modules: The data acquisition module is configured to acquire low-resolution crop tags from the source region, multi-temporal high-resolution remote sensing image sequence data of the source and target regions, and vegetation index time series data. The phenological period correction module is configured to calculate the time offset between the source region and the target region based on the low-resolution crop tags and vegetation index time series data. The pseudo-label generation module is configured to: use a first neural network model to perform super-resolution processing on the low-resolution crop labels to obtain high-resolution pseudo-labels that match the multi-temporal high-resolution remote sensing image sequence data of the source region; The model training module is configured to: perform phenological alignment processing on the multi-temporal high-resolution remote sensing image sequence data of the source region according to the time offset between the source region and the target region, to obtain the phenologically aligned source region image sequence; and train the second neural network model using the high-resolution pseudo-label and the phenologically aligned source region image sequence to obtain the crop recognition model. The crop identification application module is configured to: select an image sequence from the multi-temporal high-resolution remote sensing image sequence data of the target area that matches the phenological period of the source area image sequence after phenological alignment, based on the time offset between the source area and the target area, obtain the selected image sequence, and use the crop identification model to predict and segment the selected image sequence to generate a high-resolution crop distribution map of the target area.

[0010] The adaptive phenological period difference-based cross-regional crop identification method and system provided by this invention has the following beneficial effects: This invention addresses the dual challenges of spatial visual features and temporal phenological rhythms in cross-regional crop identification when local labels are unavailable in the target area. First, it acquires low-resolution crop labels for the source region, multi-temporal high-resolution remote sensing images covering both the source and target regions, and time-series vegetation index data for both regions. By comparing and analyzing the time-series vegetation index curves of the two regions, the temporal offset of crop phenology between the source and target regions is quantified. Simultaneously, a label super-resolution network is used to process low-resolution labels in the source domain into high-resolution pseudo-labels; based on the aforementioned... Phenological alignment was performed on multi-temporal images from the source domain, and these images were paired with high-resolution pseudo-labels to train a spatiotemporally adaptive Transformer network. Finally, the trained model was applied to images that had undergone similar phenological alignment. High-resolution crop distribution maps can be generated from phenologically aligned images of the target area.

[0011] This invention proposes and solves the problem of spatial visual differences and temporal phenological differences that exist simultaneously in remote sensing crop identification due to different geographical locations, improves the model's generalization ability and robustness in new regions, and solves the problem of adaptation to both spatiotemporal domains. This invention, through an innovative phenological period correction step and a temporal adaptive network structure, enables the model to break free from dependence on absolute calendar dates and learn the inherent laws of crop growth cycles. Even if the phenological periods of the source domain and the target domain differ by several weeks or even months, it can achieve accurate identification and achieve adaptation to phenological differences. This invention makes it possible to generate high-precision, annually updated crop distribution maps for any target area using readily available, low-resolution agricultural census data or products from other regions, reducing reliance on expensive local labels and lowering the cost of high-precision agricultural monitoring. To address the technical challenges of poor model generalization and low recognition accuracy in cross-regional remote sensing crop identification due to the inability to handle differences in spatial visual features and temporal phenological rhythms, an innovative phenological period correction mechanism is employed to overcome temporal differences and improve the accuracy and robustness of crop identification in label-free, cross-regional scenarios. Attached Figure Description

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the cross-regional crop identification method for adaptive phenological period differences provided by the present invention; Figure 2 This is a flowchart of the overall process for the cross-regional crop identification method based on adaptive phenological differences provided by the present invention. Figure 3 This is a schematic diagram of cross-regional phenological period correction; Figure 4 This is a schematic diagram of the second neural network model. Detailed Implementation

[0013] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0014] Figure 1 A schematic diagram of the cross-regional crop identification method based on adaptive phenological differences, according to this embodiment, is shown. In this embodiment, the cross-regional crop identification method based on adaptive phenological differences includes the following steps: S1: Acquire low-resolution crop tags for the source region, multi-temporal high-resolution remote sensing image sequence data of the source and target regions, and vegetation index time series data; S2: Calculate the time offset between the source region and the target region based on the low-resolution crop tags and vegetation index time series data; In one exemplary embodiment, the step of calculating the time offset between the source region and the target region includes: Using the low-resolution crop tags as masks, pure phenological curves representing the target crop in the source region are extracted from the vegetation index time series data of the source region. Extract the mixed phenological curve representing the overall vegetation growth rhythm of the target area from the vegetation index time series data of the target area, and use it as the estimated phenological curve of the target crop. Based on the pure phenological curve and the mixed phenological curve, the time offset between the source region and the target region is obtained; In one exemplary embodiment, obtaining the time offset between the source region and the target region based on the pure phenological curve and the mixed phenological curve specifically includes: Based on the pure phenological curve and the mixed phenological curve, the growth peak period of the pure phenological curve and the mixed phenological curve are determined respectively; Calculate the time difference between the growth peak periods of the pure phenological curve and the mixed phenological curve to obtain the time offset between the source region and the target region; S3: Use the first neural network model to perform super-resolution processing on the low-resolution crop tags to obtain high-resolution pseudo-tags that match the multi-temporal high-resolution remote sensing image sequence data of the source region. In one exemplary embodiment, the first neural network model is a fully convolutional network, which includes an encoder and a decoder, the decoder including at least one transposed convolutional layer for upsampling the feature map; S4: Based on the time offset between the source region and the target region, perform phenological alignment processing on the multi-temporal high-resolution remote sensing image sequence data of the source region to obtain the phenologically aligned source region image sequence; use the high-resolution pseudo-label and the phenologically aligned source region image sequence to train the second neural network model to obtain the crop recognition model; In one exemplary embodiment, the second neural network model includes a spatiotemporal feature embedding module, a Transformer encoder, and a pixel-level segmentation decoder; The spatiotemporal feature embedding module is used to convert multi-temporal high-resolution remote sensing image sequence data into a feature vector sequence containing spatial information and temporal location information. The Transformer encoder is used to capture temporal dependencies using a self-attention mechanism based on the feature vector sequence, and to separate and normalize content features and style features using an instance-batch normalization module to adapt to spatial domain visual differences. The pixel-level segmentation decoder is used to upsample the features output by the Transformer encoder to the original spatial resolution in order to generate the final segmentation map; S5: Based on the time offset between the source region and the target region, select an image sequence from the multi-temporal high-resolution remote sensing image sequence data of the target region that matches the phenological period of the source region image sequence after phenological alignment, obtain the selected image sequence, and use the crop recognition model to predict and segment the selected image sequence to generate a high-resolution crop distribution map of the target region.

[0015] This embodiment provides a cross-regional crop identification system that adapts to phenological differences, the system comprising the following modules: The data acquisition module is configured to acquire low-resolution crop tags from the source region, multi-temporal high-resolution remote sensing image sequence data of the source and target regions, and vegetation index time series data. The phenological period correction module is configured to calculate the time offset between the source region and the target region based on the low-resolution crop tags and vegetation index time series data. The pseudo-label generation module is configured to: use a first neural network model to perform super-resolution processing on the low-resolution crop labels to obtain high-resolution pseudo-labels that match the multi-temporal high-resolution remote sensing image sequence data of the source region; The model training module is configured to: perform phenological alignment processing on the multi-temporal high-resolution remote sensing image sequence data of the source region according to the time offset between the source region and the target region, to obtain the phenologically aligned source region image sequence; and train the second neural network model using the high-resolution pseudo-label and the phenologically aligned source region image sequence to obtain the crop recognition model. The crop identification application module is configured to: select an image sequence from the multi-temporal high-resolution remote sensing image sequence data of the target area that matches the phenological period of the source area image sequence after phenological alignment, based on the time offset between the source area and the target area, obtain the selected image sequence, and use the crop identification model to predict and segment the selected image sequence to generate a high-resolution crop distribution map of the target area.

[0016] In some embodiments, the above-described method for cross-regional crop identification based on adaptive phenological differences can also be implemented in the following ways.

[0017] like Figure 2 The diagram shows the overall flowchart of the cross-regional crop identification method that adapts to phenological differences; in this embodiment, the cross-regional crop identification method that adapts to phenological differences includes: Step 1: Acquisition of multi-source time-series data: Acquire low-resolution crop classification labels for the source region, as well as multi-temporal high-resolution remote sensing image sequences covering the source and target regions; simultaneously acquire high-temporal-frequency vegetation index time-series data covering these two regions.

[0018] Step 2: Cross-regional phenological period correction: Based on vegetation index time series data, extract or estimate the key phenological periods of the target crop in both the source and target regions, and calculate the time offset between the two regions accordingly. .

[0019] Step 3: High-resolution pseudo-label generation: Input the low-resolution crop classification labels of the source region into a label super-resolution network, which serves as the first neural network model, to generate high-resolution pseudo-labels with clear boundaries that are consistent with the spatial resolution of the high-resolution remote sensing image of the source region.

[0020] Step 4: Spatiotemporal Adaptive Model Training: Based on the time offset calculated in Step 2 The multi-temporal high-resolution remote sensing image sequence of the source region is subjected to phenological alignment processing, and the sequence is used as a training pair with the high-resolution pseudo-label generated in step three. The result is a training of a spatiotemporal adaptive network that serves as the second neural network model, thereby obtaining a crop recognition model with cross-regional generalization capability.

[0021] Step 5: Crop Identification in the Target Area: Based on the time offset calculated in Step 2 Images matching the phenological period of the source region during training are selected from the multi-temporal high-resolution remote sensing image sequence of the target region, and then input into the trained crop recognition model for pixel-level classification, thereby generating a high-resolution crop distribution map of the target region.

[0022] Preferably, in step two, the time offset... The specific calculation methods include: 1. Smooth the vegetation index time series curves of the source and target regions to obtain smoothed phenological curves. and ,in, Dates within the year; 2. Determine the peak growth dates of the two curves by solving for the maximum parameter:

[0023]

[0024] 3. Calculate the difference between the dates of the two peak periods to obtain the time offset:

[0025] Preferably, in step three, the label super-resolution network, which serves as the first neural network model, is a fully convolutional network that adopts an encoder-decoder structure. Its core is that the decoder uses a transposed convolutional layer to upsample the feature map extracted by the encoder to restore the spatial size of the high-resolution image, thereby realizing the generation of high-resolution pseudo-labels from low-resolution labels. Preferably, in step four, the spatiotemporal adaptive network serving as the second neural network model is specifically a spatiotemporal phenological adaptive Transformer network, the network structure of which includes: 1. Spatiotemporal Feature Embedding Module. A convolutional neural network is used to independently extract spatial features from each frame of the multi-temporal image sequence, and these features are added to a learnable temporal position code to transform the image sequence into a feature vector sequence containing spatial content and temporal order information.

[0026] 2. Transformer encoder with built-in instance-batch normalization module. This is the core of the network. It captures long-range temporal dependencies in the crop growth cycle through a self-attention mechanism, enabling the learning of the intrinsic laws governing phenological rhythms. Simultaneously, it innovatively embeds an IBN module into the feedforward network. Batch normalization is used to learn cross-regional invariant, content-related universal features, while instance normalization is used to remove stylized features of specific regions, thereby synergistically achieving dual adaptation to temporal phenological differences and spatial visual differences.

[0027] 3. Pixel-level segmentation decoder. Through cascaded upsampling modules and skip connections, the high-dimensional spatiotemporal features output by the Transformer encoder are restored to the original image resolution, generating the final pixel-level crop classification map.

[0028] Preferably, step one also includes unifying the categories of crop tags from different sources, and performing coordinate registration and data augmentation on all images and tags.

[0029] This embodiment provides a cross-regional crop identification system that adapts to phenological differences. The system includes: a data acquisition module for performing step one of the aforementioned method; a phenological correction module for performing step two of the aforementioned method; a pseudo-label generation module for performing step three of the aforementioned method; a spatiotemporal adaptive model training module for performing step four of the aforementioned method, which internally deploys the second neural network model; and a crop identification application module for performing step five of the aforementioned method.

[0030] In some embodiments, the above-described adaptive phenological period difference cross-regional crop identification system can also be implemented in the following ways.

[0031] In this embodiment, the cross-regional crop identification system that adapts to phenological differences includes: The data acquisition module is used to acquire low-resolution crop tags of the source area, multi-temporal high-resolution remote sensing image sequences covering the source / target area, and vegetation index time series data covering the source / target area; The phenological period correction module is used to determine the key phenological period of the target crop in the source area and the estimated key phenological period of the target crop in the target area based on the vegetation index time series data, and to calculate the time offset between the two. The pseudo-label generation module is used to perform super-resolution processing on the low-resolution crop labels through a first neural network model to generate high-resolution pseudo-labels. The model training module is used to perform phenological alignment processing on the multi-temporal high-resolution remote sensing image sequence of the source region according to the time offset, and pair the aligned image sequence with the high-resolution pseudo-label to train a second neural network model with spatiotemporal adaptive capability to obtain a crop recognition model. The crop identification application module is used to perform phenological alignment image selection on the multi-temporal high-resolution remote sensing image sequence of the target area based on the time offset, and input the selected images into the crop identification model to generate a high-resolution crop distribution map of the target area.

[0032] In some embodiments, the above-described method for cross-regional crop identification based on adaptive phenological differences can also be implemented in the following ways.

[0033] This embodiment aims to identify cotton across regions. It selects Xinjiang Uygur Autonomous Region, which has significant differences in cotton planting patterns and phenological rhythms, as the source domain, and Hunan Province, which lacks local labels, as the target domain. The goal is to use data from Xinjiang to generate a high-precision cotton distribution map of Hunan Province.

[0034] The specific method flow of this embodiment is as follows: Step 1: Data Acquisition and Preprocessing 1. Obtain a multi-temporal, high-resolution remote sensing image sequence covering the main cotton-producing areas of Xinjiang during the 2022 growing season, specifically GF-2 images with a spatial resolution of 2 meters. Simultaneously, obtain a cotton planting distribution map of the corresponding region, released by the agricultural department, with a spatial resolution of 30 meters, as a low-resolution crop label.

[0035] 2. Obtain a sequence of high-resolution GF-2 remote sensing images of multiple time phases covering cotton-growing areas in Hunan Province during the same period.

[0036] 3. Download Sentinel-2 NDVI products covering Xinjiang Uygur Autonomous Region and Hunan Province during the 2022 growing season from a public data platform. The spatial resolution is 10 meters and the temporal resolution is 5 days. After interpolation, a smooth vegetation index time series is obtained.

[0037] 4. Perform coordinate system I, geometric registration, and radiometric correction on all image and label data. Classify the cotton labels in the source region and crop all data into uniformly sized image blocks using a sliding window for subsequent processing.

[0038] Step 2: Cross-regional phenological period correction like Figure 3 The diagram shown is a cross-regional phenological period correction diagram; 1. Using the 30-meter cotton tag obtained in step one as a mask, extract the average NDVI of pure cotton pixels from the Sentinel-2 NDVI time series data of Xinjiang, thereby obtaining a pure phenological curve representing the standard growth rhythm of Xinjiang cotton. .

[0039] 2. Based on agricultural statistics, Changde City, Hunan Province, was selected as the macro-analysis unit. Sentinel-2 NDVI time series data for all pixels in the city were extracted, and their average values ​​were calculated to obtain a mixed phenological curve representing the overall vegetation growth rhythm of the region. .

[0040] 3. To accurately calculate the time offset Savitzky-Golay filtering was used to smooth and denoise the two curves, resulting in... and The dates within the year corresponding to the peak growth periods of the two curves are determined by solving for the independent variable (argmax) of the maximum parameter.

[0041] 4. Calculate the difference between the two peak dates to obtain the time offset. For example, if the peak growth period for cotton in Xinjiang is calculated to be on the 210th day of the year, while the peak growth period for vegetation in Hunan is on the 185th day of the year, then the time offset...

[0042] Step 3: High-resolution pseudo-tag generation Construct a fully convolutional network (FCN) for label super-resolution as the first neural network model. This network employs an encoder-decoder structure. Input the 30-meter low-resolution cotton labels from the source domain in step one into this model. The decoder core consists of a convolutional kernel with a stride of 3 and a kernel size of [missing information]. The transposed convolutional layer increases the resolution of the feature maps extracted by the encoder by 3 times, restoring the 30-meter resolution labels to a spatial resolution that matches the 2-meter GF-2 image, thereby generating high-resolution cotton pseudo-labels with clear boundaries and accurate patches.

[0043] It is particularly important to note that this embodiment compares the pure crop phenological curve of the source region with the mixed vegetation phenological curve of the target region. This is key to achieving adaptive correction without target domain labels. The technical rationale lies in the fact that the core premise of this invention is the lack of prior knowledge such as labels in the target region, making it impossible to accurately extract the pure curve of the target crop. However, within a macro-scale agricultural analysis unit, the growth rhythm of the target crop has a dominant influence on the average growth rhythm of all vegetation within that unit. Therefore, this mixed vegetation phenological curve is the best proxy signal reflecting the phenological rhythm of the target crop. This invention cleverly avoids dependence on expensive target domain labels by comparing the phase difference of the peak times of the two curves, achieving low-cost, high-efficiency adaptive correction of phenological differences.

[0044] Step 4: Spatiotemporal Adaptive Model Training 1. Based on the time offset calculated in step two. Phenological alignment was performed on a multi-temporal 2-meter high-resolution image sequence from the source region (Xinjiang). Specifically, the timestamps of the original image sequence were shifted forward to construct a training image sequence synchronized with the phenological rhythm of the target region.

[0045] 2. The training image sequence aligned with the phenological pattern described above, along with the high-resolution pseudo-labels generated in step 3, are used as training pairs and input into a spatiotemporal phenological adaptive Transformer network, which serves as a second neural network, for end-to-end training. This network transforms the image sequence into feature vectors through a spatiotemporal feature embedding module; its built-in IBN Transformer encoder learns the intrinsic temporal pattern of the cotton growth cycle through a self-attention mechanism and adapts to spatial visual differences through the IBN module; finally, the network uses a pixel-level segmentation decoder to restore the resolution and output the classification result.

[0046] like Figure 4 The diagram shown is a schematic of the second neural network model.

[0047] Step 5: Crop Identification in the Target Area 1. Based on Image sequences matching the phenological periods used during source domain training were selected from the multi-temporal high-resolution image database of Hunan Province.

[0048] 2. The selected image sequence of Hunan Province is input into the trained Transformer network model for inference. The model classifies each pixel and finally generates a high-resolution cotton distribution map of Hunan Province. Implementation Results: To verify the effectiveness of this embodiment, we conducted a comparative experiment. The baseline method directly applied a standard time-series model trained on Xinjiang data to Hunan Province without performing the phenological period correction in step two and the phenological alignment in step four. Experimental results showed that the baseline method achieved a cotton identification F1 score of only 65.3%, exhibiting severe misclassification and omissions due to both phenological and spatial differences. However, using the complete method described in this embodiment, even in Hunan Province where no real labels were available, the generated cotton distribution map, verified through small-scale field sampling, achieved an F1 score of 88.7%, an improvement of 23.4 percentage points compared to the baseline method. This demonstrates that the present invention can effectively solve the model failure problem caused by phenological and spatial visual differences, showcasing cross-regional transferability and practical application value.

[0049] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for cross-regional crop identification based on adaptive phenological differences, characterized in that, Includes the following steps: S1: Acquire low-resolution crop tags for the source region, multi-temporal high-resolution remote sensing image sequence data of the source and target regions, and vegetation index time series data; S2: Calculate the time offset between the source region and the target region based on the low-resolution crop tags and vegetation index time series data; S3: Use the first neural network model to perform super-resolution processing on the low-resolution crop tags to obtain high-resolution pseudo-tags that match the multi-temporal high-resolution remote sensing image sequence data of the source region. S4: Based on the time offset between the source region and the target region, perform phenological alignment processing on the multi-temporal high-resolution remote sensing image sequence data of the source region to obtain the phenologically aligned source region image sequence; use the high-resolution pseudo-label and the phenologically aligned source region image sequence to train the second neural network model to obtain the crop recognition model; the second neural network model includes a spatiotemporal feature embedding module, a Transformer encoder, and a pixel-level segmentation decoder; the spatiotemporal feature embedding module is used to convert the multi-temporal high-resolution remote sensing image sequence data into a feature vector sequence containing spatial information and temporal location information; the Transformer encoder is used to capture temporal dependencies using a self-attention mechanism based on the feature vector sequence, and to separate and normalize content features and style features using an instance-batch normalization module to adapt to spatial domain visual differences; the pixel-level segmentation decoder is used to upsample the features output by the Transformer encoder to the original spatial resolution to generate the final segmentation map. S5: Based on the time offset between the source region and the target region, select an image sequence from the multi-temporal high-resolution remote sensing image sequence data of the target region that matches the phenological period of the source region image sequence after phenological alignment, obtain the selected image sequence, and use the crop recognition model to predict and segment the selected image sequence to generate a high-resolution crop distribution map of the target region.

2. The cross-regional crop identification method based on adaptive phenological differences according to claim 1, characterized in that, The step of calculating the time offset between the source region and the target region includes: Using the low-resolution crop tags as masks, pure phenological curves representing the target crop in the source region are extracted from the vegetation index time series data of the source region. Extract the mixed phenological curve representing the overall vegetation growth rhythm of the target area from the vegetation index time series data of the target area, and use it as the estimated phenological curve of the target crop. Based on the pure phenological curve and the mixed phenological curve, the time offset between the source region and the target region is obtained.

3. The cross-regional crop identification method based on adaptive phenological differences according to claim 2, characterized in that, The step of obtaining the time offset between the source region and the target region based on the pure phenological curve and the mixed phenological curve specifically includes: Based on the pure phenological curve and the mixed phenological curve, the growth peak period of the pure phenological curve and the mixed phenological curve are determined respectively; The time difference between the growth peak periods of the pure phenological curve and the mixed phenological curve is calculated to obtain the time offset between the source region and the target region.

4. The cross-regional crop identification method based on adaptive phenological differences according to claim 1, characterized in that, The first neural network model is a fully convolutional network, which includes an encoder and a decoder. The decoder includes at least one transposed convolutional layer, which is used for upsampling of the feature map.

5. A cross-regional crop identification system that adapts to phenological differences, characterized in that, The system includes the following modules: The data acquisition module is configured to acquire low-resolution crop tags from the source region, multi-temporal high-resolution remote sensing image sequence data of the source and target regions, and vegetation index time series data. The phenological period correction module is configured to calculate the time offset between the source region and the target region based on the low-resolution crop tags and vegetation index time series data. The pseudo-label generation module is configured to: use a first neural network model to perform super-resolution processing on the low-resolution crop labels to obtain high-resolution pseudo-labels that match the multi-temporal high-resolution remote sensing image sequence data of the source region; The model training module is configured to: perform phenological alignment processing on the multi-temporal high-resolution remote sensing image sequence data of the source region according to the time offset between the source region and the target region, to obtain the phenologically aligned source region image sequence; train the second neural network model using the high-resolution pseudo-label and the phenologically aligned source region image sequence to obtain the crop recognition model; the second neural network model includes a spatiotemporal feature embedding module, a Transformer encoder, and a pixel-level segmentation decoder. The spatiotemporal feature embedding module is used to convert multi-temporal high-resolution remote sensing image sequence data into a feature vector sequence containing spatial information and temporal location information. The Transformer encoder is used to capture temporal dependencies using a self-attention mechanism based on the feature vector sequence, and to separate and normalize content features and style features using an instance-batch normalization module to adapt to spatial domain visual differences. The pixel-level segmentation decoder is used to upsample the features output by the Transformer encoder to the original spatial resolution in order to generate the final segmentation map; The crop identification application module is configured to: select an image sequence from the multi-temporal high-resolution remote sensing image sequence data of the target area that matches the phenological period of the source area image sequence after phenological alignment, based on the time offset between the source area and the target area, obtain the selected image sequence, and use the crop identification model to predict and segment the selected image sequence to generate a high-resolution crop distribution map of the target area.

6. The cross-regional crop identification system for adaptive phenological differences according to claim 5, characterized in that, The step of calculating the time offset between the source region and the target region includes: Using the low-resolution crop tags as masks, pure phenological curves representing the target crop in the source region are extracted from the vegetation index time series data of the source region. Extract the mixed phenological curve representing the overall vegetation growth rhythm of the target area from the vegetation index time series data of the target area, and use it as the estimated phenological curve of the target crop. Based on the pure phenological curve and the mixed phenological curve, the time offset between the source region and the target region is obtained.

7. The cross-regional crop identification system for adaptive phenological period differences according to claim 6, characterized in that, The step of obtaining the time offset between the source region and the target region based on the pure phenological curve and the mixed phenological curve specifically includes: Based on the pure phenological curve and the mixed phenological curve, the growth peak period of the pure phenological curve and the mixed phenological curve are determined respectively; The time difference between the growth peak periods of the pure phenological curve and the mixed phenological curve is calculated to obtain the time offset between the source region and the target region.

8. The cross-regional crop identification system for adaptive phenological period differences according to claim 5, characterized in that, The first neural network model is a fully convolutional network, which includes an encoder and a decoder. The decoder includes at least one transposed convolutional layer, which is used for upsampling of the feature map.

Images