Landslide detection method based on adaptive fusion of optical satellite remote sensing and DEM data

Abstract

The application discloses a landslide detection method based on adaptive fusion of optical satellite remote sensing and DEM data, and takes optical satellite remote sensing data and DEM data as core inputs to construct a heterogeneous double-branch high-resolution network: the optical branch adopts HRNet-W48 to extract spectral texture features, and the DEM branch adopts a lightweight HRNet-W18 to extract terrain geometric features. A terrain perception gating fusion module is designed, a pixel-level weight mask is dynamically generated based on the terrain features reflected by the DEM data through the gating network, and the adaptive weighted fusion of optical and terrain features is realized in the form of residual error; a CE+BCL combined loss function is proposed, and the boundary contrast loss is used to strengthen the supervision of the landslide edge area. The application effectively solves the problems of rigid fusion strategy, loss of spatial details and fuzzy boundary in the traditional method, and significantly improves the accuracy, boundary clarity and complex scene adaptability of landslide recognition.

Description

Technical Field

[0001] This invention relates to the field of landslide detection technology, specifically a landslide detection method based on adaptive fusion of optical satellite remote sensing and DEM data. Background Technology

[0002] Landslides are a common and highly destructive geological hazard, and rapid and accurate identification of landslide extent is crucial for disaster assessment, emergency rescue, and risk prevention. With the development of remote sensing technology, automated landslide detection methods based on optical imagery and digital elevation models have become a research hotspot. However, existing technologies still face a series of challenges in practical applications. The mainstream technologies in the current landslide detection field can be divided into four main categories, and the core solutions and characteristics of each type are as follows: (1) Landslide detection method based on single optical image These methods use RGB optical remote sensing imagery as the sole input data source and employ convolutional neural networks (CNNs) or visual Transformers as the backbone network to extract surface texture features, achieving semantic segmentation of landslide areas. For example, some techniques use U-Net, ResNet, or Swin-Transformer as the core architecture, capturing visual information such as texture and color of landslides through multi-scale feature extraction; other techniques enhance landslide boundary features by introducing edge detection modules, but they still essentially rely on single optical data. The core logic of these methods is to automatically learn the visual differences between landslides and the background in optical imagery using neural networks, without requiring additional topographic data.

[0003] (2) Landslide detection method based on single DEM data These methods rely solely on Digital Elevation Model (DEM) data, calculating terrain-derived features such as slope, aspect, and curvature, and combining them with traditional machine learning (e.g., support vector machines, random forests) or simple CNNs for landslide identification. For example, some techniques generate terrain slope maps from DEM data, using slope thresholds as the core criterion for landslide judgment; others utilize terrain curvature features extracted from DEMs to assist in identifying areas of abrupt terrain changes. The core idea of ​​these methods is to use terrain geometric features to reflect the topographic undulation characteristics of landslides, without relying on optical texture information.

[0004] (3) Landslide detection method based on fusion of optical and DEM data This type of method incorporates both optical imagery and DEM data, combining the advantages of both types of data through a simple fusion strategy. Common fusion methods include: first, channel stitching, where DEM data is stitched together with RGB imagery as an additional channel and then input into a single network; second, fixed-ratio weighted fusion, where features of the two types of data are linearly superimposed by manually setting weights; and third, early feature fusion, where optical features and terrain features are directly merged in the shallow layers of the network. For example, CN117274823B discloses a visual Transformer method based on DEM feature enhancement, which generates a mask matrix by calculating the roughness of DEM sub-regions, and then inputs it into the network after superimposing it with optical image patches; CN119417753A merges RGB images and DEM images into a four-channel image through a data fusion module, and then inputs it into a convolutional neural network to extract features; CN119380187A discloses an improved U-Net network, which adds an attention mechanism module after the skip connections in the last layer of the decoder, and fuses optical remote sensing features and terrain features derived from DEM; CN120997594A introduces a visual language model, enhances semantic understanding through text prompts, and fuses RGB and DEM features by combining a cross-attention mechanism.

[0005] Despite the progress made in the field of landslide detection, the following core shortcomings still urgently need to be addressed: (1) Insufficient feature representation from a single data source: Methods based on a single optical image are easily affected by factors such as changes in illumination, vegetation cover, and “different objects with the same spectrum” (e.g., terraces and landslides have similar spectra), making it difficult to capture the core geometric feature of landslides, such as topographic undulations; methods based on a single DEM data lack surface texture information and cannot distinguish landslides from other topographic undulation areas (e.g., natural steep slopes and artificially excavated slopes), resulting in a high false detection rate.

[0006] (2) Lack of adaptability in multi-source fusion strategies: Traditional multi-source fusion methods, such as channel splicing and fixed-ratio weighting, do not consider the spatial heterogeneity of terrain complexity. In areas with complex terrain (such as abrupt slope changes and landslide walls), DEM features are not fully utilized, while in flat areas, DEM noise introduces interference. Even some end-to-end fusion methods focus their attention mechanisms on the weight allocation of feature channels or spatial locations, without taking "terrain complexity" as the basis for dynamic adjustment of the fusion ratio, and thus cannot adapt to the feature requirements of different terrain scenarios.

[0007] (3) Boundary positioning ambiguity problem: Existing landslide identification schemes mostly use general classification loss functions and lack a special supervision mechanism for the geometric characteristics of landslide edges, resulting in severe "sawtooth" or positioning offset in the identification results at the landslide boundary.

[0008] Therefore, there is an urgent need for an automated segmentation method that can adaptively fuse multi-source features, maintain high-resolution details, and accurately locate landslide boundaries. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a landslide detection method based on adaptive fusion of optical satellite remote sensing and DEM data. This method achieves differentiated depth extraction of optical and terrain features through a heterogeneous dual-branch network, realizes dynamic spatial adaptation of feature fusion through a terrain-aware gating fusion mechanism, improves the ability to capture subtle landslides by maintaining high-resolution representation throughout the process, and significantly improves the positioning accuracy of landslide boundaries by introducing a boundary contrast loss function.

[0010] To address the aforementioned technical problems, embodiments of the present invention provide the following technical solution: a landslide detection method based on adaptive fusion of optical satellite remote sensing and DEM data, comprising the following steps: S1. Acquire RGB image data and corresponding DEM terrain data of the target area, and perform standardized preprocessing such as format conversion, size unification and normalization; S2. The preprocessed RGB image data and DEM topographic data are respectively input into a dual-branch high-resolution feature extraction network for processing, and output a multi-scale fusion feature map; wherein, the first branch is an optical feature extraction branch, which processes the RGB image data and outputs a first multi-scale fusion feature map; the second branch is a topographic feature extraction branch, which processes the DEM topographic data and outputs a second multi-scale fusion feature map; S3. Input the first multi-scale fused feature map and the second multi-scale fused feature map into the terrain-aware gating fusion module; the terrain-aware gating fusion module dynamically generates a weight mask corresponding to the spatial size of the first multi-scale fused feature map based on the terrain complexity information reflected by the second multi-scale fused feature map; using the weight mask, the first multi-scale fused feature map and the second multi-scale fused feature map after channel alignment are weighted and fused to generate a fused feature map; S4. Input the fused feature map into the segmentation head network. After feature enhancement, category prediction and spatial upsampling, output a landslide segmentation prediction map with the same size as the input image.

[0011] Furthermore, it also includes using a combined loss function to supervise and optimize the network parameters during the entire landslide detection model training phase. The combined loss function is:

[0012] For the combined loss function, For cross-entropy loss with class weights, For boundary contrast loss, This is the balance coefficient; Among them, the cross-entropy loss with class weights The formula is:

[0013] in, , The height and width of the feature map; For cross-entropy loss, For the model to pixels The probability of predicting it as a landslide. For pixels The true category labels, among which , For pixel coordinate index, For boundary adaptive weights, For valid pixel indication functions, when pixel The label value is set to 1 if it is not an ignored value, otherwise it is set to 0; boundary adaptive weights. The calculation formula is:

[0014] in, For boundary mask, These are the boundary weight coefficients; The formula is:

[0015] in, The spatial dimensions of the feature map; This refers to the number of effective pixels. For valid pixel indicator functions, when pixel The value is 1 if the label value is not an ignored value, and 0 otherwise. For pixels The true category label, 0 represents background, 1 represents landslide; For the model to pixels Predicted as category The probability of; This is the category weight vector.

[0016] Furthermore, in step S2, the optical feature extraction branch adopts the HRNet-W48 network structure, and the terrain feature extraction branch adopts the HRNet-W18 network structure; both branches use two-dimensional batch normalization layers for layer normalization, enhance nonlinearity through the ReLU activation function, and achieve multi-scale feature interaction through the cross-branch fusion mechanism in the high-resolution module.

[0017] Furthermore, in step S3, the terrain-aware gating fusion module consists of three parts: a DEM feature channel alignment unit, a terrain complexity gating subnet, and a dynamic fusion unit, and performs the following operations: S31, the DEM feature channel alignment unit adjusts the number of channels of the second multi-scale fusion feature map to be the same as that of the first multi-scale fusion feature map through a 1×1 convolutional layer, thereby realizing the alignment of DEM terrain data feature channels while keeping the spatial size unchanged; S32, the terrain complexity gating subnet extracts terrain complexity information from the aligned DEM terrain data features and generates an adaptive weight mask. ; S33, Dynamic fusion unit according to formula Dynamic feature map fusion is performed, where... The feature map after feature map fusion. This is the first multi-scale fused feature map. This is the second multi-scale fusion feature map after channel alignment. For the weight mask, This indicates element-wise multiplication.

[0018] Furthermore, in step S4, the segmentation head network consists of a feature enhancement unit, a category prediction unit, and a spatial upsampling unit, which perform the following operations in sequence: S41. The feature enhancement unit performs non-linear enhancement on the fused feature map through a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation function to improve feature discriminativeness. S42. The category prediction unit compresses the number of channels in the enhanced feature map to 2 through another 1×1 convolutional layer, corresponding to the category scores of the background and the landslide, respectively. S43. The spatial upsampling unit upsamples the category score map to a spatial size consistent with the input image through bilinear interpolation, generating the final landslide segmentation prediction map.

[0019] Furthermore, step S32 specifically includes: S321. The second multi-scale fusion feature map after channel alignment is input into the first convolutional layer of the terrain complexity gated subnet. The kernel size of the first convolutional layer is 3×3. The number of input channels is 720, which is the same as the number of channels of the aligned DEM feature. The number of output channels is 180. After convolution, it passes through a two-dimensional batch normalization layer and a ReLU activation function in sequence to extract intermediate features that reflect local terrain undulations. S322. The intermediate features are input into the second convolutional layer of the terrain complexity gated subnet. The kernel size of the second convolutional layer is 3×3. The number of input channels is 180, which is the same as the number of output channels of the first convolutional layer. The number of output channels is 720, which is the same as the number of channels of the first multi-scale fusion feature map. After convolution, the features pass through a two-dimensional batch normalization layer and a Sigmoid activation function. The Sigmoid activation function maps the convolutional output to the [0, 1] interval, generating a 720-channel weight mask. The weight mask It is dynamically adjusted according to the state of the DEM feature map.

[0020] Furthermore, the cross-entropy loss with class weights Boundary mask in It is obtained through the following steps: Landslide boundary region extraction is performed by using the Laplacian operator to extract the landslide boundary region from the labeled image of the target region in real time. The convolution kernel of the Laplacian operator is defined as follows:

[0021] Based on the Laplacian operator, the boundary pixel between the landslide and the background is located, and the boundary detection formula is:

[0022] in, For the label image, This indicates a convolution operation. For the detection results, a threshold is used to generate a binary boundary mask.

[0023] Boundary region dilation is achieved by performing a morphological dilation operation on the binary boundary mask to expand the scope of boundary supervision, using max pooling:

[0024] in, For boundary mask, Represented in pixels Centered on, with side length as neighborhood window, The initial boundary mask after binarization is used in the neighboring pixels The value at that location, These are the pixel coordinates within the neighborhood window.

[0025] Furthermore, the specific configuration of the HRNet-W48 network structure used in the optical feature extraction branch is as follows: Phase 1: Contains one Bottleneck module with 3 input channels and 64 output channels, followed by a batch normalization layer and a ReLU activation function to output the basic feature map; Phase 2: Contains one high-resolution module with a two-branch parallel structure, with 48 and 96 channels in each branch. It captures mid-to-low-scale texture features through multi-branch feature fusion and outputs a feature map with 144 channels. Phase 3: Contains 4 high-resolution modules, adopts a 3-branch parallel structure, with branch channels of 48, 96 and 192 respectively, and outputs a feature map with 336 channels; Phase 4: This phase consists of three high-resolution modules with a four-branch parallel structure. The number of channels in each branch is 48, 96, 192, and 384, respectively. Finally, the first multi-scale fusion feature map with 720 channels is output by feature interpolation and channel concatenation.

[0026] Furthermore, the HRNet-W18 network structure used in the terrain feature extraction branch is specifically configured as follows: Phase 1: Contains 1 Bottleneck module with 1 input channel and 64 output channels; Phase 2: Contains one high-resolution module, adopts a two-branch parallel structure, with 18 and 36 branch channels respectively; Phase 3: Contains 4 high-resolution modules, adopts a 3-branch parallel structure, and has 18, 36 and 72 branch channels respectively; Phase 4: Contains 3 high-resolution modules, adopts a 4-branch parallel structure, with branch channel numbers of 18, 36, 72 and 144 respectively, and finally outputs the second multi-scale fusion feature map with 270 channels.

[0027] Furthermore, the standardized preprocessing for format conversion, size unification, and normalization specifically includes: After the RGB image is read using OpenCV, it is converted from BGR format to RGB format; the DEM data is read in grayscale image form. If the DEM file does not exist or the reading fails, a zero-value matrix with the same size as the RGB image is automatically generated and filled. RGB images and DEM data are uniformly scaled to 512×512 pixels; The RGB images were normalized using the mean and standard deviation of the ImageNet dataset, while the DEM data were normalized independently to a mean of 0.5 and a standard deviation of 0.5.

[0028] The beneficial effects of the above-described technical solution of the present invention are as follows: 1. This invention, through its designed Terrain-Aware Gated Fusion Module (TAGF), changes the traditional method's approach of blindly stitching or using fixed weights for fusion. This module senses the intensity of surface undulations in real time through terrain branches and dynamically generates pixel-level weight masks accordingly. This mechanism allows the model to automatically enhance the guiding role of terrain geometry features in complex, landslide-prone areas, while automatically switching to optical textures in flat, interference-prone areas. This logically solves the false detection problem caused by different objects sharing the same spectrum, significantly improving the model's environmental adaptability in complex scenes.

[0029] 2. This invention employs an asymmetric dual-branch HRNet architecture, maintaining high-resolution feature maps throughout the entire feature extraction and fusion stage. Compared to the traditional encoder-decoder structure that first downsamples and compresses then upsamples to recover the data, this scheme avoids the irreversible loss of spatial details of the landslide body during hierarchical transmission. This consistently high-resolution representation makes the model highly sensitive to complex landslide shapes such as long strips and small areas, ensuring the spatial integrity of the recognition results.

[0030] 3. This invention establishes specialized supervision for landslide edges by introducing a synergistic mechanism of Boundary Contrast Loss (BCL) and Global Cross-Entropy Loss during the training phase. Utilizing the Laplacian operator for real-time extraction of topological boundaries, the model obtains enhanced boundary discrimination signals, forcing the feature space to generate more discriminative representations in the boundary regions. This solves the persistent problems of "adhesion and ambiguous localization between landslides and background areas," resulting in generated recognition boundaries that better match actual terrain transition lines and reducing the difficulty of subsequent vectorization processing.

[0031] 4. This invention employs an asymmetric computational resource allocation strategy (W48 for optical images, W18 for DEM images). Because optical images have high information entropy and complex textures, while DEM terrain features are relatively simple, this asymmetric design ensures that the model can deeply mine spectral details while avoiding computational redundancy in terrain feature extraction. Through this architectural decoupling and optimization, the model achieves optimal representation of the features of these two types of heterogeneous data with limited computational resources. Attached Figure Description

[0032] Figure 1 This is a flowchart of the landslide detection method based on adaptive fusion of optical satellite remote sensing and DEM data according to the present invention. Detailed Implementation

[0033] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0034] like Figure 1As shown, this invention proposes a landslide detection method based on adaptive fusion of optical satellite remote sensing and DEM data, comprising the following steps: S1. Data Acquisition and Preprocessing: Acquire RGB image data and corresponding DEM topographic data of the target area, and perform preprocessing such as format conversion, size unification and normalization.

[0035] S2. Heterogeneous feature extraction: The preprocessed RGB image data and DEM topographic data are respectively input into a two-branch high-resolution feature extraction network.

[0036] The first branch is the optical feature extraction branch, which processes the RGB image data and outputs the first multi-scale fusion feature map. The second branch is the terrain feature extraction branch, which processes the DEM terrain data and outputs a second multi-scale fusion feature map. S3. Terrain-aware feature fusion: Input the first multi-scale fused feature map and the second multi-scale fused feature map into the terrain-aware gating fusion module.

[0037] The terrain-aware gating fusion module dynamically generates a weight mask corresponding to the spatial size of the first multi-scale fusion feature map based on the terrain complexity information reflected by the second multi-scale fusion feature map. Using the weight mask, the first multi-scale fused feature map and the second multi-scale fused feature map after channel alignment are weighted and fused to generate a fused feature map; S4. Landslide Area Prediction: The fused feature map is input into the segmentation head network. After feature enhancement, category prediction and spatial upsampling, a landslide segmentation prediction map with the same size as the input image is output.

[0038] The working principle of each step of the present invention is explained in detail below: Step S1 mainly involves acquiring and preprocessing the dataset. Given the characteristics of multi-source data, it ensures the consistency and high quality of the input data, laying the foundation for subsequent feature extraction.

[0039] (1) Data set structure design.

[0040] The data directory structure includes four main modules: image (RGB optical imagery), dem (DEM topographic data), label (binary classification labels, 1 for landslides and 0 for background), and list (training / validation list files). The ratio of training set, validation set, and test set is 8:1:1.

[0041] (2) Preprocessing process.

[0042] Data reading and format conversion: After the RGB image is read by OpenCV, it is converted from BGR format to RGB format; DEM data is read in grayscale image form. If the DEM file does not exist or the reading fails, a zero-value matrix with the same size as the RGB image is automatically generated to fill it, so as to avoid training interruption caused by missing data; After the label image is read, it is binarized (pixel value > 128 is set to 1, otherwise it is set to 0) to ensure class consistency.

[0043] Size unification: RGB images, DEM data, and labels are uniformly adjusted to a target size of 512×512. RGB images and DEM data are interpolated using bilinear interpolation, while labels are interpolated using nearest neighbor interpolation to avoid category ambiguity caused by the interpolation process.

[0044] Normalization processing: RGB images were normalized using ImageNet parameters (mean [0.485, 0.456, 0.406], standard deviation [0.229, 0.224, 0.225]) to eliminate the dimensional differences of color channels; DEM data were normalized using independent parameters (mean [0.5], standard deviation [0.5]) to avoid interference from excessively large elevation values ​​on model training.

[0045] Through the above processing in step S1, RGB image data and DEM terrain data of the adaptive feature extraction network are obtained.

[0046] S2. Input the preprocessed RGB image data and DEM topographic data into a dual-branch high-resolution feature extraction network; where: (1) The first branch is the optical feature extraction branch, which processes the RGB image data and outputs the first multi-scale fusion feature map.

[0047] To address the feature differences between RGB optical data and DEM topographic data, a heterogeneous HRNet dual-branch architecture is designed to achieve specialized feature extraction for both types of data, balancing feature representation sufficiency and computational efficiency. HRNet-W48 is adopted as the core architecture, focusing on refined extraction of optical texture features. The specific configuration is as follows: Stage 1: Consists of one Bottleneck module, which takes 3-channel RGB preprocessed data as input and outputs a 64-channel basic feature map. The nonlinear expression is enhanced by a two-dimensional batch normalization layer (BatchNorm2d) and the ReLU activation function.

[0048] Phase 2: Contains one high-resolution module with a two-branch parallel structure. The number of channels in each branch is 48 and 96 respectively. It captures low- to medium-scale texture features through multi-branch feature fusion and outputs a feature map with 48+96=144 channels.

[0049] Phase 3: Contains 4 high-resolution modules with a 3-branch parallel structure. The number of channels in each branch is 48, 96, and 192, respectively. Feature extraction is deepened through 4 stacking of modules, and the output is a feature map with 48+96+192=336 channels.

[0050] Phase 4: Contains 3 high-resolution modules, employing a 4-branch parallel structure with branch channel counts of 48, 96, 192, and 384 respectively. It captures high, medium, and low-scale texture and spatial structure features through multi-scale feature interaction. Finally, it outputs a fused feature map with 48+96+192+384=720 channels through feature interpolation (adjusting the features of each branch to the same spatial size) and channel concatenation. ).

[0051] (2) The second branch is the terrain feature extraction branch, which processes the DEM terrain data and outputs the second multi-scale fusion feature map.

[0052] The lightweight HRNet-W18 architecture is adopted to reduce computational complexity while ensuring effective extraction of terrain geometric features. The specific configuration is as follows: the network structure is consistent with the RGB main branch, only the number of channels is reduced proportionally to adapt to the low-dimensional feature characteristics of DEM data.

[0053] Stage 1: Consists of a single Bottleneck block. The input is preprocessed data from a 1-channel DEM, and the output is a 64-channel basic terrain feature map. The nonlinear representation of terrain features is enhanced by using a two-dimensional batch normalization layer (BatchNorm2d) and the ReLU activation function, laying the foundation for subsequent multi-branch extraction.

[0054] Phase 2: Contains one high-resolution module with a two-branch parallel structure. The number of channels in each branch is 18 and 36 respectively. Multi-branch feature fusion is achieved by summing, accurately capturing low- to medium-scale terrain geometric features (such as small elevation differences and gentle slopes), and outputting a feature map with 18+36=54 channels.

[0055] Phase 3: Contains 4 high-resolution modules with a 3-branch parallel structure. The number of channels in each branch is 18, 36, and 72, respectively. The terrain feature extraction is deepened by stacking the modules 4 times, and the feature interaction of terrain structures at different scales (such as valleys and gentle slopes) is enhanced, outputting a feature map with 18+36+72=126 channels.

[0056] Phase 4: Contains three high-resolution modules with a four-branch parallel structure. The number of channels in each branch is 18, 36, 72, and 144, respectively. Through cross-resolution multi-scale feature interaction, it comprehensively captures high, medium, and low-scale terrain geometric features (such as steep slopes and large-scale terrain undulations). Finally, through feature interpolation (adjusting the features of each branch to the same spatial size) and channel concatenation, it outputs a terrain feature map with 18+36+72+144=270 channels. ).

[0057] Both branches use a two-dimensional batch normalization layer (BatchNorm2d) for layer normalization (momentum=0.1), enhance nonlinearity through the ReLU activation function, and achieve multi-scale feature interaction through the cross-branch fusion mechanism (SUM fusion method) in the high-resolution module to ensure the hierarchy and integrity of features.

[0058] Step S3's terrain-aware gated fusion module dynamically generates a weight mask corresponding to the spatial dimensions of the first multi-scale fusion feature map based on the terrain complexity information reflected in the second multi-scale fusion feature map. Using the weight mask, the first multi-scale fusion feature map and the channel-aligned second multi-scale fusion feature map are weighted and fused to generate fused features. The terrain-aware gated fusion module adaptively adjusts the fusion ratio based on terrain complexity, achieving dynamic complementary fusion of optical and terrain features, thus overcoming the adaptability limitations of traditional fixed fusion. The specific structure and workflow are as follows: (1) Overall structure of the module.

[0059] The terrain-aware gating fusion module consists of three parts: a DEM feature channel alignment unit, a terrain complexity gating subnet, and a dynamic fusion unit. The input is a 720-channel feature map of the RGB branch. ) and 270-channel feature map of DEM branch ( The output is a 720-channel fused feature map. ).

[0060] (2) Implementation of key units.

[0061] ①DEM feature channel alignment unit: due to (270 channels) and (720 channels) The number of channels is inconsistent, so a 1×1 convolutional layer is designed to achieve channel alignment, specifically: First, a two-dimensional convolutional layer is used, with the number of input channels set to 270. The number of channels is the same; the number of output channels is set to 720, which is consistent with... The number of channels is consistent; the kernel size is 1×1 to ensure that the spatial size of the feature map is not changed while adjusting the number of channels, and no bias term is added to avoid interference from additional parameters; After the convolution operation, a two-dimensional batch normalization layer (BatchNorm2d) is connected. The number of output channels of this layer is the same as that of the previous convolutional layer (720 channels). The momentum parameter of batch normalization is set to 0.1 to balance historical statistical information with current batch information and stabilize the training process. Finally, a nonlinear transformation is performed using the ReLU activation function, and the number of DEM feature channels is adjusted to 720 using in-situ operation mode while maintaining the spatial size. If there are differences in spatial size (e.g., due to network errors), bilinear interpolation is used to adjust the aligned DEM features to match the desired size. Consistent spatial dimensions (512 / 4×512 / 4=128×128).

[0062] ② Terrain Complexity Gated Subnet: Its core function is to extract terrain complexity information from the aligned DEM features and generate an adaptive weight mask. ( ).

[0063] The first convolutional layer of the gated subnet is used to extract local features of terrain complexity. Specifically, it is designed as follows: a two-dimensional convolutional layer is used, with 720 input channels, consistent with the number of channels in the aligned DEM features; the number of output channels is 180, calculated by a channel reduction ratio of 4, where 720 divided by 4 equals 180, thus compressing the feature dimension to reduce computation; the kernel size is set to 3×3 to capture the correlation of terrain features within a local 3×3 area; the edge padding is set to 1 to ensure that the spatial size of the feature map after convolution is consistent with the input, and no bias term is added. After the convolution operation, a two-dimensional batch normalization layer (BatchNorm2d) is applied. The number of output channels of this layer is the same as that of the current convolutional layer (180 channels). Batch normalization reduces the internal covariate shift and improves the stability of feature extraction. Finally, the ReLU activation function is used to introduce nonlinearity, which enhances the model's ability to express local features of terrain complexity and provides effective feature support for subsequent weight mask generation.

[0064] The first convolutional layer combines a two-dimensional batch normalization layer (BatchNorm2d(180)) with ReLU activation to extract local features of terrain complexity.

[0065] The second convolutional layer of the gated subnet is used to generate a terrain complexity weight mask. The specific design is as follows: a two-dimensional convolutional layer is used, with 180 input channels, consistent with the number of output channels in the first convolutional layer; the number of output channels is 720, consistent with the optical features. The number of channels in the input and the aligned DEM feature channels are consistent to ensure that the weight mask can match features channel by channel; the kernel size is set to 3×3 to further capture the local dependencies of terrain features; the edge padding is set to 1 to ensure that the spatial size of the output feature map is consistent with that of the input, and finally... Size matching, and no offset terms added; After the convolution operation, a two-dimensional batch normalization layer (BatchNorm2d) is connected. The number of output channels of this layer is the same as the number of output channels of the current convolutional layer (720 channels), which stabilizes the numerical distribution of the weight mask. Finally, the convolution output is mapped to the [0, 1] interval using the Sigmoid activation function to generate a 720-channel weight mask. The mask and The spatial dimensions are exactly the same. The larger the value, the higher the terrain complexity of the corresponding area, and the more the model will rely on DEM features to assist in landslide identification during subsequent fusion.

[0066] The second convolutional layer combines BatchNorm2d(720) with the Sigmoid activation function, and the output is the same as... 720-channel weighted masks of uniform size , A larger value indicates a higher level of terrain complexity in the corresponding area, requiring more reliance on DEM features.

[0067] ③ Dynamic fusion unit: Feature fusion is achieved using a residual fusion mechanism, as shown in the following formula:

[0068] in, This is the first multi-scale fused feature map. This is the second multi-scale fusion feature map after channel alignment. For the weight mask, This indicates element-wise multiplication. This mechanism ensures that optical features are the foundation, while terrain features are dynamically supplemented according to complexity, thus preserving detailed information of optical textures and suppressing misjudgments caused by foreign objects sharing the same spectrum through terrain features.

[0069] Step S4 inputs the fused feature map into the segmentation head network. After feature enhancement, category prediction, and spatial upsampling, it outputs a landslide segmentation prediction map with the same size as the input image. The segmentation head is a key link connecting the terrain perception fusion module and the final landslide classification output. Its core function is to nonlinearly enhance the high-dimensional fused features, compress the dimensions, predict the category, and restore the scale, achieving lightweight inference while ensuring feature discriminativeness. The specific structure and workflow are as follows: (1) Overall structure of the module.

[0070] The segmentation head adopts a three-stage architecture of "feature enhancement - category prediction - spatial reconstruction", consisting of a feature enhancement unit, a category prediction unit, and a spatial upsampling unit. The input is a 720-channel fused feature map output from the terrain perception fusion module. The initial prediction output is a 2-channel (landslide / background) class probability map (size...). The final output is a predicted mask with the same size as the input image (size). By employing the logic of "enhancing first and then reducing dimensionality," the model retains the differentiated semantic information between landslides and background in the fused features, while minimizing computational overhead through a 1×1 convolutional structure, ensuring that the model can balance detection accuracy and inference speed in complex mountainous environments.

[0071] (2) Implementation of key units.

[0072] ① Feature enhancement stage (first stage) The goal of this stage is to nonlinearly enhance the 720-channel fusion features to improve the inter-class distinguishability between landslide targets and background features.

[0073] Two-dimensional convolutional layer: A two-dimensional convolution with a kernel size of 1×1 is used. The number of input channels is set to 720, which is consistent with the number of channels of the TAGF output fusion features. The number of output channels is also set to 720 to keep the feature dimension unchanged and avoid information loss. No bias term is added to reduce training interference introduced by additional parameters. At the same time, the 1×1 convolutional kernel ensures that the spatial size of the feature map is not changed while enhancing the features (still 128×128). Only the channel dimension features are reorganized and enhanced.

[0074] Two-dimensional batch normalization layer (BatchNorm2d): Connected after the convolutional layer, it outputs 720 channels, and the momentum parameter for batch normalization is set to 0.1. This layer is used to stabilize the numerical distribution of high-dimensional features, alleviate the gradient vanishing problem in deep network training, and ensure that the fused features have consistent statistical properties before entering the prediction layer.

[0075] ReLU activation function: It directly modifies the original tensor in an in-situ operation mode, which introduces nonlinear transformation, strengthens the key semantic information of the landslide area in the fused features (such as the coupling features of terrain undulation and texture), and reduces memory usage, which is suitable for lightweight design requirements; nonlinear transformation can break the linear correlation between features, making it easier for the model to learn the complex distinguishing boundary between landslide and background.

[0076] ② Category prediction stage (second stage) The core objective of this stage is to accurately reduce the enhanced high-dimensional features to the category dimension and output the probability distribution of the landslide and background. The specific design is as follows: Two-dimensional convolutional layer: A two-dimensional convolution with a kernel size of 1×1 is used, with 720 input channels and 2 output channels (corresponding to the two categories of "landslide" and "background"). This layer does not have a bias term and compresses high-dimensional semantic information into a category score map through cross-channel weighted integration.

[0077] No additional activation functions or normalization layers: This stage directly outputs the raw class scores without introducing any activation functions or normalization layers. This design aims to preserve the original distribution characteristics of the predicted values ​​to ensure higher numerical accuracy of subsequent loss functions (such as weighted CE loss and BCL boundary contrast loss) when calculating gradients.

[0078] ③ Spatial upsampling stage (third stage) Bilinear interpolation restoration: Since the size of the preceding feature map is 1 / 4 of the original input ( The system introduces a bilinear interpolation algorithm to upsample the 2-channel prediction image to the original input size. Ensure that the predicted mask is precisely aligned with the original optical and topographic imagery at the pixel level.

[0079] In addition, to address issues such as sample imbalance and difficulty in boundary detection in landslide detection, a combination of loss functions was designed to ensure model convergence and detection accuracy.

[0080] (1) Combination of loss functions.

[0081] A combined loss function of "Cross-Entropy Loss (CE) + Boundary Contrast Loss (BCL)" is adopted to balance global classification accuracy and boundary detection accuracy. The combined loss function is as follows:

[0082] For the combined loss function, For cross-entropy loss with class weights, For boundary contrast loss, This is the balance coefficient; Boundary Contrast Loss (BCL) is specifically designed for the problem of ambiguous landslide boundaries. Its core mechanism is as follows: Boundary extraction. Boundary regions of the label image are extracted using the Laplacian operator (convolution kernel [[0,1,0], [1,-4,1], [0,1,0]]) to generate a boundary mask.

[0083] Boundary weighting. Higher weights are assigned to the loss for boundary regions (boundary weight = 5.0), forcing the model to focus on the classification accuracy of landslide boundary pixels.

[0084] The formula is:

[0085] in, , The height and width of the feature map; For cross-entropy loss, For the model to pixels The probability of predicting it as a landslide. For pixels The true category labels, among which , For pixel coordinate index, For boundary adaptive weights, For valid pixel indicator functions, when pixel The label value is set to 1 if it is not an ignored value, otherwise it is set to 0; boundary adaptive weights. The calculation formula is:

[0086] in, This is the boundary mask, and can be either 0 or 1. These are the boundary weight coefficients. ; Cross-entropy loss (CE) is employed with class weights, set to [1.0, 3.0] (the landslide class weight is 3.0), to alleviate the sample imbalance problem caused by the low proportion of landslide samples. The formula is:

[0087] in, The spatial dimensions of the feature map; This refers to the number of effective pixels. For valid pixel indicator functions, when pixel The value is 1 if the label value is not an ignored value, and 0 otherwise. For pixels The true category label, 0 represents background, 1 represents landslide; For the model to pixels Predicted as category The probability of; This is the category weight vector.

[0088] The above-described technical solution of the present invention achieves the following innovations: 1. Asymmetric dual-branch ORID-HRNet high-resolution parallel feature extraction architecture An asymmetric dual-branch parallel architecture of ORID-HRNet was adopted. Addressing the fundamental differences in features and information representation requirements between RGB optical data and DEM terrain data, differentiated allocation of computational power and customized architecture adaptation were implemented to achieve professional and accurate extraction of features from both types of data. The main optical branch employs a high-capacity HRNet-W48 network configuration, leveraging its multi-scale parallel feature interaction capabilities to deeply mine high-dimensional spectral texture details and spatial structure features in RGB data, ultimately outputting a 720-channel high-discriminative optical feature map. The terrain auxiliary branch uses a lightweight HRNet-W18 network configuration. This branch maintains structural consistency with the optical branch in terms of network stages, number of modules, and number of parallel branches, only reducing the number of channels to adapt to terrain features. While accurately capturing the core features of terrain geometry, this significantly reduces the computational parameters and power consumption of the terrain branch, ultimately outputting a 270-channel terrain feature map. Both branches use a 64-channel Bottleneck structure to complete the initial feature extraction in Stage 1. Starting from Stage 2, the number of channels is configured differently. The optical branch increases the number of channels at multiple scales in the order of [48, 96, 192, 384], while the terrain branch is configured to be reduced proportionally in the order of [18, 36, 72, 144]. The channel ratio is maintained at about 2.67:1, which matches the difference in information density between the two types of data.

[0089] To address the differences in feature density and information entropy between heterogeneous data (RGB and DEM), an asymmetric computational power allocation is used to achieve optimal feature representation. This architecture maintains high-resolution representation throughout the feature extraction process, avoiding the loss of landslide spatial details in traditional downsampling and ensuring high sensitivity in capturing minute landslide bodies. The asymmetric design balances feature extraction depth and computational efficiency, and the consistently high resolution significantly improves the model's ability to capture slender, irregular, small-area landslide hazards.

[0090] 2. Terrain-Aware Gated Fusion Mechanism (TAGF) This technical approach aims to address the problem of improper utilization of terrain features caused by "spatial heterogeneity" in multi-source data fusion.

[0091] Following the dual-branch feature extraction layer, a feature flow derived from terrain branches was designed. A gated network module driven by a driver. This module consists of two layers. Convolutional layers and the Sigmoid activation function. This gated network does not rely on manually set weights, but instead utilizes... Real-time calculation of pixel-level attention masks with the same spatial dimensions as the feature map (range of values) The final fused features follow the formula. .

[0092] When the DEM feature map exhibits the following pattern at a certain location, the convolution response value increases, and after passing through the Sigmoid function, α approaches 1: the feature values ​​of adjacent pixels differ greatly (corresponding to dramatic changes in terrain elevation), the feature values ​​are unevenly distributed within a local 3×3 window (corresponding to complex terrain undulations), and the overall feature activation value is relatively high (the DEM branch extracts significant terrain information in this area).

[0093] When the DEM feature map exhibits the following pattern at a certain location, the convolution response value decreases, and α approaches 0 after passing through the Sigmoid: the feature values ​​of adjacent pixels are similar (corresponding to flat terrain), the feature values ​​are evenly distributed within the local 3×3 window (corresponding to simple terrain), and the overall feature activation value is low (the DEM branch did not extract effective shape information in this area).

[0094] By using a gated subnet (composed of two convolutional layers and a sigmoid function) to perceive the geometric undulations of the local terrain, the weight of terrain features is increased at locations with dramatic slope changes, and terrain noise interference is reduced in flat areas, thus achieving efficient complementarity between optical textures and terrain geometric features.

[0095] 3. CE+BCL Combined Loss Monitoring Mechanism Boundary extraction: During model training, the Laplacian operator is used to perform real-time convolution on the ground-value landslide labels to accurately locate the topological boundary region where the landslide meets the background. A combined loss function is used, as shown in the following formula:

[0096] in, The global cross-entropy loss, weighted by landslide category, is responsible for the basic classification of the entire graph. To compare the loss at the landslide boundary, a higher loss weight is applied to the boundary area than to the non-boundary area.

[0097] In landslide remote sensing detection scenarios, the ratio of landslide pixels to background pixels is severely imbalanced. Taking the Bijie landslide dataset used in this solution as an example, landslide pixels typically account for less than 10%. If standard cross-entropy loss is directly applied for training, the model tends to predict most pixels as background to minimize the loss value, resulting in a high false negative rate for landslide areas. To address this issue, a class weighting mechanism is introduced to adjust the cross-entropy loss. The formula is as follows:

[0098] in, The spatial dimensions of the feature map (after sampling from the segmentation head and the original). Figure 1 (All are 512×512). The number of valid pixels (excluding the ignored area with a label value of 255); For valid pixel indicator functions, when pixel The value is 1 if the label value is not an ignored value, and 0 otherwise. For pixels The true category label (0 represents background, 1 represents landslide); For the model to pixels Predicted as category The probability (obtained by Softmax normalization after the TAGF fusion feature is output by the segmentation head); This is the category weight vector.

[0099] Category weight settings: Set the category weight vector to... ,in, Background class weights, used as benchmark weights; The weight of the landslide class is higher than that of the background class, which makes the model impose a greater penalty on the classification error of landslide samples during training. This weight configuration makes the contribution of the classification error of one landslide pixel to the loss value equivalent to the classification error of three background pixels, effectively improving the model's sensitivity to the minority class (landslide) and reducing the false negative rate.

[0100] Synergistic effect with TAGF fusion features: The predicted probability map is the result of processing the fused features directly applied to the TAGF module output by the segmentation head. Since the TAGF fused features integrate optical texture information and DEM terrain geometry information... The supervisory signal is backpropagated to simultaneously optimize the parameters of the RGB branch, DEM branch and TAGF gated subnet, so that the three can learn the most effective feature representation method for landslide identification.

[0101] Landslide boundaries are crucial for distinguishing landslide bodies from their surroundings, and their detection accuracy directly impacts the accuracy of landslide area estimation and risk assessment. However, in actual remote sensing imagery, landslide boundaries often face the following challenges: At the optical image level: vegetation cover blurs the boundary's spectral characteristics, and the similar textures of bare soil and landslide bodies lead to misjudgments. At the DEM data level: topographic transition lines are unclear in low-resolution DEMs, and gradual changes in boundary elevation result in smooth transitions in topographic features. To address these issues, a boundary contrast loss is designed as an auxiliary supervisory signal. By applying higher loss weights to boundary pixels, the model is forced to generate more discriminative feature representations in the boundary region.

[0102] Landslide boundary region extraction: The Laplacian operator is used to extract landslide boundary regions from the labeled image in real time. The Laplacian operator is a second-order differential operator that can detect regions in an image where grayscale values ​​change abruptly. Its convolution kernel is defined as follows:

[0103] Applying this operator to the labeled image (background value 0, landslide value 1) can accurately locate the boundary pixels between the landslide and the background. The boundary detection formula is as follows:

[0104] in, For the label image, This represents a convolution operation. For the detection results, a threshold is used to generate a binary boundary mask.

[0105] Boundary region dilation: The original boundary mask only contains a boundary line with a width of one pixel, limiting its monitoring range. To expand the scope of boundary monitoring, a morphological dilation operation is performed on the binary boundary mask, implemented using max pooling.

[0106] in, For boundary mask, Represented in pixels Centered on, with side length as neighborhood window, The initial boundary mask after binarization is used in the neighboring pixels The value at that location, These are the pixel coordinates within the neighborhood window. This scheme sets the boundary width parameter to 3 pixels, corresponding to the dilation kernel size. This ensures that the final boundary region covers a range of 3 pixels on each side of the original boundary line, forming a strip-shaped boundary supervision region.

[0107] The formula is as follows:

[0108] in, For pixels Cross-entropy loss, The boundary adaptive weights are calculated using the following formula:

[0109] in, The boundary weight coefficient is set to 5.0 in this scheme. This formula makes the non-boundary region ( ): Standard loss weights are used; boundary region ( ): The loss weighting is increased to 5 times.

[0110] Synergistic effect with TAGF fusion features: The supervisory signal generates a gradient intensity five times greater in the boundary region than in the non-boundary region. Backpropagation is used to optimize the feature fusion strategy of the TAGF module at the boundary pixels. Specifically, the boundary region typically corresponds to the back wall or side wall of a landslide with dramatic topographic relief. The TAGF gating weights... In such areas, the boundary tends to increase, making full use of DEM terrain features to assist in boundary localization. Enhanced boundary supervision enables the RGB branch to learn clearer boundary texture features, while the DEM branch learns more accurate terrain transition line features. The two are then adaptively fused through TAGF to form a fused representation with stronger boundary discrimination capabilities.

[0111] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A landslide detection method based on adaptive fusion of optical satellite remote sensing and DEM data, characterized in that, Includes the following steps: S1. Acquire RGB image data and corresponding DEM terrain data of the target area, and perform standardized preprocessing such as format conversion, size unification and normalization; S2. The preprocessed RGB image data and DEM topographic data are respectively input into a dual-branch high-resolution feature extraction network for processing, and output a multi-scale fusion feature map; wherein, the first branch is an optical feature extraction branch, which processes the RGB image data and outputs a first multi-scale fusion feature map; the second branch is a topographic feature extraction branch, which processes the DEM topographic data and outputs a second multi-scale fusion feature map; S3. Input the first multi-scale fused feature map and the second multi-scale fused feature map into the terrain-aware gating fusion module; the terrain-aware gating fusion module dynamically generates a weight mask corresponding to the spatial size of the first multi-scale fused feature map based on the terrain complexity information reflected by the second multi-scale fused feature map; using the weight mask, the first multi-scale fused feature map and the second multi-scale fused feature map after channel alignment are weighted and fused to generate a fused feature map; S4. Input the fused feature map into the segmentation head network. After feature enhancement, category prediction and spatial upsampling, output a landslide segmentation prediction map with the same size as the input image.

2. The method according to claim 1, characterized in that, This also includes, during the entire landslide detection model training phase, using a combined loss function to supervise and optimize the network parameters. The combined loss function is: For the combined loss function, For cross-entropy loss with class weights, For boundary contrast loss, This is the balance coefficient; Among them, the cross-entropy loss with class weights The formula is: in, , The height and width of the feature map; For cross-entropy loss, For the model to pixels The probability of predicting it as a landslide. For pixels The true category labels, among which , For pixel coordinate index, For boundary adaptive weights, For valid pixel indication functions, when pixel The label value is set to 1 if it is not an ignored value, otherwise it is set to 0; boundary adaptive weights The calculation formula is: in, For boundary mask, These are the boundary weight coefficients; The formula is: in, The spatial dimensions of the feature map; This refers to the number of effective pixels. For valid pixel indication functions, when pixel The value is 1 if the label value is not an ignored value, and 0 otherwise. For pixels The true category label, 0 represents background, 1 represents landslide; For the model to pixels Predicted as category The probability of; This is the category weight vector.

3. The method according to claim 1, characterized in that, In step S2, the optical feature extraction branch adopts the HRNet-W48 network structure, and the terrain feature extraction branch adopts the HRNet-W18 network structure. Both branches use two-dimensional batch normalization layers for layer normalization, enhance nonlinearity through the ReLU activation function, and achieve multi-scale feature interaction through the cross-branch fusion mechanism in the high-resolution module.

4. The method according to claim 1, characterized in that, In step S3, the terrain-aware gating fusion module consists of three parts: a DEM feature channel alignment unit, a terrain complexity gating subnet, and a dynamic fusion unit, and performs the following operations: S31, the DEM feature channel alignment unit adjusts the number of channels of the second multi-scale fusion feature map to be the same as that of the first multi-scale fusion feature map through a 1×1 convolutional layer, thereby realizing the alignment of DEM terrain data feature channels while keeping the spatial size unchanged; S32, the terrain complexity gating subnet extracts terrain complexity information from the aligned DEM terrain data features and generates an adaptive weight mask. ; S33, Dynamic fusion unit according to formula Dynamic feature map fusion is performed, where... The feature map after feature map fusion. This is the first multi-scale fused feature map. This is the second multi-scale fusion feature map after channel alignment. For the weight mask, This indicates element-wise multiplication.

5. The method according to claim 1, characterized in that, In step S4, the segmentation head network consists of a feature enhancement unit, a category prediction unit, and a spatial upsampling unit, and performs the following operations in sequence: S41. The feature enhancement unit performs non-linear enhancement on the fused feature map through a 1×1 convolutional layer, a batch normalization layer, and a ReLU activation function to improve feature discriminativeness. S42. The category prediction unit compresses the number of channels in the enhanced feature map to 2 through another 1×1 convolutional layer, corresponding to the category scores of the background and the landslide, respectively. S43. The spatial upsampling unit upsamples the category score map to a spatial size consistent with the input image through bilinear interpolation, generating the final landslide segmentation prediction map.

6. The method according to claim 4, characterized in that, Step S32 specifically involves: S321. The second multi-scale fusion feature map after channel alignment is input into the first convolutional layer of the terrain complexity gated subnet. The kernel size of the first convolutional layer is 3×3. The number of input channels is 720, which is the same as the number of channels of the aligned DEM feature. The number of output channels is 180. After convolution, it passes through a two-dimensional batch normalization layer and a ReLU activation function in sequence to extract intermediate features that reflect local terrain undulations. S322. The intermediate features are input into the second convolutional layer of the terrain complexity gated subnet. The kernel size of the second convolutional layer is 3×3. The number of input channels is 180, which is the same as the number of output channels of the first convolutional layer. The number of output channels is 720, which is the same as the number of channels of the first multi-scale fusion feature map. After convolution, the features pass through a two-dimensional batch normalization layer and a Sigmoid activation function. The Sigmoid activation function maps the convolutional output to the [0, 1] interval, generating a 720-channel weight mask. The weight mask It is dynamically adjusted according to the state of the DEM feature map.

7. The method according to claim 2, characterized in that, The cross-entropy loss with class weights Boundary mask in It is obtained through the following steps: Landslide boundary region extraction is performed by using the Laplacian operator to extract the landslide boundary region in real time from the labeled image of the target area. The convolution kernel of the Laplacian operator... The definition is as follows: Based on the Laplacian operator, the boundary pixel between the landslide and the background is located, and the boundary detection formula is: in, For the label image, This indicates a convolution operation. For the detection results, a threshold is used to generate a binary boundary mask. Boundary region dilation is achieved by performing a morphological dilation operation on the binary boundary mask to expand the scope of boundary supervision, using max pooling: in, For boundary mask, Represented in pixels Centered on, with side length as neighborhood window, The initial boundary mask after binarization is used in the neighboring pixels The value at that location, These are the pixel coordinates within the neighborhood window.

8. The method according to claim 3, characterized in that, The specific configuration of the HRNet-W48 network structure used in the optical feature extraction branch is as follows: Phase 1: Contains one Bottleneck module with 3 input channels and 64 output channels, followed by a batch normalization layer and a ReLU activation function to output the basic feature map; Phase 2: Contains one high-resolution module with a two-branch parallel structure, with 48 and 96 channels in each branch. It captures mid-to-low-scale texture features through multi-branch feature fusion and outputs a feature map with 144 channels. Phase 3: Contains 4 high-resolution modules, adopts a 3-branch parallel structure, with branch channels of 48, 96 and 192 respectively, and outputs a feature map with 336 channels; Phase 4: This phase consists of three high-resolution modules with a four-branch parallel structure. The number of channels in each branch is 48, 96, 192, and 384, respectively. Finally, the first multi-scale fusion feature map with 720 channels is output by feature interpolation and channel concatenation.

9. The method according to claim 3, characterized in that, The HRNet-W18 network structure used in the terrain feature extraction branch is specifically configured as follows: Phase 1: Contains 1 Bottleneck module with 1 input channel and 64 output channels; Phase 2: Contains one high-resolution module, adopts a two-branch parallel structure, with 18 and 36 branch channels respectively; Phase 3: Contains 4 high-resolution modules, adopts a 3-branch parallel structure, and has 18, 36 and 72 branch channels respectively; Phase 4: Contains 3 high-resolution modules, adopts a 4-branch parallel structure, with branch channel numbers of 18, 36, 72 and 144 respectively, and finally outputs the second multi-scale fusion feature map with 270 channels.

10. The method according to claim 1, characterized in that, The standardized preprocessing, which includes format conversion, size unification, and normalization, specifically includes: After the RGB image is read using OpenCV, it is converted from BGR format to RGB format; the DEM data is read in grayscale image form. If the DEM file does not exist or the reading fails, a zero-value matrix with the same size as the RGB image is automatically generated and filled. RGB images and DEM data are uniformly scaled to 512×512 pixels; The RGB images were normalized using the mean and standard deviation of the ImageNet dataset, while the DEM data were normalized independently to a mean of 0.5 and a standard deviation of 0.5.

Images