Multi-task mapping extraction system and method for remote sensing image farmland shelterbelt

Abstract

The application discloses a kind of multi-task mapping extraction system and method for remote sensing image farmland shelter forest, including time window determination module, image acquisition and pretreatment module, annotation dataset construction module, training dataset construction module, improved semantic segmentation model construction module, prediction post-processing and mapping module and multi-task collaborative dataset verification module;The present application is aimed at the characteristics that the structure of farmland shelter forest under remote sensing image presents distribution broken, irregular form, complex boundary and is easily confused with surrounding ground objects, by improving semantic segmentation model, introducing multiple attention mechanisms and jump connection structure, the ability of description to small target and complex boundary is strengthened, effectively solve the problem of fuzzy boundary recognition of traditional model, insufficient fidelity, ensure the accuracy on pixel level scale and the integrity on spatial structure of extraction result.

Description

Technical Field

[0001] This invention relates to the intersection of deep learning semantic segmentation model algorithm improvement and remote sensing mapping technology, specifically a multi-task mapping and extraction system and method for farmland shelterbelts in remote sensing images. Background Technology

[0002] In forestry surveys and monitoring, the delineation of traditional farmland shelterbelts relies on manual labor or field techniques, which is time-consuming and labor-intensive. With the development of remote sensing (RS) technology, remote sensing-based delineation methods have become an efficient alternative, enabling detailed delineation of large areas.

[0003] Farmland shelterbelts, as seen in remote sensing imagery, exhibit a highly fragmented landscape and irregular distribution. Accurate extraction of their cover information is a primary indicator for precision agriculture, and accurate boundary information is crucial for delineating forest distribution. Deep learning (DL) methods are favored due to their superior feature description capabilities and accurate image detection. Compared to traditional methods relying on manually designed features or simple machine learning (ML) algorithms, convolutional neural networks (CNNs) can simultaneously extract shallow structural features and deep semantic features, improving the accuracy and efficiency of image segmentation. They have become the mainstream method for accurate extraction from remote sensing images. Deep learning, through multi-scale feature fusion, attention mechanisms, dilated convolutions, and encoder-decoder structures, captures contextual information at different scales, improving segmentation accuracy, boundary fidelity, and the extraction of complex boundaries and small targets.

[0004] However, existing deep learning methods still have many shortcomings in extracting farmland shelterbelts from remote sensing images: when the target boundaries in the image have complex shapes, blurred transitions, or mixed pixels, conventional models have difficulty accurately distinguishing the boundaries of ground features; areas with irregular target boundaries or rich local features are prone to breakage or mismatch, resulting in insufficient boundary fidelity; multi-scale feature fusion is insufficient, with a strong dependence on local features, failing to fully capture the contextual relationships between targets, affecting the completeness and semantic consistency of the extraction; training the model requires a large amount of dataset, and the creation of datasets is cumbersome and labor-intensive, with high costs for professional annotation, making it difficult to apply to large-scale remote sensing image mapping tasks. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-task mapping and extraction system and method for farmland shelterbelts in remote sensing images, so as to solve the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-task mapping and extraction system for farmland shelterbelts in remote sensing images, comprising a time window determination module, an image acquisition and preprocessing module, a labeled dataset construction module, a training dataset construction module, an improved semantic segmentation model construction module, a prediction post-processing and mapping module, and a multi-task collaborative dataset verification module;

[0007] The time window determination module is used to determine the optimal mapping time window based on the phenological differences between farmland shelterbelts and surrounding land features;

[0008] The image acquisition and preprocessing module is used to acquire remote sensing images of a specified type from the Google Earth Engine platform and provide standardized data based on the preprocessing foundation of the images.

[0009] The labeled dataset construction module adopts a two-stage approach to construct the labeled dataset. In the first stage, preliminary labeled data is generated through manual interpretation and a random forest model is trained to obtain the initial label dataset. In the second stage, the initial label dataset is processed by boundary handling and manually corrected to obtain a high-quality label dataset.

[0010] The training dataset construction module is used to perform format conversion, normalization, cropping and filtering, and proportional division of the labeled data to form training sets and validation sets;

[0011] The improved semantic segmentation model building module is based on DeepLabV3+ and optimizes the network by replacing the backbone network, introducing multiple attention mechanisms and skip connection structures.

[0012] The prediction post-processing and mapping module is used to reassemble and stitch together the model prediction results and perform geographic registration to complete mapping and spatial analysis.

[0013] The multi-task collaborative dataset verification module is used to verify the rationality of the dataset construction method through quantitative evaluation and spatial consistency analysis.

[0014] Furthermore, the optimal mapping time window determined by the time window determination module is May of each year, during which the leaves of farmland shelterbelts grow ahead of time, while crops have not yet fully emerged or are in a fallow phase.

[0015] Furthermore, the remote sensing image acquired by the image acquisition and preprocessing module is a Sentinel-2 Level-2A surface reflectance product. This remote sensing image has undergone radiometric calibration, atmospheric correction, geometric correction, orthorectification, and multi-band spatial registration preprocessing, and is accompanied by quality assessment information on cloud and land cover types.

[0016] Furthermore, in the first stage of the labeled dataset construction module, vector labeled files (.shp) are created using ArcGIS software, and then manually interpreted by professionals to form preliminary labeled data. This data is used to train a random forest model to obtain an initial labeled dataset. In the second stage, the initial labeled dataset undergoes boundary smoothing and adjacent patch merging. In conjunction with the results of field surveys, misclassified and missed areas are manually corrected to form a high-quality labeled dataset.

[0017] Furthermore, the specific operations of the training dataset construction module include: converting the .shp format labeled data into .tif format raster data and performing normalization processing; cropping the image and label data to a size of 256×256 pixels and removing invalid samples; and dividing the training set and validation set into a 7:3 ratio.

[0018] Furthermore, the backbone network of the improved semantic segmentation model construction module is a ResNeSt-50 model that introduces a split-attention mechanism; the various attention mechanisms include an efficient multi-scale attention mechanism, a reverse attention mechanism, and a parallel differential boundary feature attention mechanism; the network structure includes a skip connection structure.

[0019] Furthermore, the prediction post-processing and mapping module generates a mapping result covering the study area by recombining and stitching the prediction results, and uses the "Define Projection" tool of ArcGIS software to assign geographic coordinate information consistent with the original remote sensing image to the prediction result.

[0020] Furthermore, the quantitative evaluation method of the multi-task collaborative dataset verification module calculates the overall accuracy OA, user accuracy UA, producer accuracy PA, and Kappa coefficient by constructing a confusion matrix; spatial consistency analysis statistically analyzes the overlapping area of ​​manually interpreted labels and model predicted labels as "consensus region labels".

[0021] The extraction method for a multi-task mapping extraction system of farmland shelterbelts in remote sensing imagery includes the following steps:

[0022] Step 1: Determine May of each year as the optimal time window for cartography;

[0023] Step 2: Acquire Sentinel-2 Level-2A surface reflectance remote sensing imagery from the GEE platform;

[0024] Step 3: Construct the labeled dataset using a two-stage approach;

[0025] Step 4: After converting, normalizing, cropping and filtering the labeled data, divide the training set and validation set into a 7:3 ratio;

[0026] Step 5: Based on DeepLabV3+, replace the backbone network, introduce multiple attention mechanisms and skip connection structures to build an improved semantic segmentation model;

[0027] Step 6: Reassemble the splicing model prediction results and add geographic coordinate information to complete the mapping;

[0028] Step 7: Verify the rationality of the dataset construction method through quantitative evaluation and spatial consistency analysis.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] This invention addresses the characteristics of farmland shelterbelts in remote sensing imagery: fragmented distribution, irregular morphology, complex boundaries, and easy confusion with surrounding features. By improving the semantic segmentation model and introducing multiple attention mechanisms and skip connection structures, it enhances the ability to characterize small targets and complex boundaries, effectively solving the problems of blurred boundary recognition and insufficient fidelity in traditional models. This ensures the accuracy of the extracted results at the pixel scale and the integrity of the spatial structure. The two-stage dataset construction method—manual interpretation, random forest model assistance, and secondary manual correction—significantly reduces the reliance on large-scale, high-precision, purely manually labeled data. While maintaining dataset quality, it greatly reduces the time and manpower required for sample construction, overcoming the cumbersome dataset creation in traditional deep learning methods. The proposed multi-task collaborative mapping framework does not rely on a single model structure. Through optimized model design and dataset construction processes, it can adapt to the application needs of remote sensing imagery in different regions and environments. Multi-regional experiments have verified that it exhibits stable mapping performance in various scenarios and has broad applicability. By introducing an efficient multi-scale attention mechanism and a skip connection structure, the contextual relationships between targets are fully captured, reducing the problem of missing feature context, improving the semantic consistency and overall integrity of the extraction results, and avoiding problems such as breakage and incorrect matching caused by insufficient feature fusion in traditional models. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method in an embodiment of the present invention;

[0032] Figure 2 This is a diagram showing the differences in phenology among different typical land features in this embodiment of the invention;

[0033] Figure 3 This is a structural diagram of the improved semantic segmentation model in an embodiment of the present invention;

[0034] Figure 4 This is a detailed and intuitive visualization of the farmland shelterbelt datasets created using two different methods in this embodiment of the invention. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0036] Please see Figure 1-4 This invention provides a multi-task mapping and extraction system for farmland shelterbelts in remote sensing images, including a time window determination module, an image acquisition and preprocessing module, a labeled dataset construction module, a training dataset construction module, an improved semantic segmentation model construction module, a prediction post-processing and mapping module, and a multi-task collaborative dataset verification module.

[0037] The time window determination module is used to determine the optimal mapping time window based on the phenological differences between farmland shelterbelts and surrounding land features.

[0038] The image acquisition and preprocessing module is used to acquire remote sensing images of a specified type from the Google Earth Engine platform and provide standardized data based on the preprocessing capabilities of the images themselves.

[0039] The labeled dataset construction module adopts a two-stage approach to build the labeled dataset. In the first stage, preliminary labeled data is generated through manual interpretation and a random forest model is trained to obtain the initial label dataset. In the second stage, the initial label dataset is processed for boundaries and manually corrected to obtain a high-quality label dataset.

[0040] The training dataset construction module is used to convert, normalize, crop, filter, and proportionally divide the labeled data to form training and validation sets.

[0041] The semantic segmentation model building module is improved based on DeepLabV3+, and the network is optimized by replacing the backbone network, introducing multiple attention mechanisms and skip connection structures.

[0042] The prediction post-processing and mapping module is used to reassemble and stitch together the model prediction results and perform georegistration to complete mapping and spatial analysis.

[0043] The multi-task collaborative dataset validation module is used to verify the rationality of the dataset construction method through quantitative evaluation and spatial consistency analysis.

[0044] The optimal mapping time window determined by the time window determination module is May of each year. During this time window, the leaves of farmland shelterbelts grow early, while crops have not yet fully emerged or are in the fallow stage.

[0045] The remote sensing image acquired by the image acquisition and preprocessing module is the Sentinel-2 Level-2A surface reflectance product. This remote sensing image has undergone radiometric calibration, atmospheric correction, geometric correction, orthorectification and multi-band spatial registration preprocessing, and is accompanied by quality assessment information on cloud and land cover types.

[0046] The first stage of the labeled dataset construction module involves creating vector label files (.shp) using ArcGIS software. These files are then manually interpreted by professionals to form preliminary labeled data. This data is used to train a random forest model to obtain the initial label dataset. The second stage involves smoothing the boundaries of the initial label dataset, merging adjacent patches, and manually correcting misclassified and missed areas based on field survey results, thus forming a high-quality label dataset.

[0047] The specific operations of the training dataset construction module include: converting the .shp format labeled data into .tif format raster data and performing normalization processing; cropping the image and label data to a size of 256×256 pixels and removing invalid samples; and dividing the training set and validation set into a 7:3 ratio.

[0048] The backbone network of the improved semantic segmentation model construction module is the ResNeSt-50 model with a split-attention mechanism; the various attention mechanisms include an efficient multi-scale attention mechanism, a reverse attention mechanism, and a parallel differential boundary feature attention mechanism; the network structure includes a skip connection structure.

[0049] The prediction post-processing and mapping module generates mapping results covering the study area by recombining and stitching the prediction results, and uses ArcGIS software's "Define Projection" tool to assign geographic coordinate information consistent with the original remote sensing imagery to the prediction results.

[0050] The quantitative evaluation method for the multi-task collaborative dataset validation module is to construct a confusion matrix and calculate the overall accuracy (OA), user accuracy (UA), producer accuracy (PA), and Kappa coefficient. Spatial consistency analysis is to statistically analyze the overlapping area between manually interpreted labels and model predicted labels as "consensus region labels".

[0051] The extraction method for a multi-task mapping extraction system of farmland shelterbelts in remote sensing imagery includes the following steps:

[0052] Step 1: Determine the optimal time window for cartography

[0053] By analyzing the changes in remote sensing features of farmland shelterbelts, crops, bare soil, and other typical land features at different stages throughout the year, May was determined to be the optimal mapping window. At this time, the leaves of farmland shelterbelts have grown earlier, while crops have not yet fully emerged or are in a fallow stage, resulting in minimal background interference. This approach effectively identifies significant temporal differences between farmland shelterbelts and surrounding crops, enhancing the classification accuracy and boundary detail recognition capabilities of subsequent models.

[0054] Step 2: Image Acquisition and Preprocessing

[0055] Sentinel-2 Level-2A surface reflectance remote sensing imagery was acquired from the Google Earth Engine (GEE) platform. This imagery has undergone preprocessing operations such as radiometric calibration, atmospheric correction, geometric correction, orthorectification, and multi-band spatial registration, and includes quality assessment information on cloud and land cover types, providing basic data support for subsequent image selection, feature extraction, and model training.

[0056] Step 3: Construct the labeled dataset

[0057] A two-stage approach was used to construct the farmland shelterbelt annotation dataset:

[0058] 3.1 Vector annotation files (.shp) are created using ArcGIS software. These files are then manually interpreted by professionals with remote sensing interpretation experience to determine the land cover categories and generate preliminary annotation data. This data is used as training samples for a random forest model, and the initial farmland shelterbelt label dataset is obtained through classification using the random forest model.

[0059] 3.2 Based on the initial labeled dataset, the boundaries of farmland shelterbelts were smoothed and adjacent patches were merged. Combined with the results of field surveys and verification, misclassified and omitted areas were manually corrected to form a high-quality farmland shelterbelt labeled dataset after secondary optimization and correction.

[0060] Step 4: Construct the training dataset

[0061] 4.1 Convert the .shp format annotation data obtained in the two methods in step 3 into .tif format raster data and normalize the label data;

[0062] 4.2 While maintaining the same spatial resolution and size as the original remote sensing image, the image and corresponding label data are cropped to a fixed size of 256×256 pixels, and sample data that do not contain valid label information or have irregular edges are removed.

[0063] 4.3 The selected sample data were divided into a model training set and a validation set in a 7:3 ratio to construct a high-quality dataset for training and performance evaluation of the semantic segmentation model for farmland shelterbelts.

[0064] Step 5: Construct the improved semantic segmentation model

[0065] Using DeepLabV3+ as the baseline semantic segmentation framework, its network structure is improved in a targeted manner:

[0066] 5.1 In the backbone network, the traditional residual network structure is replaced with the ResNeSt-50 model that introduces the Split-Attention mechanism to enhance the model's adaptive modeling ability for key channel features;

[0067] 5.2 An efficient multi-scale attention mechanism is introduced into the overall model framework encoder to effectively preserve the local detail information of high-level and low-level features during the multi-scale feature fusion process;

[0068] 5.3 Before the feature fusion decoding stage, a reverse attention mechanism is introduced to differentiate the target area and background area of ​​farmland shelterbelt, thereby improving the model's ability to distinguish the target area and the fidelity of feature extraction.

[0069] 5.4 By using a parallel differential boundary feature attention mechanism, the boundary and contour information of regions with significant feature changes are highlighted, thereby enhancing the model's ability to express the boundary structure of farmland shelterbelts.

[0070] 5.5 An additional skip connection structure is introduced into the network structure to enhance the information transmission and correlation between features at different levels, improve the ability to retain fine-grained features during convolution, and achieve high-precision boundary extraction and segmentation of small targets such as farmland shelterbelts in remote sensing images.

[0071] Step 6: Post-processing and plotting of prediction results

[0072] 6.1 The prediction results output by the trained semantic segmentation model are reorganized and stitched together according to the spatial position relationship of each prediction block in the original remote sensing image to generate a continuous farmland shelterbelt distribution map covering the entire study area.

[0073] 6.2 Using the "Define Projection" tool in ArcGIS software, the stitched .tif format prediction results are reassigned with geographic coordinate information consistent with the original remote sensing image, realizing the spatial positioning and geographic registration of the prediction results, and completing the mapping and spatial analysis of farmland shelterbelts.

[0074] Step 7: Multi-task collaborative dataset validation

[0075] We used a combination of quantitative evaluation methods and spatial consistency statistical analysis to verify the rationality and feasibility of two methods for constructing labeled datasets:

[0076] 7.1 Confusion Matrix is ​​constructed, using the labeled data obtained by manual interpretation as the real labels and the classification results of the random forest model as the predicted labels. Evaluation metrics such as overall accuracy (OA), user accuracy (UA), producer accuracy (PA) and Kappa coefficient are calculated to objectively evaluate the reliability of the classification results of the random forest model.

[0077] 7.2 Spatial overlay analysis is performed on manually interpreted labels and model predicted labels. The spatially overlapping area is defined as the "consensus region label". Statistical analysis is conducted on its distribution characteristics to verify the consistency and stability between the two dataset construction methods at the spatial level, thereby enhancing the objectivity and rigor of the verification results.

[0078] Example:

[0079] This invention combines a multi-task collaborative mapping framework with an improved semantic segmentation model, and applies it to the task of extracting and mapping farmland shelterbelts from remote sensing images. Through a dataset construction method of "model + manual correction" and an optimized model structure, it achieves high-efficiency and high-precision extraction and identification of farmland shelterbelts.

[0080] Detailed implementation process:

[0081] Step 1: Determining the optimal mapping time window

[0082] By statistically analyzing and comparing the changes in remote sensing features of farmland shelterbelts, crops, bare soil, and other typical land features at different stages throughout the year in the study area, the spectral differences and separability of various land features in different months were comprehensively examined. The results showed that in May, farmland was in the bare soil stage or the crops had not yet fully emerged, and some plots were in a fallow state. However, farmland shelterbelts, due to the relatively early budding and leaf growth of trees or shrubs, exhibited higher vegetation coverage and obvious spectral response characteristics in remote sensing images. This effectively reduced the land feature confusion caused by the vigorous growth period of crops, thereby enhancing the identifiability and separability of farmland shelterbelts in remote sensing images. Therefore, May was determined to be the optimal mapping time window.

[0083] Step 2: Image Acquisition and Preprocessing

[0084] Remote sensing images of the Sentinel-2 satellite's Level-2A surface reflectance product were obtained from the Google Earth Engine (GEE) cloud computing platform. Before publication, these images underwent standardized preprocessing operations such as radiometric calibration, atmospheric correction, geometric correction, orthorectification, and spatial registration between multiple bands. This effectively eliminates the impact of sensor differences and atmospheric environment on image quality, providing high-quality basic data for subsequent feature extraction and model training.

[0085] Step 3: Construction of labeled dataset

[0086] 3.1 Initial Labeling Dataset Construction: Vector label files (.shp) were created using ArcGIS software. Experienced professionals in remote sensing interpretation manually interpreted and labeled the land cover types within the study area, forming preliminary vector labeling data for farmland shelterbelts. This preliminary labeled data was used as training samples for a random forest model. After training, the trained model was used to classify remote sensing images of the study area, outputting the initial farmland shelterbelt label dataset.

[0087] 3.2 Construction of High-Quality Labeled Dataset: Based on the initial label dataset, ArcGIS software was used to smooth the boundaries of farmland shelterbelt patches, and adjacent patches that met the merging criteria were reasonably merged. Subsequently, staff conducted field surveys and on-site verifications, manually correcting and supplementing the labeling of misclassified areas (areas that were not farmland shelterbelts but were mistakenly identified as farmland shelterbelts) and omitted areas (areas of unidentified farmland shelterbelts) in the initial label data, ultimately forming a high-quality farmland shelterbelt label dataset after secondary optimization and correction.

[0088] Step 4: Training Dataset Construction

[0089] 4.1 Data Format Conversion and Normalization: The two types of farmland shelterbelt label data obtained in Step 3 were uniformly converted into .tif format raster label data using professional tools. To meet the input data requirements of the deep learning semantic segmentation model, the converted raster label data was normalized to ensure that the data values ​​were within a uniform and reasonable range.

[0090] 4.2 Data Cropping and Filtering: During the conversion process, it is ensured that the generated raster label data maintains consistency with the corresponding original remote sensing image in terms of spatial resolution, spatial reference, and image size, guaranteeing the spatial alignment accuracy between the image data and the label data. Based on this, the remote sensing image data and corresponding label data are cropped to a fixed size of 256×256 pixels to generate sample data blocks. Simultaneously, the cropped samples are manually filtered to remove samples that do not contain valid farmland shelterbelt label information, or that have irregular edges or obvious data anomalies, thereby improving the effectiveness and consistency of the training data.

[0091] 4.3 Training and Validation Set Division: The selected valid sample data is divided into a training set and a validation set in a 7:3 ratio. 70% of the samples are used to improve the training process of the semantic segmentation model, and 30% of the samples are used for performance evaluation and parameter adjustment during the model training process, thus constructing a high-quality dataset for training and performance evaluation of the semantic segmentation model for farmland shelterbelts.

[0092] Step 5: Improve the semantic segmentation model construction

[0093] Using DeepLabV3+ as the basic semantic segmentation framework, while maintaining its overall encoding-decoding structure, targeted improvements and optimizations were made to key network modules:

[0094] 5.1 Backbone Network Replacement: The traditional residual network structure is replaced with a ResNeSt-50 model that introduces a split-attention mechanism, enabling the network to adaptively model the importance of different channels during feature extraction, thereby enhancing its ability to express key spectral and structural features of farmland shelterbelts.

[0095] 5.2 Introduction of efficient multi-scale attention mechanism: An efficient multi-scale attention mechanism is introduced into the overall model framework. This mechanism can adjust the weights of features at different scales during the multi-scale feature fusion process, effectively preserve local detailed feature information, reduce the impact of scale changes on the segmentation results, and enable the model to capture both global features and local detailed features of farmland shelterbelts.

[0096] 5.3 Introduction of Reverse Attention Mechanism: Before the feature fusion decoding stage, a reverse attention mechanism is introduced. By differentiating and strengthening the target area and background area of ​​farmland shelterbelt, the feature interference of the background area is suppressed, the model’s attention to the target area and the fidelity of feature extraction are improved, and the model can more accurately identify the farmland shelterbelt area.

[0097] 5.4 Introduction of Parallel Differential Boundary Feature Attention Mechanism: By using the parallel differential boundary feature attention mechanism, the boundary and contour information of the fused feature change area (i.e. the boundary area between farmland shelterbelt and surrounding land features) is highlighted, which enhances the model's ability to express the boundary structure and morphological features of farmland shelterbelt and improves the accuracy of boundary extraction.

[0098] 5.5 Introduction of Skip Connection Structure: Skip connection structure is introduced into the network structure to strengthen the information transmission and fusion between features at different levels, improve the collaborative expression ability of deep semantic features and shallow spatial detail features, thereby improving the model's boundary extraction accuracy and overall segmentation effect for small targets such as farmland shelterbelts in remote sensing images.

[0099] Step 6: Post-processing and plotting of prediction results

[0100] 6.1 Recombination and stitching of prediction results: The prediction results obtained by the trained farmland shelterbelt semantic segmentation model on remote sensing images are recombined and stitched according to the spatial position relationship of each prediction block in the original remote sensing image to generate a continuous farmland shelterbelt distribution map covering the entire study area, ensuring the integrity of the map results.

[0101] 6.2 Geographic Registration and Positioning: For the stitched .tif format prediction results, the "Define Projection" tool in ArcGIS software is used to reassign them with a coordinate reference system and geographic projection information consistent with the corresponding original remote sensing image. This achieves spatial positioning and geographic registration between the prediction results and the original image, ensuring the spatial accuracy of the mapping results and completing the spatial representation of farmland shelterbelt distribution information.

[0102] Step 7: Multi-task collaborative dataset validation

[0103] 7.1 Quantitative Evaluation: A quantitative evaluation method was used to compare and analyze the consistency between the prediction results of the random forest model and the results of human visual interpretation. A confusion matrix was constructed, using the labeled data obtained from human interpretation as the true labels and the classification results of the random forest model as the predicted labels. Evaluation metrics such as Overall Accuracy (OA), User Accuracy (UA), Producer Accuracy (PA), and Kappa coefficient were calculated to objectively assess the reliability of the random forest model's classification results.

[0104] 7.2 Spatial Consistency Analysis: Spatial overlay analysis is performed on manually interpreted labels and model predicted labels. The spatially overlapping area is defined as the "consensus region label". The distribution characteristics and area of ​​the "consensus region label" are statistically analyzed to further verify the consistency and stability between the two dataset construction methods from a spatial perspective, thereby enhancing the objectivity and rigor of the verification results of the multi-task collaborative dataset.

[0105] Experimental results:

[0106] This invention verifies the superiority of the improved model and dataset construction method through comparative experiments and quantitative evaluation. The specific experimental results are shown in the table below:

[0107] Table 1. Model Performance Comparison Results

[0108]

[0109] Table 2 Comparison of the accuracy of the datasets created using the two methods for model training

[0110]

[0111] Table 3. Accuracy comparison of the two datasets used for the confusion matrix.

[0112]

[0113] Table 4. Comparison of Geographic Geometric Information Based on Two Dataset Production Methods

[0114]

[0115] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-task mapping and extraction system for farmland shelterbelts from remote sensing images, characterized in that: It includes a time window determination module, an image acquisition and preprocessing module, a labeled dataset construction module, a training dataset construction module, an improved semantic segmentation model construction module, a prediction post-processing and mapping module, and a multi-task collaborative dataset validation module; The time window determination module is used to determine the optimal mapping time window based on the phenological differences between farmland shelterbelts and surrounding land features; The image acquisition and preprocessing module is used to acquire remote sensing images of a specified type from the Google Earth Engine platform and provide standardized data based on the preprocessing foundation of the images. The labeled dataset construction module adopts a two-stage approach to construct the labeled dataset. In the first stage, preliminary labeled data is generated through manual interpretation and a random forest model is trained to obtain the initial label dataset. In the second stage, the initial label dataset is processed by boundary handling and manually corrected to obtain a high-quality label dataset. The training dataset construction module is used to perform format conversion, normalization, cropping and filtering, and proportional division of the labeled data to form training sets and validation sets; The improved semantic segmentation model building module is based on DeepLabV3+ and optimizes the network by replacing the backbone network, introducing multiple attention mechanisms and skip connection structures. The prediction post-processing and mapping module is used to reassemble and stitch together the model prediction results and perform geographic registration to complete mapping and spatial analysis. The multi-task collaborative dataset verification module is used to verify the rationality of the dataset construction method through quantitative evaluation and spatial consistency analysis.

2. The multi-task mapping and extraction system for farmland shelterbelts from remote sensing images according to claim 1, characterized in that: The optimal mapping time window determined by the time window determination module is May of each year. During this time window, the leaves of farmland shelterbelts grow early, while crops have not yet fully emerged or are in the fallow stage.

3. The multi-task mapping and extraction system for farmland shelterbelts from remote sensing images according to claim 1, characterized in that: The remote sensing image acquired by the image acquisition and preprocessing module is a Sentinel-2 Level-2A surface reflectance product. This remote sensing image has undergone radiometric calibration, atmospheric correction, geometric correction, orthorectification, and multi-band spatial registration preprocessing, and is accompanied by quality assessment information on cloud and land cover types.

4. The multi-task mapping and extraction system for farmland shelterbelts from remote sensing images according to claim 1, characterized in that: The first stage of the labeled dataset construction module involves creating vector label files (.shp) using ArcGIS software. These files are then manually interpreted by professionals to form preliminary labeled data. This data is used to train a random forest model to obtain the initial label dataset. The second stage involves smoothing the boundaries of the initial label dataset, merging adjacent patches, and manually correcting misclassified and missed areas based on field survey results, thus forming a high-quality label dataset.

5. The multi-task mapping and extraction system for farmland shelterbelts from remote sensing images according to claim 1, characterized in that: The specific operations of the training dataset construction module include: converting .shp format labeled data into .tif format raster data and performing normalization processing; cropping the image and label data to a size of 256×256 pixels and removing invalid samples; and dividing the training set and validation set into a 7:3 ratio.

6. The multi-task mapping and extraction system for farmland shelterbelts from remote sensing images according to claim 1, characterized in that: The backbone network of the improved semantic segmentation model construction module is a ResNeSt-50 model with a split attention mechanism; the various attention mechanisms include an efficient multi-scale attention mechanism, a reverse attention mechanism, and a parallel differential boundary feature attention mechanism; the network structure includes a skip connection structure.

7. The multi-task mapping and extraction system for farmland shelterbelts from remote sensing images according to claim 1, characterized in that: The prediction post-processing and mapping module generates mapping results covering the study area by recombining and stitching the prediction results, and uses the "Define Projection" tool of ArcGIS software to assign geographic coordinate information consistent with the original remote sensing image to the prediction results.

8. The multi-task mapping and extraction system for farmland shelterbelts from remote sensing images according to claim 1, characterized in that: The quantitative evaluation method of the multi-task collaborative dataset verification module is to construct a confusion matrix and calculate the overall accuracy OA, user accuracy UA, producer accuracy PA and Kappa coefficient; the spatial consistency analysis is to perform statistical analysis by taking the overlapping area of ​​manually interpreted labels and model predicted labels as "consensus area labels".

9. The extraction method of the multi-task mapping extraction system for farmland shelterbelts in remote sensing imagery according to any one of claims 1-8, characterized in that: Includes the following steps: Step 1: Determine May of each year as the optimal time window for cartography; Step 2: Acquire Sentinel-2 Level-2A surface reflectance remote sensing imagery from the GEE platform; Step 3: Construct the labeled dataset using a two-stage approach; Step 4: After converting, normalizing, cropping and filtering the labeled data, divide the training set and validation set into a 7:3 ratio; Step 5: Based on DeepLabV3+, replace the backbone network, introduce multiple attention mechanisms and skip connection structures to build an improved semantic segmentation model; Step 6: Reassemble the splicing model prediction results and add geographic coordinate information to complete the mapping; Step 7: Verify the rationality of the dataset construction method through quantitative evaluation and spatial consistency analysis.

Images