Automatic sample construction detection model design method for forest and grassland wet change

Abstract

This invention belongs to the field of remote sensing image processing and data mining technology. It proposes an automatic sample construction and detection model design method for forest and grassland wetland changes. The method forms a dataset by screening images after sliding window cropping and constructs a change detection model. It adopts a dual-temporal shared weight encoding + multi-branch decoding structure, and separately constructs encoders to realize feature extraction, a preliminary discrimination branch to output the probability value of each pixel belonging to the forest and grassland wetland range, and a degradation change detection branch to output the prediction result of forest and grassland wetland degradation change. This invention solves the problem of traditional general change detection models misreporting other land types as degradation, and significantly improves the specificity and accuracy of forest and grassland wetland degradation monitoring.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and data mining technology. Background Technology

[0002] Forest, grassland, and wetland resources are core components of terrestrial ecosystems. Timely identification and quantification of changes that degrade into non-forest, grassland, and wetland areas have become an urgent need for natural resource management.

[0003] Currently, the detection methods for forest and grassland wetland changes mainly revolve around two technical approaches. The first category is time-series analysis methods based on medium- and low-resolution remote sensing imagery. These methods rely on time-series data such as Landsat, Sentinel-2, or MODIS to detect gradual changes in land cover by constructing long-term vegetation indices (such as NDVI and EVI) or spectral change analysis models. However, limited by their spatial resolution (typically 10 meters to several hundred meters), these methods struggle to accurately identify small-scale, sudden changes and have inherent limitations in locating change boundaries and identifying change types.

[0004] The second category is deep learning methods based on two periods of high-resolution remote sensing imagery. This method utilizes sub-meter resolution satellite or aerial imagery and employs deep learning models such as Siamese networks, Transformer, and Mamba to directly learn the differences between the two periods of imagery, achieving pixel-level or object-level change recognition. Because this method performs well in refining change boundary extraction and patch integrity, it is gradually becoming a major technical direction for refined detection. However, this method is extremely dependent on a large number of high-quality, paired-annotated training samples. Forestry, grassland, and wetland resources are widely distributed and have strong environmental heterogeneity; manual visual interpretation of samples requires enormous human, material, and time costs, and annotation consistency is difficult to guarantee. Furthermore, existing deep learning models are mostly general change detection frameworks and do not embed the core rule that "the target in the early detection stage must be forestry, grassland, or wetland" into the model learning process. This leads to the model identifying changes that are not the target of detection, such as "non-forestry, grassland, or wetland → non-forestry, grassland, or wetland" (e.g., from bare land to buildings) or "non-forestry, grassland, or wetland → forestry, grassland, or wetland" (e.g., revegetation projects), as "changes," resulting in a large number of operationally significant false detections and seriously affecting the specificity and accuracy of degradation detection.

[0005] In summary, there is currently a lack of a specialized detection technology for forest and grassland degradation changes that can effectively address the two key issues of automatic sample construction and deep integration of domain knowledge into the model. Summary of the Invention

[0006] In order to build a dedicated deep learning model that can deeply understand the business logic that "only when the forest and grassland wetland changes to a non-forest and grassland wetland does it belong to the degradation that should be detected", and to solve the technical problem that the current training sample formation cost is too high, this invention proposes an automatic sample construction and detection model design method for forest and grassland wetland changes.

[0007] An automated sample construction and detection model design method for forest and grassland wetland change includes the following steps: Step 1: Preparation and preprocessing of multi-source forest and grassland wetland monitoring data. The monitoring data includes early and late remote sensing images of the same area, vector data of the previous year's forest and grassland wetland resource base map, and vector data of multiple periods of forest and grassland wetland resource change monitoring results to ensure pixel matching between early and late images. Step 2: Set a fixed-size sliding window and traversal step size, and read the image from Step 1 to perform sliding window cropping; Step 3: Filter the images after sliding window cropping to form a dataset for training the change detection model; Step 4: Construct a change detection model: A dual-temporal shared weight encoding + multi-branch decoding structure is adopted. The encoder is constructed to realize feature extraction, the early discrimination branch outputs the probability value of each pixel belonging to the forest and grassland wetland range, and the degradation change detection branch outputs the prediction result of forest and grassland wetland degradation change. Step 5: Train and optimize the parameters of the change detection model based on the multi-task joint loss function.

[0008] Furthermore, step 1 specifically includes: Data preparation and input: Collect previous and subsequent remote sensing images of the same area, vector data of the previous year's forest and grassland wetland resource base map, and vector data of multiple periods of forest and grassland wetland resource change monitoring results. The previous and subsequent remote sensing images are time-series high-resolution remote sensing images, and have undergone preprocessing such as radiometric correction and orthorectification; various vector data have undergone preprocessing operations such as coordinate transformation and map cropping.

[0009] Spatial alignment and inspection: Check whether the spatial resolution, geographic transformation parameters, and coordinate system are consistent between the earlier and later imagery. If discrepancies exist, use the later imagery as a reference and perform geometric correction and spatial alignment on the earlier imagery using reprojection and resampling algorithms to ensure strict pixel matching between the two imagery sets. Check whether the vector and data range boundaries and coordinate references are consistent. If not, use the imagery as a reference and perform coordinate transformation and range cropping.

[0010] Furthermore, step 2 specifically includes: Based on the spatially aligned data, a fixed-size sliding window and traversal step size are set to perform sliding window cropping on the image. The specific steps for each window position are as follows: Read the image data of corresponding areas in the earlier and later images respectively; The vector data of change detection results are used to generate an initial change label map containing only two types of labels: "change" and "unchanged". The changed areas are assigned a value of 1, and the unchanged areas are assigned a value of 0. The base map vector data of forest and grassland wetland resources is rasterized to generate a forest and grassland wetland range mask, with a value of 1 within the range and a value of 255 outside the range; The initial change label map is fused with the forest-grassland-wet range mask to generate the final sample label map. The rules are as follows: within the forest-grassland-wet range, the initial change label (0 or 1) is retained; outside the forest-grassland-wet range, it is uniformly labeled as 255, marking the "background" area, which does not participate in the loss calculation during model training.

[0011] Furthermore, step 3 specifically involves: The collected forest and grassland wetland degradation samples were screened, and the samples were output and the dataset was divided for subsequent model training. The specific steps are as follows: Sample screening: To alleviate the sample imbalance problem, the proportion of "changing" pixels in each sample to the total number of pixels is calculated; if the proportion is lower than the set minimum change threshold, the sample is considered to have insufficient change information and is removed. Sample Output: The filtered pre-image slices, post-image slices, and corresponding sample label slices are saved as image format files. All samples are named according to predetermined rules and stored in corresponding directories to form a structured change detection training sample set; The dataset is divided into training set: validation set: test set in a ratio of 4:1:1.

[0012] Furthermore, step 4 specifically involves: Encoder Construction: A deep learning backbone network based on a self-attention mechanism was selected as the feature encoder. A dual-branch structure with shared weights was adopted to extract features from earlier and later remote sensing images respectively. The input image size was consistent with the sample cropping size, and the multi-channel feature map was output after encoding.

[0013] Early-stage discrimination branch construction: Using the encoded features of the earlier images as input, an early-stage discrimination branch is constructed. This branch consists of multiple convolutional operations, and the output is a single-channel probability map with the same spatial resolution as the input features, used to represent the probability value of each pixel belonging to the forest-grassland-wetland area. This probability map is used to explicitly characterize the spatial distribution of forest-grassland-wetland areas in the early monitoring stage.

[0014] Degradation Change Detection Branch Construction: Pixel-by-pixel absolute difference calculation is performed on the features of the earlier and later images to obtain an earlier-to-later feature difference map. Then, this feature difference map is concatenated with the probability map output by the earlier forestry and grassland wetland discrimination branch in the channel dimension to form a fused feature. Based on this fused feature, a directional degradation change detection branch is constructed to output the predicted results of forestry and grassland wetland degradation changes.

[0015] Furthermore, step 5 specifically includes: The model training phase employs a multi-task joint loss function for optimization, with the specific steps as follows: Early forest and grassland wet discrimination loss: Based on the rule that non-255 pixels in the label are forest and grassland wet, a binary classification cross-entropy loss is adopted, and supervision is only carried out in the non-background area to constrain the model to correctly identify the range of early forest and grassland wet; Degradation change detection loss: Based on the final label map, a binary classification cross-entropy loss is adopted, and supervision is only performed on the real change pixels (Y=1) to ensure that the model learns degradation change features only in the forest and grassland wetland area; The total loss is a weighted combination of the two losses mentioned above. The weights are optimized through cross-validation on the validation set to achieve business rule-driven targeted change detection.

[0016] Technical effects: This invention fully utilizes historical monitoring results (such as resource base maps and past change patches) to automatically generate forest and grassland wetland change samples by overlaying the base map and change map, including "unchanged," "degraded," and "uninterested areas." Compared with traditional manual visual annotation, this method greatly reduces the cost and time of sample acquisition, ensures the consistency of sample annotation standards, and provides a reliable data foundation for efficient model training and iterative updates.

[0017] By introducing a pre-construction forest-grassland wetland discrimination branch into the model and explicitly integrating the discrimination results into the degradation change detection branch, false detection responses in non-forest-grassland wetland areas are suppressed from the ground up. Pre-construction and post-construction images are input separately to obtain pre-construction forest-grassland wetland discrimination results and degradation change detection results. Figure 3 As shown, Figure 3 In the graph, red represents areas outside the forest-grassland wetland zone, green represents areas of change within the forest-grassland wetland zone, and blue-purple represents areas without change within the forest-grassland wetland zone. The system retains only pixels that simultaneously meet the degradation change discrimination criteria within areas initially identified as forest-grassland wetlands as the final forest-grassland wetland degradation change result. All other areas are uniformly classified as non-degradation changes, thus completely suppressing false detections in areas of non-operational interest. This solves the problem of traditional general change detection models misreporting other land types as degradation, significantly improving the specificity and accuracy of forest-grassland wetland degradation monitoring.

[0018] The core idea of ​​this invention, "using prior knowledge graphs to generate samples and constrain models," is not only applicable to monitoring forest and grassland degradation, but can also be applied to other change detection scenarios with clear land use conversion definitions, such as the "non-agriculturalization" of arable land, illegal expansion of construction land, and wetland encroachment. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall process of this invention.

[0020] Figure 2 This is a block diagram of the internal network structure of the change detection model.

[0021] Figure 3 This is a diagram showing the detection results output by this invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0023] like Figure 1 As shown in the figure, this embodiment proposes an automatic sample construction and detection model design method for forest and grassland wetland change, including the following steps: Step 1: Based on existing forest and grassland wetland resource monitoring results, complete the preparation and spatial alignment of multi-source forest and grassland wetland monitoring data. The specific steps are as follows: Step 1.1, Data Preparation and Input: Collect the following three types of data, all derived from actual forestry and grassland wetland resource monitoring operations: a) Pre- and post-monitoring remote sensing images (a) of the same area at different time phases, preprocessed with radiometric and orthorectification corrections, and stored in GeoTIFF format; b) a vector base map of the previous year's forestry and grassland wetland resource extent, defining the spatial distribution range of forest, grassland, and wetland resources to be monitored, stored in Shapefile format; c) Vector data of multi-period forestry and grassland wetland resource change monitoring results, stored in Shapefile format, identifying change patches that meet monitoring standards and transform from forestry and grassland wetlands to non-forestry and grassland wetlands. Perform coordinate transformation and map cropping preprocessing operations on the above vector data.

[0024] Step 1.2, Spatial Alignment and Check: Spatial metadata from both the earlier and later imagery sets is retrieved using a geographic information processing (GIS) library. This metadata includes the Coordinate Reference System (CRS), geographic transformation parameters (Transform), and image dimensions. A custom program automatically checks whether the width and height differences between the two images are within a 3-pixel tolerance and whether the CRS and Transform are completely consistent. If the check fails, the image is considered misaligned, and an alignment process is automatically triggered: using the later imagery's CRS and Transform as a reference, reprojection and bilinear resampling are performed on each band of the earlier imagery to generate a new earlier imagery file that strictly corresponds one-to-one with the spatial location of the later imagery pixels, ensuring that both have the same number of rows and columns. The vectors used are read, and the data range boundaries and coordinate references between vectors and between vectors and the image are checked for consistency. If inconsistent, coordinate transformation and range clipping are performed using the image as a reference.

[0025] Step 2: Based on the spatially aligned data, set a fixed-size sliding window and traversal step size, and perform sliding window cropping on the image. The specific steps are as follows: A fixed sliding window size (tile_size) of 512×512 pixels and a sliding step of 500 pixels are set. The traversable window range is calculated with the top-left corner of the image as the origin (0, 0). For each window's starting coordinates (xoff, yoff), where xoff is the column index of the top-left pixel in the image and yoff is the row index, a geographic window is first constructed: based on xoff, yoff, and the sliding window size, the preceding and following image band data for the corresponding area are read. Next, change information is rasterized: the change monitoring vector patch covering the current geographic window is rasterized into a two-dimensional array tile_label of size tile_size × tile_size. Pixels covered by the vector patch are assigned a value of 1 ("change"), and other pixels are assigned a value of 0 ("no change"). Forest, grassland, and wetland range mask generation: the forest, grassland, and wetland resource range vector base map covering the current window is rasterized into a range mask tile_range. Pixels within the range are assigned a value of 1, and pixels outside the range are assigned a value of 255. Finally, the final label map is generated by fusion: pixel-level logical fusion is performed to generate labels. That is, only when a pixel is within the forest / grass / wet range (tile_range value is 1) will its final label inherit the value of tile_label (0 or 1); for all pixels outside the forest / grass / wet range (tile_range value is 255), its final label will be uniformly forced to be 255. The value 255 is defined as "ignore_index" in subsequent model training, and the corresponding pixel will not participate in the calculation of the loss function and gradient update.

[0026] Step 3: Screen the collected forest and grassland wetland degradation samples, and output and divide the samples for subsequent model training. The specific steps are as follows: Step 3.1, Sample Screening: To alleviate the sample imbalance problem and increase the proportion of changing pixels, after generating the tile_label for each window, the proportion of changing pixels is calculated as: number of pixels with a value of 1 in the tile_label / total number of pixels in the tile_label. A minimum change threshold (0.001) is preset. If the proportion of changing pixels in the current window is less than this threshold, it is determined that the window does not contain enough change information, and the window is skipped without image data reading, label fusion, or subsequent saving operations. The next sliding window is then processed directly.

[0027] Step 3.2, Sample Output: For windows that pass the sample filtering, the previous image patch, the later image patch, and the final label image are processed and saved using the `_save_tile` method. This method adjusts the data dimensions to the standard image channel order and converts it to an 8-bit unsigned integer format. Finally, it saves the data as a PNG file using `PIL.Image.fromarray`, thus forming a well-structured dataset for detecting forest and grassland wetland degradation changes that can be directly read by the data loader of a deep learning framework.

[0028] Furthermore, the dataset is divided into training set:validation set:test set in a ratio of 4:1:1. During the division, it is ensured that varied samples are evenly distributed across different subsets, and that training samples do not appear in the validation or test subsets to avoid data leakage. The training set is used for learning model parameters, the validation set is used for parameter tuning and early stopping detection, and the test set is used for final model performance evaluation.

[0029] Step 4: As Figure 2 As shown, a change detection model for detecting forest and grassland wetland degradation is constructed. The overall model adopts a dual-temporal shared weight encoding + multi-branch decoding structure. The specific steps are as follows: Step 4.1: The model input consists of two consecutive remote sensing images of the same geographic area, denoted as the previous image. With post-production images Both are RGB three-channel orthophotos: B represents the batch size, and H and W are the spatial dimensions. The corresponding label image Y has pixel values ​​defined as follows: Y=1: forest and grassland wetland degradation change area; Y=0: forest and grassland wetland unchanged area; Y=255: non-forest and grassland wetland area.

[0030] Step 4.2: Constructing a shared weight encoder: To extract high-level semantic features with global modeling capabilities, ViT-L based on DINOv3 pre-trained weights is used as the backbone network to process previous images. With post-production images Feature encoding is performed separately for both images. The two images share the same set of network parameters during the input stage to ensure that the feature representations are in a unified semantic space, thereby guaranteeing physical consistency and semantic comparability in difference modeling. The feature extraction process is represented as follows: .

[0031] Specifically, the encoder takes a 512×512 image as input and uses a 16×16 patch partitioning method, dividing each image into 32×32 non-overlapping image blocks. Each patch is linearly projected onto a 1024-dimensional embedding space to form a sequential feature representation. The encoder contains a 24-layer Transformer encoder structure, with each layer configured with 16 multi-head self-attention heads to model long-distance spatial dependencies and enhance global contextual representation capabilities. To improve the stability and generalization ability of positional encoding at different scales, a rotation positional encoding rescaling mechanism is introduced, and four storage tokens are set to enhance global information aggregation and semantic memory capabilities. For feature output, intermediate features from layers 6, 12, 18, and 24 are selected as multi-scale semantic representations for subsequent difference modeling and degradation detection. The final output features satisfy: Where B is the batch size, C=1024, W′=H′=32.

[0032] Step 4.3, Pre-detection branch: To explicitly reflect the core business rule that "the target in the early detection stage must be forest, grassland, or wetland," a pre-detection branch is set up after the shared weight encoder to perform pixel-level binary classification probability modeling on the early features. This branch is used to output the probability that each pixel in the early image belongs to the forest, grassland, or wetland category, providing explicit constraint information for subsequent targeted degradation change detection.

[0033] Specifically, the input to the preliminary discrimination branch is the feature representation of the highest semantic level among the multi-scale features output by the backbone network, namely the output features of the 24th layer Transformer encoder. The initial discrimination branch employs a lightweight convolutional classifier structure, which includes: a 3×3 convolutional layer (reducing the number of channels from 1024 to 256), a batch normalization layer, and a ReLU activation function for local feature integration and channel compression; followed by a 1×1 convolutional layer to map the number of channels to a 1D output; finally, a Sigmoid activation function is used to output the probability value. Its calculation process can be represented as follows: , in, This represents the initial forest and grassland wet classification head composed of the above two convolutional layers. Using the Sigmoid activation function, the output probability map is... The value of each pixel location represents the confidence level that it belongs to the forest-grassland wetland area.

[0034] Step 4.4: To construct a representation of the changes between the early and late images, feature-level difference modeling is performed based on the output features of the shared weight encoder. Early features and later features All features are extracted from the intermediate features of the Transformer encoder layers 6, 12, 18, and 24 of the backbone network, and after scale alignment, they form a multi-scale fused feature of uniform size. Therefore, pixel-by-pixel difference modeling can be performed directly, and its expression is as follows: .in It represents the difference of the pixel in the high-dimensional semantic feature vector. Its magnitude and direction reflect the degree of comprehensive change of the position in terms of spectral distribution, texture structure and spatial context.

[0035] Step 4.5, Degradation Change Detection Branch: To achieve the goal of "detecting only the degradation changes of forests and grasslands in wetlands," this embodiment constructs a degradation change detection branch, whose input consists of two parts: one is a representation of the differences in features between the early and late stages. First, it is used to depict potential changes; second, it is used for early-stage forest and grassland wet probability maps. This is used to provide business constraint information. First, the probability map is extended along the channel dimension and then concatenated and fused with the differential features to obtain the fused features: in, After fusion, UPerHead is used to perform multi-scale pooling on high-level features to enhance global context awareness. Then, a top-down feature fusion path is used to progressively upsample and fuse the multi-scale features, improving the ability to identify degraded patches at different scales. Finally, a 1×1 convolution is applied to a single-channel output, and a pixel-level degradation change probability map is obtained after passing through a Sigmoid activation function. in: , which represents the probability value of the corresponding pixel undergoing a degradation change from "forest and grassland wet" to "non-forest and grassland wet".

[0036] By explicitly embedding the previous forest and grassland wet probability map into the change detection head input and combining it with the multi-scale contextual modeling capabilities of UPerHead, this embodiment implements a directional constraint mechanism at the network structure level that "whether a change is valid must be based on the premise that the previous forest and grassland were wet," thereby fundamentally improving the directionality and accuracy of degradation change detection.

[0037] Step 5: During the model training phase, a multi-task joint loss function is used for parameter optimization. By constraining the previous forest-grassland wetland discrimination task and the targeted degradation change detection task, it is ensured that the model only performs degradation change identification in the previous forest-grassland wetland area, thereby reducing false detections. The specific steps are as follows: Step 5.1: First, calculate the previous forest / grassland wetness discrimination loss, which is the probability that a pixel in the previous image belongs to the forest / grassland wetness category. Only in non-background areas ( The gradient contribution from the background region is ignored when participating in the loss calculation. Binary Cross Entropy (BCE) loss is used as a constraint, and its calculation formula is as follows: , in, This represents an indicator function, which takes the value 1 when the condition is met and 0 otherwise. This loss ensures that the model makes accurate probabilistic predictions of forest and grassland wetlands in the early stages, providing explicit constraints for subsequent degradation and change assessment.

[0038] Step 5.2: Output pixel-level degradation change probability from the degradation change detection branch. The supervision is only calculated for pixels in the forest, grassland, and wetland areas in the early stage (i.e., pixels with label Y=1), ignoring the gradient contribution of non-forest, grassland, and wetland areas or background areas, ensuring that the model does not misclassify non-business-related areas as changes. A binary classification cross-entropy loss is used, defined as follows: This loss constraint model only learns the degradation and change characteristics of forest, grassland and wetland areas in the early stage, and structurally realizes the business rule that "whether the change is valid must be based on the premise that the area was forest, grassland and wetland in the early stage".

[0039] Step 5.3: The final training loss is a weighted combination of the previous forest-grassland wet discrimination loss and degradation change detection loss. , in , The weighting coefficients are used to balance the importance of the initial land cover classification task and the degradation change classification task, and can be adjusted based on training experience or cross-validation. Through this joint loss function, the model simultaneously optimizes the initial forestry, grassland, and wetland identification capabilities and degradation change classification capabilities during training, achieving business rule-driven targeted change detection.

[0040] The early and late images of the area to be detected are input into the trained model. After forward inference, two outputs are obtained simultaneously: the early forest and grassland wet probability map. and probability diagram of degradation change Both are single-channel probability maps with the same spatial resolution as the input image, and their values ​​range from [0, 1]. Explicit logical constraints are imposed on the inference results to ensure that degradation changes only occur in areas that were previously forested / grassland wet. Specifically, a previous forested / grassland wet probability threshold τ (e.g., 0.5) is set for... Binarization is performed to generate a forest and grassland wet cover. = 1( >τ); Subsequently, the degradation change probability map is multiplied pixel by pixel with the mask to obtain the final degradation change probability map. = ⊙ This operation forces the degradation change response in non-forestry and grassland wetlands to zero, completely eliminating false detections in non-business-related areas from the output level. Binarization is performed (e.g., a threshold of 0.5) to obtain the final binary map of forest and grassland wetland degradation changes. Morphological filtering can be performed as needed to remove isolated noise points, and the results can be vectorized or exported as a GeoTIFF file with georeferenced data for subsequent business scenarios.

[0041] All content not described in detail in this specification belongs to the prior art known to those skilled in the art. Furthermore, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for designing an automatic sample construction and detection model for forest and grassland wetland changes, characterized in that, Includes the following steps: Step 1: Preparation and preprocessing of multi-source forest and grassland wetland monitoring data. The monitoring data includes early and late remote sensing images of the same area, vector data of the previous year's forest and grassland wetland resource base map, and vector data of multiple periods of forest and grassland wetland resource change monitoring results to ensure pixel matching between early and late images. Step 2: Set a fixed-size sliding window and traversal step size, and read the image from Step 1 to perform sliding window cropping; Step 3: Filter the images after sliding window cropping to form a dataset for training the change detection model; Step 4: Construct a change detection model: A dual-temporal shared weight encoding + multi-branch decoding structure is adopted. The encoder is constructed to realize feature extraction, the early discrimination branch outputs the probability value of each pixel belonging to the forest and grassland wetland range, and the degradation change detection branch outputs the prediction result of forest and grassland wetland degradation change. Step 5: Train and optimize the parameters of the change detection model based on the multi-task joint loss function.

2. The automatic sample construction and detection model design method for forest and grassland wetland change according to claim 1, characterized in that, Step 2 specifically involves: Image data of corresponding areas in the earlier and later remote sensing images were read respectively; The vector data of change detection results are used to generate an initial change label map containing only two types of labels: "change" and "unchanged" through spatial rasterization. The changed areas are assigned a value of 1, and the unchanged areas are assigned a value of 0. The base map vector data of forest and grassland wetland resources is rasterized to generate a forest and grassland wetland range mask, with a value of 1 within the range and a value of 255 outside the range; The initial change label map is merged with the forest and grassland wet range mask to generate the final sample label map. The rule is: within the forest and grassland wet range, the initial change label is retained as 0 or 1; outside the forest and grassland wet range, it is uniformly marked as the 255 background area.

3. The automatic sample construction and detection model design method for forest and grassland wetland change according to claim 1, characterized in that, Step 4 specifically involves: Encoder construction: A backbone network with a self-attention mechanism is selected as the feature encoder. A dual-branch structure with shared weights is adopted to extract features from early and late remote sensing images respectively. After encoding, a multi-channel feature map is output. The initial discrimination branch is constructed by taking the encoded features of the previous remote sensing image as input and consisting of multiple convolution operations. The output is a single-channel probability map with the same spatial resolution as the input features, which is used to represent the probability value of each pixel belonging to the forest, grassland and wetland range. The degradation change detection branch is constructed, and pixel-by-pixel absolute difference operation is performed on the features of the early and late images to obtain the feature difference map between the early and late images; The feature difference map is concatenated with the probability map output by the previous discrimination branch in the channel dimension to form a fused feature. Based on the fused feature, a degradation change detection branch is constructed to output the prediction results of forest and grassland degradation changes.

4. The automatic sample construction and detection model design method for forest and grassland wetland change according to claim 3, characterized in that, The backbone network of the encoder is ViT-L based on DINOv3 pre-trained weights. The encoder contains a 24-layer Transformer encoder structure, with 16 multi-head self-attention heads configured in each layer. A rotation position encoding rescaling mechanism is introduced, and 4 storage tokens are set. In terms of feature output, the intermediate features of the 6th, 12th, 18th and 24th layers are selected as multi-scale semantic representations.

5. The automatic sample construction and detection model design method for forest and grassland wetland change according to claim 3, characterized in that, The initial discrimination branch structure includes a 3×3 convolution layer, a batch normalization layer, and a ReLU activation function for local feature integration and channel compression, followed by a 1×1 convolution layer.

6. The automatic sample construction and detection model design method for forest and grassland wetland change according to claim 3, characterized in that, After fusion, the degradation change detection branch uses UPerHead to perform multi-scale pooling on high-level features. Then, through a top-down feature fusion path, the multi-scale features are upsampled and fused step by step. Finally, the 1×1 convolution is mapped to a single-channel output, and the pixel-level degradation change probability map is obtained through the Sigmoid activation function.

7. The automatic sample construction and detection model design method for forest and grassland wetland change according to claim 1, characterized in that, Step 5 first calculates the early forest and grassland wetland discrimination loss, using binary cross-entropy loss, supervised only in non-background areas, to constrain the model to correctly identify the early forest and grassland wetland range; and the degradation change detection loss, using binary cross-entropy loss, supervised only on real change pixels, to ensure that the model learns degradation change features only in forest and grassland wetland areas; the total loss is a weighted combination of the early forest and grassland wetland discrimination loss and the degradation change detection loss.

8. The automatic sample construction and detection model design method for forest and grassland wetland change according to claim 1, characterized in that, The dataset is divided into training set: validation set: test set in a ratio of 4:1:1.

Images