Transform mechanism-based Sentinel2 satellite remote sensing image wheat planting information extraction method

Abstract

The invention belongs to the technical field of agricultural monitoring, and particularly relates to a Sentinel2 satellite remote sensing image wheat planting information extraction method based on a Transform mechanism, which comprises the following specific steps: data acquisition and preprocessing, sample data making, segmentation model construction based on Transform, model evaluation and optimization, and result output and application. In the data acquisition stage, acquisition of images covering wheat growth key periods is clearly emphasized, so that the classification precision is greatly improved; the data preprocessing steps are complete, the operation method and principle are elaborated in each link from radiometric calibration and atmospheric correction to image cutting, splicing and registration, high quality of data is ensured, and a solid foundation is laid for subsequent model training.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural monitoring, and in particular to a method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on a Transformer mechanism. Background Art

[0002] In the field of agricultural monitoring, accurate wheat planting information is crucial for grain yield estimation and agricultural resource management. Sentinel-2 satellite remote sensing imagery, due to its inclusion of a red edge band, offers unique advantages for crop species identification. However, traditional remote sensing image classification methods have limitations and inaccuracies in extracting wheat planted area.

[0003] Currently, we do not perform traditional remote sensing data preprocessing such as atmospheric and geometric correction. Instead, we directly download L2A data for remote sensing interpretation and then use the results for practical applications. If we must perform preprocessing, we should download Sentinel 2L1c data, as L2A data is an atmospherically corrected product. We use ArcGIS tools to annotate the data based on the field survey data of the previous year, with reference to Google's high-resolution data, historical data, and NDVI thresholds. We also perform multi-person cross-validation annotations.

[0004] Therefore, a method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism was invented. Summary of the Invention

[0005] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0006] The wheat planting information extraction method from Sentinel2 satellite remote sensing images based on the Transformer mechanism includes the following specific steps:

[0007] S1, data acquisition and preprocessing: first acquire Sentinel-2 satellite image data, then perform data preprocessing;

[0008] S2, sample data production: first label the wheat samples, then construct the sample dataset;

[0009] S3, Transformer-based segmentation model construction: first perform model selection and architecture design, then perform model training;

[0010] S4, model evaluation and optimization: first calculate the model evaluation index, then optimize and improve the model;

[0011] S5, result output and application: first generate wheat planting information extraction results, then apply and verify the results.

[0012] As a preferred solution of the method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism of the present invention, the steps of acquiring Sentinel-2 satellite image data in S1 are as follows:

[0013] S11, determine the research area: identify the geographical area where wheat planting information needs to be extracted and obtain the corresponding Sentinel-2 satellite image data;

[0014] S12, Data download: Download Sentinel-2 L1C or L2A image data covering the study area through the ESA distribution platform, ensuring that the time range of the image covers the critical period of wheat growth to improve classification accuracy.

[0015] As a preferred solution of the method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism of the present invention, the data preprocessing steps in S1 are as follows:

[0016] S13, radiometric calibration: converting digital quantized values recorded by satellite sensors into absolute radiometric brightness values to eliminate the effects of sensor response differences and atmospheric scattering and absorption factors on image radiometric information;

[0017] S14, Atmospheric Correction: Use atmospheric correction algorithms to remove the scattering and absorption effects of the atmosphere on the image, convert the radiance value into surface reflectivity, and improve the authenticity and interpretability of the image;

[0018] S15, Image cropping and stitching: Crop the downloaded Sentinel-2 image according to the boundary vector file of the study area to remove irrelevant areas. If the study area is large, multiple images need to be stitched together to generate an image dataset that fully covers the study area.

[0019] S16, Image Registration: Based on high-precision geographic reference data, the pre-processed images are registered to ensure spatial consistency between images of different phases and different bands, with the error controlled at the sub-pixel level.

[0020] As a preferred solution of the method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism of the present invention, the steps of labeling wheat samples in S2 are as follows:

[0021] S21, Visual Interpretation: Using high-resolution Sentinel-2 imagery, combined with field survey data and historical agricultural data, and through professional visual interpretation, the polygonal boundaries of wheat-growing areas are accurately marked on the imagery.

[0022] S22, sample diversification: To improve the generalization ability of the model, ensure that the labeled wheat samples cover wheat-growing areas with different geographical locations, terrain conditions, growth conditions and soil backgrounds within the study area; at the same time, label a certain number of non-wheat land samples as negative samples for the classification model.

[0023] As a preferred solution of the method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism described in the present invention, the steps of constructing the sample dataset in S2 are as follows:

[0024] S23, Sample Division: Divide the labeled sample data into training set, validation set, and test set according to a certain ratio. During the division process, the distribution of various samples in each subset should be relatively balanced to avoid overfitting or underfitting of the model.

[0025] S24, data enhancement: Perform data enhancement operations on training set samples to increase sample diversity, expand the amount of training data, and improve the robustness and generalization ability of the model.

[0026] As a preferred solution of the method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism described in the present invention, the steps of model selection and architecture design in S3 are as follows:

[0027] S31, select Transformer architecture: adopt the Transformer-based semantic segmentation model;

[0028] S32, model architecture design: Based on the selected Transformer model, the model's input layer, encoder, decoder, and output layers are appropriately adjusted and optimized according to the characteristics of Sentinel-2 imagery and the requirements of the wheat planting information extraction task.

[0029] As a preferred solution of the method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism described in the present invention, the steps of model training in S3 are as follows:

[0030] S33, set training parameters: determine the hyperparameters for model training, including learning rate, number of iterations, batch size, and optimizer. Through experimental comparison and parameter tuning, select the optimal hyperparameter combination to improve model training efficiency and segmentation accuracy.

[0031] S34, loss function definition: Use a loss function suitable for the semantic segmentation task, or combine multiple loss functions to construct a composite loss function to better measure the difference between the model prediction results and the true labels, and guide model training;

[0032] S35, model training process: The divided training set samples are input into the constructed Transformer model for training. During the training process, the performance indicators of the model on the validation set are monitored in real time. The training parameters are adjusted according to the performance of the validation set to prevent the model from overfitting. The model weight file with the best performance during the training process is saved for subsequent prediction and evaluation.

[0033] As a preferred solution of the method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism of the present invention, the steps of calculating the model evaluation index in S4 are as follows:

[0034] S41, using the test set evaluation: Apply the trained model to the test set samples, obtain the model's prediction results, and calculate the evaluation indicators to comprehensively evaluate the accuracy and reliability of the model in extracting wheat planting information;

[0035] S42, Comparative Accuracy Analysis: The evaluation results are compared with those of traditional remote sensing image classification methods to intuitively demonstrate the advantages of the Transformer-based method in improving the accuracy of wheat planting area extraction.

[0036] As a preferred solution of the method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism of the present invention, the steps of model optimization and improvement in S4 are as follows:

[0037] S43, model fine-tuning: Based on the evaluation results, analyze the prediction errors of the model in different regions and landform types, and fine-tune the model accordingly;

[0038] S44, Model Fusion: Try to fuse the Transformer-based model with the auxiliary model, comprehensively utilize the advantages of different models, and further improve the accuracy of wheat planting information extraction;

[0039] S45, post-processing optimization: perform post-processing operations on the model's prediction results to remove isolated noise points and small misclassified areas, smooth the classification boundaries, and improve the integrity and accuracy of the classification results.

[0040] As a preferred solution of the method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism of the present invention, the step of generating the wheat planting information extraction result in S5 is as follows:

[0041] S51, classification map production: The optimized and evaluated model was applied to the Sentinel-2 imagery data of the entire study area to obtain a classification map of the wheat planting area;

[0042] S52, Area Statistics and Report Generation: Based on the classification map, count the wheat planting area in the study area and generate a detailed report;

[0043] The steps for applying and verifying the results in S5 are as follows:

[0044] S53, Field Verification and Feedback: Combine field surveys and ground-based measurement data to conduct field verification of the extracted wheat planting information. Compare and analyze the field observation results with the remote sensing interpretation results to evaluate the accuracy and reliability of the interpretation results. Any problems and errors discovered during the field verification process will be promptly fed back to the data processing and model optimization stages to further improve the accuracy and practicality of the wheat planting information extraction.

[0045] S54, Agricultural Production Monitoring and Management: Apply the extracted wheat planting information to agricultural production monitoring and management. By regularly conducting remote sensing monitoring of wheat planting areas, we can timely grasp the growth status and changing trends of wheat, and provide accurate guidance services for agricultural production.

[0046] Compared with existing technologies:

[0047] 1. Comprehensive and detailed data processing: During the data acquisition phase, explicit emphasis was placed on obtaining images covering the critical period of wheat growth, which greatly improved classification accuracy. The data preprocessing steps were complete, from radiometric calibration and atmospheric correction to image cropping, stitching, and registration. Each step detailed the operating methods and principles to ensure high data quality and lay a solid foundation for subsequent model training.

[0048] 2. Scientific and reasonable sample preparation: Wheat sample annotation is achieved through visual interpretation combined with multi-source data to ensure accuracy. At the same time, emphasis is placed on sample diversity, covering different terrains, growth conditions, and other land types. Stratified random sampling is used when dividing training, validation, and test sets to ensure balanced sample distribution. In addition, a variety of data augmentation methods are used to effectively improve model generalization and robustness.

[0049] 3. Targeted model construction: We selected a suitable Transformer architecture model and rationally adjusted and optimized each model layer based on the characteristics of Sentinel-2 imagery and the wheat planting information extraction task. During model training, hyperparameter setting, loss function definition, and training process monitoring were clearly planned, which helped improve model training results.

[0050] 4. Comprehensive Evaluation and Optimization: A variety of model evaluation metrics comprehensively measure model accuracy and reliability, while also highlighting advantages through comparative analysis with traditional methods. In terms of model optimization, from model fine-tuning and fusion to post-processing optimization, model performance is enhanced across multiple dimensions, resulting in more accurate extraction results.

[0051] 5. Widely applicable and closed-loop results: The output section details the process of creating classification maps and generating area statistics reports. Applications encompass multifaceted monitoring and management of agricultural production. Field validation and feedback form a closed-loop optimization process, continuously improving the practicality and accuracy of the solution. DETAILED DESCRIPTION

[0052] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0053] The present invention provides a method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism, which includes the following specific steps:

[0054] S1, data acquisition and preprocessing: first acquire Sentinel-2 satellite image data, then perform data preprocessing;

[0055] The specific steps of S1 are as follows:

[0056] S11. Determine the study area: Identify the geographic area where wheat planting information needs to be extracted and obtain the corresponding Sentinel-2 satellite imagery data. The Sentinel-2 satellite has high spatial resolution (10-60 meters) and multispectral bands (including the red edge band), which can effectively identify crop species.

[0057] S12, Data download: Download Sentinel-2 L1C or L2A imagery covering the study area through the European Space Agency (ESA) Copernicus Open Access Hub or other data distribution platforms; ensure that the imagery covers the critical period of wheat growth to improve classification accuracy;

[0058] S13, radiometric calibration: converting the digital quantization values (DN) recorded by satellite sensors into absolute radiometric brightness values to eliminate the effects of sensor response differences and atmospheric scattering and absorption on image radiometric information;

[0059] S14, Atmospheric Correction: Use atmospheric correction algorithms such as the 6S model and ACOLITE to remove atmospheric scattering and absorption effects on images, convert radiance values into surface reflectance, and improve image authenticity and interpretability.

[0060] The atmospheric correction steps of the 6S model are as follows:

[0061] Enter basic parameters: Open the 6S model software and, in the corresponding parameter setting interface, accurately enter the imaging time of the Sentinel-2 image, accurate to the second. This determines the position of the sun and the atmospheric state. Next, enter the geographic location of the image, including longitude and latitude, to ensure that the model can accurately simulate the local atmospheric characteristics. Also, specify that the sensor type corresponding to the image is Sentinel-2MSI to adapt the model to the corresponding radiation transfer model.

[0062] Obtain and input ground elevation data: Obtain digital elevation model (DEM) data for the study area from authoritative data sources, such as the Shuttle Radar Topography Mission (SRTM). Import the DEM data into the 6S model, which adjusts parameters such as atmospheric density and pressure based on elevation information. This is because the composition and density of the atmosphere at different altitudes vary, affecting the absorption and scattering of radiation by the atmosphere.

[0063] Set atmospheric parameters: Set the aerosol type in the model. Common types include continental and oceanic types, and the type should be selected based on the actual environmental characteristics of the study area. Determine the aerosol optical depth by referring to local atmospheric monitoring data or using the model's built-in inversion algorithm for preliminary estimation. Also, set cloud cover and cloud height and other cloud-related parameters. If the image is covered by clouds, accurately setting these parameters will help the model more accurately correct for atmospheric effects.

[0064] Run the model calculation: After completing the above parameter settings, click the Run button. The 6S model is based on the radiation transfer equation and simulates the entire process of solar radiation entering from outside the atmosphere, being scattered and absorbed by the atmosphere, interacting with the surface, and then passing through the atmosphere again to be received by satellite sensors. The model uses complex mathematical operations to eliminate atmospheric interference with radiation and outputs corrected surface reflectance data.

[0065] The ACOLITE algorithm atmospheric correction steps are as follows:

[0066] Prepare the Python environment and related libraries: Ensure the ACOLITE library and related dependent libraries, such as numpy and scipy, are installed in the Python environment. If not, install them using the pipinstall command. Also, prepare the Sentinel-2 image data to be calibrated. The data format must meet the requirements of the ACOLITE library and is generally a common remote sensing image format, such as .tif.

[0067] Loading image data: Write a Python script and use the data reading functions in the ACOLITE library to load Sentinel-2 image data. For example, use the acolite.io.read function to specify the image file path and read the image data into an array format that can be processed by Python to facilitate subsequent operations.

[0068] Set atmospheric correction parameters: Set the key parameters required for atmospheric correction in the script; set the aerosol type, such as marine aerosol type in coastal areas; determine the solar zenith angle, which can be obtained from the image metadata through simple calculation and conversion; also set the observation zenith angle. If the image is observed vertically, the observation zenith angle is 0; in addition, set parameters such as the water body flag (if the study area contains water bodies) to optimize the atmospheric correction results;

[0069] Perform atmospheric correction: Call the atmospheric correction function in the ACOLITE library, such as acolite.atmcorr.run, and pass in the loaded image data and set parameters. The function performs atmospheric correction on the image based on the ACOLITE algorithm. The algorithm analyzes the radiance values of different image bands and, combined with the set parameters, eliminates the effects of atmospheric scattering and absorption, converting the image's radiance values into surface reflectance.

[0070] Output correction results: After atmospheric correction is completed, use the ACOLITE library's output function to save the corrected surface reflectance data in a suitable format, such as .tif format. When saving, you can specify the output path and file name to facilitate further analysis and processing in remote sensing image processing software or geographic information system (GIS) software.

[0071] S15, Image cropping and stitching: Crop the downloaded Sentinel-2 image according to the boundary vector file of the study area to remove irrelevant areas. If the study area is large, multiple images need to be stitched together to generate an image dataset that fully covers the study area.

[0072] S16, Image Registration: Using high-precision geographic reference data (such as digital elevation models (DEMs) and high-precision maps) as a benchmark, the pre-processed images are registered to ensure spatial consistency between images of different phases and bands, with errors controlled at the sub-pixel level.

[0073] S2, sample data production: first label the wheat samples, then construct the sample dataset;

[0074] The specific steps of S2 are as follows:

[0075] S21, Visual Interpretation: Using high-resolution Sentinel-2 imagery, combined with field survey data and historical agricultural data, professionals visually interpret the imagery to accurately delineate the polygonal boundaries of wheat-growing areas. Care must be taken to distinguish wheat from other landforms (e.g., other crops, forests, water bodies, and residential areas).

[0076] S22, Sample Diversification: To improve the generalization ability of the model, ensure that the labeled wheat samples cover wheat-growing areas with different geographical locations, terrain conditions, growth conditions, and soil backgrounds within the study area; at the same time, label a certain number of non-wheat ground samples as negative samples for the classification model;

[0077] S23, Sample Partitioning: Divide the labeled sample data into training, validation, and test sets according to a certain ratio (e.g., 70% training set, 15% validation set, and 15% test set). During the partitioning process, the distribution of various samples in each subset should be relatively balanced to avoid overfitting or underfitting of the model.

[0078] S24, Data Augmentation: Perform data augmentation operations on training set samples, such as random rotation, flipping, cropping, scaling, etc., to increase sample diversity, expand the amount of training data, and improve the robustness and generalization ability of the model;

[0079] S3, Transformer-based segmentation model construction: first perform model selection and architecture design, then perform model training;

[0080] The specific steps for S3 are as follows:

[0081] S31, choose the Transformer architecture: Use Transformer-based semantic segmentation models, such as SegFormer and SwinTransformer. These models effectively capture long-range dependencies in images through the self-attention mechanism and can better process complex spatial information in remote sensing images.

[0082] S32, Model Architecture Design: Based on the selected Transformer model, the model's input layer, encoder, decoder, and output layers are appropriately adjusted and optimized according to the characteristics of Sentinel-2 imagery and the requirements of the wheat planting information extraction task. For example, the input layer is adjusted to adapt to the input format of multispectral imagery data, and a suitable decoder structure is designed to achieve accurate pixel-level classification output.

[0083] S33, set training parameters: determine the hyperparameters for model training, including learning rate, number of iterations, batch size, optimizer (such as AdamW), etc.; through experimental comparison and parameter tuning, select the optimal hyperparameter combination to improve model training efficiency and segmentation accuracy;

[0084] S34, loss function definition: Use a loss function suitable for semantic segmentation tasks, such as the cross-entropy loss function (CrossEntropyLoss) and the Dice loss function, or combine multiple loss functions to construct a composite loss function to better measure the difference between the model prediction results and the true labels, and guide model training;

[0085] Cross entropy and dice loss function or other functions are combined for optimization, and clear parameters or adaptive weights are set. total =λL Dice +(1-λ)L CE , and give the parameter range (such as λ = 0.5);

[0086] S35, model training process: Input the divided training set samples into the constructed Transformer model for training; During the training process, monitor the model performance indicators (such as accuracy, recall rate, F1 value, intersection over union ratio, etc.) on the validation set in real time, adjust the training parameters according to the validation set performance to prevent model overfitting; Save the model weight file with the best performance during the training process for subsequent prediction and evaluation;

[0087] S4, model evaluation and optimization: first calculate the model evaluation index, then optimize and improve the model;

[0088] The specific steps of S4 are as follows:

[0089] S41, using the test set for evaluation: Apply the trained model to the test set samples to obtain the model's prediction results; calculate a series of evaluation indicators, such as overall accuracy (OverallAccuracy), class accuracy (ClassAccuracy), recall (Recall), F1 value (F1-score), intersection over union (IoU), etc., to comprehensively evaluate the accuracy and reliability of the model in extracting wheat planting information;

[0090] S42, Comparative Accuracy Analysis: Compare and analyze the evaluation results of this method with those of traditional remote sensing image classification methods (such as support vector machines (SVMs) and random forests (RFs), visually demonstrating the advantages of the Transformer-based method in improving the accuracy of wheat area extraction. For example, by comparing the actual wheat area with the areas interpreted by different methods, calculate the error rate, and evaluate the effectiveness of this method in reducing area extraction errors.

[0091] S43, Model Fine-tuning: Based on the evaluation results, analyze the model's prediction errors in different regions and terrain types, and fine-tune the model accordingly. For example, for specific terrain or areas with mixed terrain that the model is prone to misjudging, add corresponding sample data and retrain the model to optimize its recognition ability in these areas.

[0092] S44, Model Fusion: Try to fuse the Transformer-based model with other auxiliary models (such as models based on convolutional neural networks (CNNs)) to comprehensively utilize the advantages of different models and further improve the accuracy of wheat planting information extraction. Model fusion can use weighted averaging, voting and other methods to integrate the prediction results of multiple models to obtain the final classification result.

[0093] S45, post-processing optimization: Perform post-processing operations on the model's prediction results, such as morphological filtering (dilation, erosion, etc.), connected domain analysis, etc., to remove isolated noise points and small misclassified areas, smooth classification boundaries, and improve the integrity and accuracy of the classification results;

[0094] S5, result output and application: first generate wheat planting information extraction results, then apply and verify the results;

[0095] The specific steps of S5 are as follows:

[0096] S51, Classification Map Creation: Apply the optimized and evaluated model to Sentinel-2 imagery data for the entire study area to generate a classification map of wheat-growing areas. The classification map uses different colors or labels to identify wheat-growing areas and other landform types, visually displaying the spatial distribution of wheat within the study area.

[0097] S52, Area Statistics and Report Generation: Based on the classification map, the wheat planting area within the study area is counted and a detailed report is generated. The report includes statistical results of wheat planting area, area distribution in different regions, and comparative analysis with historical data, providing a scientific basis for agricultural production management and decision-making.

[0098] S53, Field Verification and Feedback: Combine field surveys and ground-based measurement data to conduct field verification of the extracted wheat planting information. Compare and analyze field observations with remote sensing interpretation results to assess the accuracy and reliability of the interpretations. Provide timely feedback to data processing and model optimization based on any problems and errors discovered during the field verification process to further improve the accuracy and practicality of wheat planting information extraction.

[0099] S54, Agricultural Production Monitoring and Management: Apply the extracted wheat planting information to agricultural production monitoring and management, such as crop growth monitoring, yield estimation, irrigation and fertilization decision-making, etc.; through regular remote sensing monitoring of wheat planting areas, timely grasp the growth status and changing trends of wheat, and provide accurate guidance services for agricultural production.

[0100] Although the present invention has been described above with reference to embodiments, various modifications may be made thereto and equivalent components may be substituted without departing from the scope of the present invention. In particular, as long as there are no structural conflicts, the various features of the embodiments disclosed herein may be combined with each other in any manner, and the omission of an exhaustive description of such combinations in this specification is solely for the sake of space and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism, characterized by: The specific steps are as follows: S1, data acquisition and preprocessing: first acquire Sentinel-2 satellite image data, then perform data preprocessing; S2, sample data production: first label the wheat samples, then construct the sample dataset; S3, Transformer-based segmentation model construction: first perform model selection and architecture design, then perform model training; S4, model evaluation and optimization: first calculate the model evaluation index, then optimize and improve the model; S5, result output and application: first generate wheat planting information extraction results, then apply and verify the results.

2. The method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism according to claim 1 is characterized in that: The steps for acquiring Sentinel-2 satellite image data in S1 are as follows: S11, determine the research area: identify the geographical area where wheat planting information needs to be extracted and obtain the corresponding Sentinel-2 satellite image data; S12, Data download: Download Sentinel-2 L1C or L2A image data covering the study area through the ESA distribution platform, ensuring that the time range of the image covers the critical period of wheat growth to improve classification accuracy.

3. The method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism according to claim 2 is characterized in that: The steps of data preprocessing in S1 are as follows: S13, radiometric calibration: converting digital quantized values recorded by satellite sensors into absolute radiometric brightness values to eliminate the effects of sensor response differences and atmospheric scattering and absorption factors on image radiometric information; S14, Atmospheric Correction: Use atmospheric correction algorithms to remove the scattering and absorption effects of the atmosphere on the image, convert the radiance value into surface reflectivity, and improve the authenticity and interpretability of the image; S15, image cropping and stitching: Crop the downloaded Sentinel-2 image according to the boundary vector file of the study area to remove irrelevant areas; If the study area is large, multiple images need to be stitched together to generate an image dataset that completely covers the study area; S16, Image Registration: Based on high-precision geographic reference data, the pre-processed images are registered to ensure spatial consistency between images of different phases and different bands, with the error controlled at the sub-pixel level.

4. The method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism according to claim 1, characterized in that: The steps for labeling wheat samples in S2 are as follows: S21, Visual Interpretation: Using high-resolution Sentinel-2 imagery, combined with field survey data and historical agricultural data, and through professional visual interpretation, the polygonal boundaries of wheat-growing areas are accurately marked on the imagery. S22, sample diversification: To improve the generalization ability of the model, ensure that the labeled wheat samples cover wheat-growing areas with different geographical locations, terrain conditions, growth conditions and soil backgrounds within the study area; at the same time, label a certain number of non-wheat land samples as negative samples for the classification model.

5. The method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism according to claim 4 is characterized in that: The steps for constructing the sample dataset in S2 are as follows: S23, Sample Division: Divide the labeled sample data into training set, validation set, and test set according to a certain ratio. During the division process, the distribution of various samples in each subset should be relatively balanced to avoid overfitting or underfitting of the model. S24, data enhancement: Perform data enhancement operations on training set samples to increase sample diversity, expand the amount of training data, and improve the robustness and generalization ability of the model.

6. The method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism according to claim 1, characterized in that: The steps for model selection and architecture design in S3 are as follows: S31, select Transformer architecture: adopt the Transformer-based semantic segmentation model; S32, model architecture design: Based on the selected Transformer model, the model's input layer, encoder, decoder, and output layers are appropriately adjusted and optimized according to the characteristics of Sentinel-2 imagery and the requirements of the wheat planting information extraction task.

7. The method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism according to claim 6 is characterized in that: The steps of model training in S3 are as follows: S33, set training parameters: determine the hyperparameters for model training, including learning rate, number of iterations, batch size, and optimizer. Through experimental comparison and parameter tuning, select the optimal hyperparameter combination to improve model training efficiency and segmentation accuracy. S34, loss function definition: Use a loss function suitable for the semantic segmentation task, or combine multiple loss functions to construct a composite loss function to better measure the difference between the model prediction results and the true labels, and guide model training; S35, model training process: The divided training set samples are input into the constructed Transformer model for training. During the training process, the performance indicators of the model on the validation set are monitored in real time. The training parameters are adjusted according to the performance of the validation set to prevent the model from overfitting. The model weight file with the best performance during the training process is saved for subsequent prediction and evaluation.

8. The method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism according to claim 1, characterized in that: The steps for calculating the model evaluation index in S4 are as follows: S41, using the test set evaluation: Apply the trained model to the test set samples, obtain the model's prediction results, and calculate the evaluation indicators to comprehensively evaluate the accuracy and reliability of the model in extracting wheat planting information; S42, Comparative Accuracy Analysis: The evaluation results are compared with those of traditional remote sensing image classification methods to intuitively demonstrate the advantages of the Transformer-based method in improving the accuracy of wheat planting area extraction.

9. The method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism according to claim 8, characterized in that: The steps for model optimization and improvement in S4 are as follows: S43, model fine-tuning: Based on the evaluation results, analyze the prediction errors of the model in different regions and landform types, and fine-tune the model accordingly; S44, Model Fusion: Try to fuse the Transformer-based model with the auxiliary model, comprehensively utilize the advantages of different models, and further improve the accuracy of wheat planting information extraction; S45, post-processing optimization: perform post-processing operations on the model's prediction results to remove isolated noise points and small misclassified areas, smooth the classification boundaries, and improve the integrity and accuracy of the classification results.

10. The method for extracting wheat planting information from Sentinel2 satellite remote sensing images based on the Transformer mechanism according to claim 1, characterized in that: The steps for generating the wheat planting information extraction result in S5 are as follows: S51, classification map production: The optimized and evaluated model was applied to the Sentinel-2 imagery data of the entire study area to obtain a classification map of the wheat planting area; S52, Area Statistics and Report Generation: Based on the classification map, count the wheat planting area in the study area and generate a detailed report; The steps for applying and verifying the results in S5 are as follows: S53, Field Verification and Feedback: Combine field surveys and ground-based measurement data to conduct field verification of the extracted wheat planting information. Compare and analyze the field observation results with the remote sensing interpretation results to evaluate the accuracy and reliability of the interpretation results. Any problems and errors discovered during the field verification process will be promptly fed back to the data processing and model optimization stages to further improve the accuracy and practicality of the wheat planting information extraction. S54, Agricultural Production Monitoring and Management: Apply the extracted wheat planting information to agricultural production monitoring and management. By regularly conducting remote sensing monitoring of wheat planting areas, we can timely grasp the growth status and changing trends of wheat, and provide accurate guidance services for agricultural production.