Three-branch semantic segmentation method for ski resort land cover classification

Abstract

The invention relates to a three-branch semantic segmentation method for ski resort land cover classification, and relates to the technical field of computer vision and remote sensing intelligent analysis, and the method comprises the steps: obtaining a typical ski resort high-resolution image; building an overall framework, a backbone module, a neck feature fusion module and a multi-task segmentation head of the three-branch semantic segmentation network GDBNet; and carrying out model training on the GDBNet by adopting the ski resort surface coverage classification data set for subsequent surface coverage classification. According to the method, a very simple module design is adopted, and a CNN and Transform coupled three-branch network GDBNet is constructed to model global, local and edge context information. The model provided by the invention reaches the SOTA level, and the overall F1 value is higher than 90%, which is superior to other advanced algorithms. Compared with a baseline model, the method has the advantage that the F1 index of ground feature classification in the skiing resort area is improved by 5.78% to 90.4%.

Description

Technical Field

[0001] This invention relates to the field of computer vision and remote sensing intelligent analysis technology, and in particular to a three-branch semantic segmentation method for land cover classification in ski resorts. Background Technology

[0002] Timely and accurate feature classification in ski resorts will benefit various aspects of ski resort operation and management, mountain disaster prevention, and natural resource planning. Currently, many feature classification results exist, including medium-resolution and high-resolution versions. However, ski resorts are typically small in area, with fine-scale features and unique features such as ski runs. Conventional medium-resolution feature classification results are coarse-grained and cannot accurately describe the features of ski resorts. The latest meter-scale feature classification results lack a ski run category and have low overall accuracy, making them poorly suited for the fine-scale features of ski resorts. Therefore, it is necessary to develop refined feature classification mapping specifically for the unique scenario of ski resorts.

[0003] However, current land cover classification in ski resorts faces several key technical bottlenecks: First, the problem of spectral feature confusion leads to a severe lack of inter-class discrimination—snow, ice, and clouds have highly similar reflectance in the visible light band (all >85%), while compacted hardened ski slopes and bare rock surfaces exhibit nearly identical absorption characteristics in the near-infrared band, resulting in a high misclassification rate for traditional classification methods in such scenarios. Second, the complexity of multi-scale and multi-shaped targets restricts classification integrity—small linear elements such as snow slopes and roads are interspersed among large areal elements such as vegetation and building areas, and the large differences in features make it difficult for models to cope. In addition, the lack of adaptability to seasonal and climatic dynamic environments leads to a decline in the model's generalization ability—the seasonal differences of different elements and the strong light variations in high-altitude areas (such as the shadow areas on steep slopes and the strong reflective areas of snow surfaces) cause the existing models to have unstable results on cross-phenological and cross-temporal data. Summary of the Invention

[0004] This invention addresses the technical problems of high misclassification rates, complex multi-scale and multi-shaped targets, and decreased model generalization ability caused by the lack of adaptability to seasonal and climatic dynamic environments in existing ski resort land cover classification methods. It provides a three-branch semantic segmentation method for ski resort land cover classification.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A three-branch semantic segmentation method for land cover classification in ski resorts includes the following steps:

[0007] Step S1: Obtain high-resolution images of typical ski resorts, annotate them, and construct a ski resort land cover classification dataset;

[0008] Step S2: Construct the overall framework of the three-branch semantic segmentation network GDBNet;

[0009] Step S3: Construct the backbone module of the three-branch semantic segmentation network GDBNet;

[0010] Step S4: Construct the neck feature fusion module of the three-branch semantic segmentation network GDBNet;

[0011] Step S5: Construct the multi-task segmentation head of the three-branch semantic segmentation network;

[0012] Step S6: Train the GDBNet model using the ski resort land cover classification dataset for subsequent land cover classification.

[0013] In the above technical solution, in step S1, the land cover types in the ski resort land cover classification dataset include: roads, water bodies, vegetation, farmland, ski slopes, and building areas.

[0014] In the above technical solution, in step S2, the overall framework adopts the general Backbone-Neck-Head framework, wherein:

[0015] The backbone section employs a three-branch design to extract global, detailed, and boundary features;

[0016] Global features are encoded using Vision Transformer, while detailed and boundary features are encoded using lightweight ResNet18.

[0017] The Neck section employs a pixel-level attention fusion module and a boundary-guided attention fusion module to fuse features from different branches; the Neck section has a hierarchical structure.

[0018] The Head section employs a multi-task supervised design, supervising global interactions, local interactions, and edge extraction to assist each branch in effectively learning different task representations.

[0019] In the above technical solution, in step S3, the backbone module includes the following in the Backbone section: Global branch, Detail branch, and Boundary branch;

[0020] The Global branch is coded using the standard Vision Transformer;

[0021] Both the Detail and Boundary branches are encoded using lightweight ResNet18 and output hierarchical multi-scale features from four stacked residual blocks in the backbone.

[0022] In the above technical solution, in step S3, for an image x, the backbone module is calculated as follows:

[0023]

[0024] in, and These represent the hierarchical multi-scale feature maps output by the i-th module in the Global, Detail, and Boundary branches, respectively, where i ∈ [0,3]. ViT(x) represents the encoding of image x using the Vision Transformer. ResNet18 D (x) and ResNet18 B (x) represents the encoding of image x using ResNet18.

[0025] In the above technical solution, step S4 specifically includes:

[0026] Step S41: Build the first neck feature fusion module, PG-AFM module;

[0027] First, the size and number of channels of the hierarchical multi-scale feature maps output by each branch of the backbone module are adjusted. A 1×1 convolution and interpolation module is used to process the hierarchical multi-scale feature maps output by the Global branch. Hierarchical multi-scale feature maps output by the Detail branch Adjustments were made to obtain feature maps with the same feature map size and number of channels. and

[0028] Next, an attention mechanism is used to fuse the features of the two branches, the specific calculation of which is as follows:

[0029]

[0030] Finally, the multi-scale fusion features of PGL-FM are obtained. This represents the pixel activation value; Sigmoid() represents the sigmoid calculation.

[0031] Step S42: Construct the second neck feature fusion module, the BG-AFM module;

[0032] The Global and Detail branches are fused using Boundary branch features; the hierarchical multi-scale feature map output from the Boundary branch is then processed. Adjusting the size and channels to output features To match the features output by the other two branches;

[0033] A gating mechanism is employed to make the model rely more on the Detail branch in boundary regions and more on the semantic information of the Global branch in other regions. The specific calculation is as follows:

[0034]

[0035] In the formula, It is the probability distribution of the edges of ground features in the image;

[0036] Using the probability distribution of ground feature edges in the image as a gating mechanism, the Global and Detail branches are selected and aggregated to output the multi-scale fusion features of BG-AFM.

[0037] In the above technical solution, step S5 specifically includes:

[0038] The BG-AFM module, which fuses features from all three branches, will serve as the primary output task. An UpperNet semantic segmentation head will be added after the BG-AFM module to optimize the entire network; its loss will be denoted as l0. Similarly, an FCN semantic segmentation head will be added after the PG-AFM module to assist in optimizing the Global and Detail branches; its loss will be denoted as l1. All semantic losses will be calculated using cross-entropy loss, as shown in the following formula:

[0039] l0=-∑y×log y gdb

[0040] l1=-∑y×log y gd

[0041] Where y represents the actual label of land cover classification, y gdb and y gd These represent the outputs of the segmentation heads that connect to the BG-AFM module and the PG-AFM module, respectively.

[0042] The calculation is performed using edge-weighted cross-entropy loss, and the formula is as follows:

[0043] l2=-∑w pixel ×y edge ×log y b

[0044] Among them, y edge This represents the true edge labeling of land cover classification; y b Represents the edge segmentation head output; w pixel L2 represents pixel weights; L2 represents edge loss.

[0045] Ultimately, GDBNet's loss is:

[0046] LOSS = a0×l0 + a1×l1 + a2×l2

[0047] Where a0, a1, and a2 represent the loss weight parameters of l0, l1, and l2, respectively.

[0048] In the above technical solution, in step S6, after the model training is completed, the accuracy of the model is evaluated using a validation sample set, specifically as follows:

[0049] The F1 score is used for model evaluation. TP: positive samples predicted as positive by the model, FP: negative samples predicted as positive by the model, and FN: positive samples predicted as negative by the model. The F1 score is calculated as follows:

[0050]

[0051] Here, precision represents accuracy, and recall represents recall.

[0052] The present invention has the following beneficial effects:

[0053] This invention presents a three-branch semantic segmentation method for land cover classification in ski resorts. It employs a minimalist modular design, constructing a three-branch network GDBNet coupled with CNN and Transformer to model global, local, and edge contextual information. Specifically, the neck region utilizes pixel-guided and edge-guided feature aggregation modules to fuse multi-scale feature representations from different branches. The head region employs a multi-task approach combining semantic and edge supervision to enhance the model's ability to identify fine features. In dataset evaluation, the proposed model achieves state-of-the-art (SOTA) performance, with an overall F1 score exceeding 90%, outperforming other advanced algorithms. Compared to the baseline model, this invention improves the F1 score for ski resort land cover classification by 5.78% to 90.4%. Attached Figure Description

[0054] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0055] Figure 1 This is a flowchart illustrating the three-branch semantic segmentation method for land cover classification in ski resorts according to the present invention.

[0056] Figure 2 This is a schematic diagram of a ski resort.

[0057] Figure 3 This is a schematic diagram of the overall architecture of GDBNet.

[0058] Figure 4This is a schematic diagram of the PG-AFM module.

[0059] Figure 5 This is a schematic diagram of the BG-AFM module. Detailed Implementation

[0060] The inventive concept of this invention is as follows:

[0061] This invention presents a three-branch semantic segmentation method for land cover classification in ski resorts, primarily applied to the digital and refined management of ski areas. Through intelligent analysis of high-resolution remote sensing images, it achieves comprehensive classification of all elements, including ski slopes, vegetation-covered areas, bare rock zones, and man-made structures. In terms of dynamic monitoring of snow and ice resources, this invention supports quantitative assessment of snow cover, analysis of snow quality degradation, and avalanche risk early warning. Simultaneously, it provides core data support for assessing the ecological carrying capacity of resorts and for ski slope planning and maintenance, serving industrial applications such as ski resort operation and management, mountain disaster prevention, and cold-region environmental protection.

[0062] The present invention will now be described in detail with reference to the accompanying drawings.

[0063] The three-branch semantic segmentation method for land cover classification in ski resorts, as described in this invention, relates to the field of remote sensing technology, such as... Figure 1 As shown, it includes the following steps:

[0064] Step S1: Obtain high-resolution images of typical ski resorts, annotate them, and construct a ski resort land cover classification dataset;

[0065] Step S2: Construct the overall framework of the three-branch semantic segmentation network GDBNet;

[0066] Step S3: Construct the backbone module of the three-branch semantic segmentation network GDBNet;

[0067] Step S4: Construct the neck feature fusion module of the three-branch semantic segmentation network GDBNet;

[0068] Step S5: Construct the multi-task segmentation head of the three-branch semantic segmentation network;

[0069] Step S6: Train the GDBNet model using the ski resort land cover classification dataset for subsequent land cover classification.

[0070] The following section details the three-branch semantic segmentation method of the present invention for land cover classification in ski resorts, using specific examples.

[0071] Step S1: Obtain high-resolution images of typical ski resorts, annotate them, and construct a ski resort land cover classification dataset. Specifically:

[0072] In this example, the present invention uses seven major ski resorts in Jilin Province as the target area and collects high-resolution optical images using the Jilin-1 satellite constellation as the data source. These images have a resolution of approximately 0.75 meters. Considering that seasonal effects can significantly impact the representation of different land features, the present invention collects a total of eight high-resolution remote sensing images for four quarters in 2023 and 2024, based on the boundary range of each ski resort. The images have three channels: red, green, and blue. Example images of the target area and ski resorts are shown below. Figure 2 As shown.

[0073] After image collection, image annotation was performed. Based on typical landform features of ski resorts, six common land cover types were defined, including roads, water bodies, vegetation, farmland, ski slopes, and built-up areas. Annotators used polygon features to label all objects in the six categories (excluding background). After the first round of annotation, self-checking and cross-checking were conducted to correct incorrect labels, missing objects, and inaccurate boundaries. Using image-label pairs, 10,000 samples were finally generated using grid cropping, with a pixel size of 512*512. These samples were randomly allocated to the training and test sets in an 8:2 ratio, with 8,000 images in the training set and 2,000 images in the test set.

[0074] Step S2: Construct the overall framework of the three-branch semantic segmentation network GDBNet. Specifically:

[0075] To address the challenges of land cover classification in high-resolution remote sensing imagery, this invention proposes a novel three-branch semantic segmentation model, GDBNet, which combines the advantages of CNN and Transformer. The overall framework adopts a general Backbone-Neck-Head framework to adapt to the modular design of current research methods, such as... Figure 3 As shown, the Backbone section employs a three-branch design to extract global, detailed, and boundary features. Global features are encoded using Vision Transformer, while detailed and boundary features are encoded using lightweight ResNet18 to avoid significantly increasing parameters. The Neck section proposes a pixel-level attention fusion module and a boundary-guided attention fusion module to fuse features from different branches. To adapt to dense prediction tasks like semantic segmentation, the Neck section has a hierarchical structure to maintain the same feature scale and spatial resolution, ensuring consistent representations across different feature levels. The Head section employs a multi-task supervised design, supervising global interactions, local interactions, and edge extraction to assist each branch in effectively learning different task representations.

[0076] Step S3: Build the backbone module of the three-branch semantic segmentation network GDBNet. Specifically:

[0077] like Figure 3 As shown, a three-branch design is adopted in the backbone part. The Global branch is encoded using the standard VisionTransformer. In this invention, the ViT-B model of DINOv2 is used for weight initialization. To ensure hierarchical multi-scale feature output, layers 3, 6, 9, and 12 of ViT-B, totaling four levels of features, are selected to output hierarchical multi-scale features. Both the Detail and Boundary branches are encoded using a lightweight ResNet18, and hierarchical multi-scale features are output from four stacked residual blocks in the backbone model. For an image x, the backbone modules of the three-branch skeleton are calculated as follows:

[0078]

[0079] in, and These represent the hierarchical multi-scale feature maps output by the i-th module in the Global, Detail, and Boundary branches, respectively, where i ∈ [0,3]. ViT(x) represents the encoding of image x using the Vision Transformer. ResNet18 D (x) and ResNet18 B (x) represents the encoding of image x using ResNet18.

[0080] Step S4: Construct the neck feature fusion module of the three-branch semantic segmentation network GDBNet; specifically including:

[0081] Step S41: Build the first neck feature fusion module, PG-AFM module;

[0082] Feature fusion can enhance the representational power of a model, effectively improving its representational capabilities. In GDBNet, the Global branch Transformer structure can extract rich global semantic information, while the Detail branch CNN structure can provide fine-grained local semantic information. Therefore, effectively fusing the information from both is crucial. This invention proposes the PGL-FM module, whose structure is as follows: Figure 4 As shown. First, the size and number of channels of the hierarchical multi-scale feature maps output by each branch of the backbone module are adjusted, that is, a 1×1 convolution and interpolation module is used to adjust the hierarchical multi-scale feature maps output by the Global branch. Hierarchical multi-scale feature maps output by the Detail branch Adjustments were made to obtain features with the same feature map size and number of channels, i.e. and Next, an attention mechanism is used to fuse the two-branch features, the specific calculation of which is as follows:

[0083]

[0084] in, This represents the pixel activation value; Sigmoid() represents the sigmoid calculation.

[0085] The core concept of this design stems from the gating principle: for corresponding pixels originating from the Global and Detail branches, the Sigmoid function is used to calculate the product of the two. This can be considered as a probability distribution that simultaneously identifies pixels as foreground objects across two branches, acting as a gating switch to select between the Global and Detail branches. If... The higher the level, the more preferred the Global branch, because the Global branch uses the ViT architecture, resulting in richer and more precise object semantics; if The lower the value, the more favor the Detail branch is chosen, because inconsistencies are more likely to occur in detailed areas such as object edges, where CNN branches are more advantageous. Therefore, the final result is the multi-scale fusion feature obtained from PGL-FM.

[0086] Step S42: Construct the second neck feature fusion module, the BG-AFM module;

[0087] Edge features are crucial for remote sensing land cover classification; therefore, this invention designs a boundary-guided attention feature fusion module to assist in the fusion of global and local contextual features. Specifically, boundary branch features are used to fuse the Global and Detail branches, as detailed in the following design: Figure 5 As shown. Similarly, this module first processes the hierarchical multi-scale feature map output by the Boundary branch. Adjusting the size and channels to output features This is used to match the features output by the other two branches. Considering that the Global branch is accurate in contextual semantics but loses too much in spatial and geometric details, especially in boundary regions and small objects, while the Detail branch better preserves spatial details, this invention also employs a gating mechanism. This mechanism makes the model rely more on the Detail branch in boundary regions and more on the semantic information of the Global branch in other regions. The specific calculation is as follows:

[0088]

[0089] In the formula, This can be considered as the probability distribution of ground feature edges in an image. The larger the value, the more likely the pixel region is to be an edge detail region; the smaller the pixel region, the more likely it is to be other regions. This value is used as a gating mechanism to select and aggregate the Global and Detail branches, outputting the multi-scale fusion features of BG-AFM.

[0090] Step S5: Construct the multi-task segmentation head of the three-branch semantic segmentation network, specifically:

[0091] To meet different task requirements, this invention designs three segmentation heads to drive different branches to learn their respective features. The BG-AFM module, which fuses the features of the three branches, will serve as the main output task. This invention adds an UpperNet semantic segmentation head after the BG-AFM module to optimize the entire network, with a loss denoted as l0. Simultaneously, a simple FCN semantic segmentation head is added after the PG-AFM module to assist in optimizing the Global and Detail branches, with a loss denoted as l1. All semantic losses utilize cross-entropy loss, calculated as follows:

[0092] l0=-∑y×log y gdb

[0093] l1=-∑y×log y gd

[0094] Where y represents the actual label of land cover classification, y gdb and y gd These represent the outputs of the segmentation heads that connect to the BG-AFM module and the PG-AFM module, respectively.

[0095] In addition, to improve the accuracy of edge information extraction by the boundary-aware branch, an edge-weighted cross-entropy loss is used for calculation, and the calculation formula is as follows:

[0096] l2=-∑w pixel ×y edge ×log y b

[0097] Among them, y edge This represents the true edge labels of land cover classification, calculated using the Canny operator. b This represents the output of the edge segmentation head. pixel 1 represents pixel weights, which are set inversely proportional to the number of edge and non-edge pixels in the labeled image; l2 represents edge loss.

[0098] Ultimately, GDBNet's loss is:

[0099] LOSS = a0×l0 + a1×l1 + a2×l2

[0100] Based on experience, a0, a1, and a2 represent the loss weight parameters of l0, l1, and l2, respectively. In this example, the loss weight parameters are set to a0 = 1.0, a1 = 1.0, and a2 = 20.0.

[0101] Step S6: Train the GDBNet model using the ski resort land cover classification dataset for subsequent land cover classification and conduct accuracy evaluation; specifically including:

[0102] Step S61: Model Training. Training and validation were performed using the PyTorch platform on an NVIDIA GeForce RTX 3090 GPU. This example sets up a total of 80,000 training iterations with a batch size of 4 and no data augmentation. The Adam optimizer was used with an initial learning rate of 6e. -5 The weights are decayed to 0.01 to perform iterative training.

[0103] Step S62: After model training is complete, the model's accuracy is evaluated using a validation sample set. This use case uses the F1 score for model evaluation. First, three concepts are defined: TP: positive samples predicted as positive by the model; FP: negative samples predicted as positive by the model; FN: positive samples predicted as negative by the model. The F1 score is calculated as follows:

[0104]

[0105] Here, precision represents the actual correctness of the model's predictions among positive samples, and recall represents the percentage of positive samples in the true set that were correctly predicted by the model. The F1 score is obtained from precision and recall, and the F1 score is generally between [0,1].

[0106] This invention presents a three-branch semantic segmentation method for land cover classification in ski resorts. It employs a minimalist modular design, constructing a three-branch network GDBNet coupled with CNN and Transformer to model global, local, and edge contextual information. Specifically, the neck region utilizes pixel-guided and edge-guided feature aggregation modules to fuse multi-scale feature representations from different branches. The head region employs a multi-task approach combining semantic and edge supervision to enhance the model's ability to identify fine features. In dataset evaluation, the proposed model achieves state-of-the-art (SOTA) performance, with an overall F1 score exceeding 90%, outperforming other advanced algorithms. Compared to the baseline model, this invention improves the F1 score for ski resort land cover classification by 5.78% to 90.4%.

[0107] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A three-branch semantic segmentation method for land cover classification in ski resorts, characterized in that, Includes the following steps: Step S1: Obtain high-resolution images of typical ski resorts, annotate them, and construct a ski resort land cover classification dataset; Step S2: Construct the overall framework of the three-branch semantic segmentation network GDBNet; Step S3: Construct the backbone module of the three-branch semantic segmentation network GDBNet; Step S4: Construct the neck feature fusion module of the three-branch semantic segmentation network GDBNet; Step S5: Construct the multi-task segmentation head of the three-branch semantic segmentation network; Step S6: Train the GDBNet model using the ski resort land cover classification dataset for subsequent land cover classification.

2. The three-branch semantic segmentation method for land cover classification in ski resorts according to claim 1, characterized in that, In step S1, the land cover classification dataset for ski resorts includes the following land cover types: roads, water bodies, vegetation, farmland, ski slopes, and built-up areas.

3. The three-branch semantic segmentation method for land cover classification in ski resorts according to claim 1, characterized in that, In step S2, the overall framework adopts the general Backbone-Neck-Head framework, wherein: The backbone section employs a three-branch design to extract global, detailed, and boundary features; Global features are encoded using Vision Transformer, while detailed and boundary features are encoded using lightweight ResNet18. The Neck section employs a pixel-level attention fusion module and a boundary-guided attention fusion module to fuse features from different branches; the Neck section has a hierarchical structure. The Head section employs a multi-task supervised design, supervising global interactions, local interactions, and edge extraction to assist each branch in effectively learning different task representations.

4. The three-branch semantic segmentation method for land cover classification in ski resorts according to claim 3, characterized in that, In step S3, the backbone module includes the following branches in the Backbone section: Global branch, Detail branch, and Boundary branch; The Global branch is coded using the standard Vision Transformer; Both the Detail and Boundary branches are encoded using lightweight ResNet18 and output hierarchical multi-scale features from four stacked residual blocks in the backbone.

5. The three-branch semantic segmentation method for land cover classification in ski resorts according to claim 4, characterized in that, In step S3, for an image x, the backbone module is calculated as follows: in, and These represent the hierarchical multi-scale feature maps output by the i-th module in the Global, Detail, and Boundary branches, respectively, where i ∈ [0,3]. ViT(x) represents the encoding of image x using the Vision Transformer. ResNet18 D (x) and ResNet18 B (x) represents the encoding of image x using ResNet18.

6. The three-branch semantic segmentation method for land cover classification in ski resorts according to claim 5, characterized in that, Step S4 specifically includes: Step S41: Build the first neck feature fusion module, PG-AFM module; First, the size and number of channels of the hierarchical multi-scale feature maps output by each branch of the backbone module are adjusted. A 1×1 convolution and interpolation module is used to process the hierarchical multi-scale feature maps output by the Global branch. Hierarchical multi-scale feature maps output by the Detail branch Adjustments were made to obtain feature maps with the same feature map size and number of channels. and Next, an attention mechanism is used to fuse the features of the two branches, the specific calculation of which is as follows: Finally, the multi-scale fusion features of PGL-FM are obtained. This represents the pixel activation value; Sigmoid() represents the sigmoid calculation. Step S42: Construct the second neck feature fusion module, the BG-AFM module; The Global and Detail branches are fused using Boundary branch features; the hierarchical multi-scale feature map output from the Boundary branch is then processed. Adjusting the size and channels to output features To match the features output by the other two branches; A gating mechanism is employed to make the model rely more on the Detail branch in boundary regions and more on the semantic information of the Global branch in other regions. The specific calculation is as follows: In the formula, It is the probability distribution of the edges of ground features in the image; Using the probability distribution of ground feature edges in the image as a gating mechanism, the Global and Detail branches are selected and aggregated to output the multi-scale fusion features of BG-AFM.

7. The three-branch semantic segmentation method for land cover classification in ski resorts according to claim 6, characterized in that, Step S5 is as follows: The BG-AFM fusion of three-branch features will be used as the main output task. An UpperNet semantic segmentation head will be added after the BG-AFM module to optimize the entire network. Its loss is denoted as l0. An FCN semantic segmentation header is added after the PG-AFM module to assist in optimizing the Global and Detail branches, and its loss is denoted as l1; the semantic loss is all based on cross-entropy loss, and the calculation formula is as follows: l0=-∑y×logy gdb l1-∑y×logy gd Where y represents the actual label of land cover classification, y gdb and y gd These represent the outputs of the segmentation heads that connect to the BG-AFM module and the PG-AFM module, respectively. The calculation is performed using edge-weighted cross-entropy loss, and the formula is as follows: l2=-∑w pixel xy edge xlogy b Among them, y edge This represents the true edge labeling of land cover classification; y b Represents the edge segmentation head output; w pixel L2 represents pixel weights; L2 represents edge loss. Ultimately, GDBNet's loss is: LOSS = a0×l0 + a1×l1 + a2×l2 Where a0, a1, and a2 represent the loss weight parameters of l0, l1, and l2, respectively.

8. The three-branch semantic segmentation method for land cover classification in ski resorts according to claim 1, characterized in that, In step S6, after the model training is complete, the accuracy of the model is evaluated using the validation sample set, specifically as follows: The F1 score is used for model evaluation. TP: positive samples predicted as positive by the model, FP: negative samples predicted as positive by the model, and FN: positive samples predicted as negative by the model. The F1 score is calculated as follows: Here, precision represents accuracy, and recall represents recall.

Images