A landslide identification method and system based on bidirectional cross-attention calibration fusion

Abstract

This invention discloses a landslide recognition method and system based on bidirectional cross-attention calibration fusion, belonging to the field of landslide recognition technology. The invention obtains a first feature map list with multi-resolution local features through a multi-scale convolutional hybrid attention algorithm, and a second feature map list with global context features through a state-space algorithm. It then fuses these feature maps using a bidirectional cross-attention fusion algorithm to obtain a more accurate and comprehensive third feature map list where local details are enhanced under the guidance of global semantics, and the global context is enriched by specific local evidence. Finally, it obtains a fourth feature map list through a bidirectional calibration fusion algorithm, which can fuse both high-resolution spatial details and strong semantic features at low resolution scales. Finally, it performs image segmentation on the fourth feature map list using an attention mechanism and an object context representation mechanism, achieving accurate and comprehensive landslide recognition.

Description

Technical Field

[0001] This invention belongs to the field of landslide recognition technology, and in particular relates to a landslide recognition method and system based on bidirectional cross-attention calibration fusion. Background Technology

[0002] Landslides are among the most common and destructive geological hazards, primarily triggered by seismic activity and heavy rainfall. They are typically characterized by their suddenness and wide-ranging impact, leading to disruptions in transportation networks and damage to critical infrastructure. Against the backdrop of drastic global climate change, the frequency and severity of landslides are expected to continue to rise, making landslide identification crucial. Remote sensing imagery can clearly reveal the morphology, boundaries, and surface texture of landslide bodies, providing an irreplaceable data foundation for accurately defining landslide extent and identifying hazard details, making it a key tool for precise landslide investigation and monitoring.

[0003] Currently, landslides exhibit significant scale variations and irregular morphological features in high-resolution remote sensing images. These characteristics make it essential to simultaneously capture fine-grained local features and global contextual relationships for accurate segmentation. Therefore, there are four main landslide identification methods based on remote sensing images: visual interpretation methods, change detection-based methods, machine learning-based methods, and deep learning-based methods. Visual interpretation methods rely on analyzing the texture and geometric features of landslides in remote sensing images to detect landslide boundaries. This method largely depends on the analyst's expertise, resulting in low efficiency and accuracy. Change detection techniques rapidly capture changes in landform, vegetation cover, or land structure caused by landslides through multi-temporal remote sensing images, thereby assessing the occurrence and evolution of landslides. However, this method usually depends on manually set thresholds or rules, and its resistance to noise and nonlinear changes is poor, leading to frequent false alarms and missed detections during landslide monitoring, thus affecting the accuracy and reliability of the detection results. Machine learning methods classify pixels using feature extraction functions and sample learning, identifying landslides while managing massive datasets. However, they overemphasize low-level features, potentially neglecting the broader semantic context of images. This deficiency manifests directly in the prevalent noise and blurred landslide boundaries in the resulting images, severely limiting their accuracy in landslide identification. Unlike traditional methods relying on manual visual interpretation, change detection, and machine learning, deep learning techniques improve the reliability of results by extracting images and building mathematical models, providing end-to-end extraction of image semantic features from low to high. However, due to inherent limitations in local receptive fields, these deep learning models are naturally insufficient in capturing global information. Furthermore, because the models cannot effectively integrate key contextual information from distant but semantically relevant regions, they often lead to incomplete segmentation of large-scale landslides and missed detections in complex terrain, affecting the comprehensiveness and accuracy of landslide identification. Summary of the Invention

[0004] This invention aims to provide a landslide identification method and system based on bidirectional cross-attention calibration fusion to solve the above-mentioned technical problems. By processing feature maps at different resolution scales through bidirectional cross-attention fusion algorithm and bidirectional calibration fusion algorithm, the accuracy and comprehensiveness of landslide identification are improved.

[0005] To address the aforementioned technical problems, this invention provides a landslide identification method based on bidirectional cross-attention calibration fusion, comprising: acquiring a remote sensing image of a mountain; extracting features from the remote sensing image to obtain an initial feature map; performing multi-resolution local feature extraction on the initial feature map based on a preset multi-scale convolutional hybrid attention algorithm to obtain a first feature map of the remote sensing image at each preset resolution scale, thereby determining a first feature map list of the remote sensing image; performing global context feature extraction on the initial feature map based on a preset state space algorithm to obtain a second feature map of the remote sensing image at each resolution scale, thereby determining a second feature map list of the remote sensing image; fusing the first and second feature map lists based on a preset bidirectional cross-attention fusion algorithm and resolution scale to determine a third feature map list of the remote sensing image at each resolution scale; performing bidirectional calibration fusion on every two adjacent third feature maps in the third feature map list sequentially based on a preset bidirectional calibration fusion algorithm and resolution scale to update every two adjacent third feature maps in the third feature map list, thereby obtaining a fourth feature map list of the remote sensing image; and performing image segmentation on the fourth feature map list based on an attention mechanism and an object context representation mechanism to determine the landslide identification result of the mountain.

[0006] Understandably, this invention extracts multi-resolution local features using a multi-scale convolutional hybrid attention algorithm to obtain a first feature map list with multi-resolution local features, and extracts global context features using a state-space algorithm to obtain a second feature map list with global context features. Then, a bidirectional cross-attention fusion algorithm is used to fuse the first and second feature map lists, resulting in a more accurate and comprehensive third feature map list where local details are enhanced under the guidance of global semantics, and the global context is enriched by specific local evidence. Next, a bidirectional calibration fusion algorithm is used to perform bidirectional calibration fusion on every two adjacent third feature maps in the third feature map list, resulting in a fourth feature map list that can fuse both more spatial details at high-resolution scales and strong feature semantics at low-resolution scales. Finally, an attention mechanism and an object context representation mechanism are used to perform image segmentation on the fourth feature map list, achieving accurate and comprehensive landslide identification. This invention improves the accuracy and comprehensiveness of landslide identification by processing feature maps at different resolution scales using bidirectional cross-attention fusion and bidirectional calibration fusion algorithms. Attached Figure Description

[0007] Figure 1 A flowchart illustrating the steps of a landslide identification method based on bidirectional cross-attention calibration fusion, as provided in this embodiment of the invention; Figure 2 A schematic diagram of a dual-branch collaborative landslide identification model provided in an embodiment of the present invention; Figure 3 A schematic diagram of a multi-scale convolutional hybrid attention block provided in an embodiment of the present invention; Figure 4 This is a comparative schematic diagram of multi-resolution feature extraction provided in an embodiment of the present invention; wherein, Figure 4 (a) is a schematic diagram of a remote sensing image provided in an embodiment of the present invention. Figure 4 (b) A schematic diagram of a truth label for a remote sensing image provided in an embodiment of the present invention; Figure 4 (c) A schematic diagram of a conventional feature extraction method provided in an embodiment of the present invention; Figure 4 (d) is a schematic diagram of the multi-scale convolutional hybrid attention block output provided in an embodiment of the present invention; Figure 5 A schematic diagram of a visual basic block provided in an embodiment of the present invention; Figure 6 A schematic diagram of a selective scanning module provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of parallel scanning calculation for four-way scanning provided in an embodiment of the present invention; Figure 8 A schematic diagram of a bidirectional cross-attention fusion module provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of a bidirectional calibration fusion module provided in an embodiment of the present invention; Figure 10 A schematic diagram of a residual block branch provided in an embodiment of the present invention; Figure 11 A schematic diagram of an object context representation module based on SimAM enhancement provided in an embodiment of the present invention; Figure 12 This is a visual identification diagram illustrating different landslide identification methods provided in embodiments of the present invention; wherein, Figure 12 (a) A schematic diagram of a remote sensing image provided in an embodiment of the present invention; Figure 12 (b) A truth map of a remote sensing image provided in an embodiment of the present invention; Figure 12 (c) A landslide identification result diagram of an existing FCN model provided in an embodiment of the present invention; Figure 12 (d) A landslide identification result diagram of an existing Unet provided in an embodiment of the present invention; Figure 12 (e) A landslide identification result diagram of an existing PSP-net model provided in an embodiment of the present invention; Figure 12 (f) A landslide identification result diagram of an existing UPerNet model provided in an embodiment of the present invention; Figure 12 (g) A landslide identification result diagram of an existing DeeplabV3+ model provided in an embodiment of the present invention; Figure 12(h) is a landslide identification result diagram of an existing HRnet model provided in an embodiment of the present invention; Figure 12 (i) A landslide identification result diagram of an existing HRFormer model provided in an embodiment of the present invention; Figure 12 (j) is a landslide identification result diagram of an existing HRTrans model provided in an embodiment of the present invention; Figure 12 (k) is a landslide identification result diagram of an existing ConvNeXt model provided in an embodiment of the present invention; Figure 12 (l) A landslide identification result diagram of an existing TransUNet model provided in an embodiment of the present invention; Figure 12 (m) is a landslide identification result diagram of an existing Vmamba model provided in an embodiment of the present invention; Figure 12 (n) is a landslide identification result diagram of an existing Swin-Unet model provided in an embodiment of the present invention; Figure 12 (o) is a landslide identification result diagram of the dual-branch collaborative landslide identification model provided in the embodiment of the present invention; Figure 13 This is a schematic diagram of a landslide identification system based on bidirectional cross-attention calibration fusion, provided as an embodiment of the present invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] Example 1 Please refer to Figure 1 , Figure 1 The flowchart of a landslide identification method based on bidirectional cross-attention calibration fusion provided in this embodiment of the invention includes steps S101 to S106. Figure 2 This is a schematic diagram of a dual-branch collaborative landslide identification model provided in an embodiment of the present invention.

[0010] Step S101: Obtain remote sensing images of the mountain, extract features from the remote sensing images, and obtain an initial feature map.

[0011] Step S102: Based on the preset multi-scale convolutional hybrid attention algorithm, perform multi-resolution local feature extraction on the initial feature map to obtain the first feature map of the remote sensing image at each preset resolution scale, and determine the list of the first feature maps of the remote sensing image.

[0012] Step S103: Based on the preset state space algorithm, perform global context feature extraction on the initial feature map to obtain the second feature map of the remote sensing image at each resolution scale, and then determine the list of second feature maps of the remote sensing image.

[0013] Step S104: Based on the preset bidirectional cross-attention fusion algorithm and resolution scale, perform feature map fusion on the first feature map list and the second feature map list to determine the third feature map list of the remote sensing image at each resolution scale.

[0014] Step S105: Based on the preset bidirectional calibration fusion algorithm and resolution scale, perform bidirectional calibration fusion on every two adjacent third feature maps in the third feature map list in sequence to update every two adjacent third feature maps in the third feature map list, and obtain the fourth feature map list of the remote sensing image.

[0015] Step S106: Based on the attention mechanism and object context representation mechanism, perform image segmentation on the fourth feature map list to determine the landslide identification result of the mountain.

[0016] In an optional embodiment, the dual-branch collaborative landslide identification model includes: a start block (Stem), an HRnet branch, a Mamba branch, a bidirectional calibration fusion module (BCFM), a bidirectional cross-attention fusion module (BiCAF), and an object context representation module based on SimAM enhancement (OCR-Head with SimAM Enhancement). The Mamba branch includes three vision selective state blocks (VSSBlocks) of the Mamba state space model. The HRnet branch includes four stages (Stage 1, Stage 2, Stage 3, and Stage 4). Further, Stem is used to implement step S101, the HRnet branch is used to implement step S102, the Mamba branch is used to implement step S103, the bidirectional cross-attention fusion module is used to implement step S104, the bidirectional calibration fusion module is used to implement step S105, and the object context representation module based on SimAM enhancement is used to implement step S106. The preset resolution scales include: H / 4×W / 4×C1, H / 8×W / 8×C2, H / 16×W / 16×C3 and H / 32×W / 32×C4.

[0017] In one optional embodiment, feature extraction is performed on the remote sensing image to obtain an initial feature map, including: the initial block comprising two consecutive 3×3 convolutional layers; and the obtained dimension is... The remote sensing image is input into a starting block. Two 3×3 convolutional layers in the starting block downsample and extract features from the remote sensing image, resulting in a dimensionality of... The initial feature map.

[0018] This embodiment employs a multi-scale convolutional hybrid attention algorithm to extract multi-resolution local features, resulting in a first feature map list with multi-resolution local features. A state-space algorithm is then used to extract global context features, resulting in a second feature map list with global context features. Subsequently, a bidirectional cross-attention fusion algorithm is used to fuse the first and second feature map lists, yielding a more accurate and comprehensive third feature map list where local details are enhanced under the guidance of global semantics, and the global context is enriched by specific local evidence. Next, a bidirectional calibration fusion algorithm is used to perform bidirectional calibration fusion on every two adjacent third feature maps in the third feature map list, resulting in a fourth feature map list that can fuse both high-resolution spatial details and strong semantic features at low-resolution scales. Finally, an attention mechanism and an object context representation mechanism are used to perform image segmentation on the fourth feature map list, achieving accurate and comprehensive landslide identification. This invention improves the accuracy and comprehensiveness of landslide identification by processing feature maps at different resolution scales using bidirectional cross-attention fusion and bidirectional calibration fusion algorithms.

[0019] In this embodiment, multi-resolution local feature extraction is performed on the initial feature map based on a preset multi-scale convolutional hybrid attention algorithm to obtain the first feature map of the remote sensing image at each preset resolution scale, and the first feature map list of the remote sensing image is determined, including: extracting features from the initial feature map based on preset residual blocks to determine the fifth feature map; initializing the fifth feature map list based on the fifth feature map; determining the number of multi-resolution feature extractions and the resolution scale corresponding to each multi-resolution feature extraction based on the number of preset resolution scales; and refining the fifth feature map list based on the preset multi-scale convolutional hybrid attention algorithm and the resolution scale corresponding to each multi-resolution feature extraction. Iterative multi-resolution feature extraction is performed until the required number of extractions is reached, thus updating the fifth feature map list. Based on the updated fifth feature map list, the first feature map of the remote sensing image at each preset resolution scale is obtained, thereby determining the first feature map list of the remote sensing image. In each iteration of multi-resolution feature extraction, a fifth feature map is added to the fifth feature map list based on the resolution scale corresponding to the current multi-resolution feature extraction. A preset multi-scale convolutional hybrid attention algorithm is used to extract features from each fifth feature map in the fifth feature map list, and the fifth feature map in the fifth feature map list is updated based on the feature extraction results.

[0020] This embodiment extracts features from the initial feature map using residual blocks to obtain the fifth feature map, thereby initializing the list of fifth feature maps required for multi-resolution local feature extraction. Then, the number of multi-resolution feature extraction iterations is determined by the number of resolution scales. The list of fifth feature maps is iteratively updated using a multi-scale convolutional attention algorithm to reach the required number of multi-resolution feature extraction iterations, resulting in the first feature map of the remote sensing image at each preset resolution scale, and thus determining the first feature map list of the remote sensing image. This approach avoids the limitations and deviations in feature extraction caused by overemphasizing low-level features and single-resolution scale convolution in traditional machine learning methods. It significantly improves the accuracy and comprehensiveness of local feature extraction across multiple resolution scales, thereby enhancing the accuracy and comprehensiveness of subsequent landslide identification.

[0021] In this embodiment, in each iteration of multi-resolution feature extraction, a fifth feature map is added to the fifth feature map list based on the resolution scale corresponding to the current multi-resolution feature extraction. A preset multi-scale convolutional hybrid attention algorithm is used to extract features from each fifth feature map in the fifth feature map list, and each fifth feature map in the fifth feature map list is updated based on the feature extraction results. This includes: in each iteration of multi-resolution feature extraction, downsampling the fifth feature map based on the resolution scale corresponding to the current multi-resolution feature extraction to form a fifth feature map at the resolution scale corresponding to the current multi-resolution feature extraction, thereby updating the fifth feature map list; performing continuous dilated convolution on each fifth feature map, and performing normalization and ReLU activation after each dilated convolution to determine the sixth feature map corresponding to each fifth feature map; and performing an identity residual connection between the fifth feature map and the sixth feature map to determine the sixth feature map corresponding to each fifth feature map. The seventh feature map corresponds to the fifth feature map; each feature channel in each seventh feature map is calibrated based on the channel attention mechanism to generate the attention weight of each feature channel in each seventh feature map; the attention weight of each feature channel in each seventh feature map is multiplied channel-by-channel with the corresponding fifth feature map to obtain the eighth feature map corresponding to each fifth feature map; a spatial attention weight map for each seventh feature map is generated based on the edge attention mechanism, and the spatial attention weight map of each seventh feature map is multiplied element-by-element and residual-connected with the corresponding fifth feature map to obtain the ninth feature map corresponding to each fifth feature map; the eighth and ninth feature maps corresponding to each fifth feature map are weighted and fused based on preset gating parameters to obtain the tenth feature map corresponding to each fifth feature map; the tenth feature map corresponding to each fifth feature map is used as the new fifth feature map to update each fifth feature map in the fifth feature map list.

[0022] This embodiment introduces a new resolution scale to the fifth feature map list by adding a fifth feature map to the list in each iteration. Through continuous dilated convolution, normalization, and ReLU activation, the receptive field is explicitly expanded, and the ability to distinguish targets of different sizes and shapes is improved, resulting in a more accurate and comprehensive sixth feature map. Then, a channel attention mechanism automatically learns and calibrates the importance weights of each feature channel, thereby strengthening the response of the eighth feature map to landslide-discriminating feature channels and suppressing interference from irrelevant or noisy channels in the eighth feature map. An edge attention mechanism is used for edge perception, allowing the ninth feature map to better highlight the edge information of the image. Finally, a weighted fusion of the eighth and ninth feature maps is achieved through gating parameters, enabling the tenth feature map to synergistically fuse channel attention and spatial attention dedicated to edge perception. This avoids the problem of local feature loss caused by sparse sampling in traditional techniques, improving the accuracy and comprehensiveness of the fifth feature map update, and thus enhancing the accuracy of landslide identification.

[0023] In one alternative embodiment, please refer to Figure 3 , Figure 3 This is a schematic diagram of a multi-scale convolutional hybrid attention block provided in an embodiment of the present invention; this embodiment implements steps S1021 to S1024 through HRNet branches; as shown in Table 1, Table 1 is a schematic table of the HRNet branch architecture provided in an embodiment of the present invention; HRNet includes 4 stages (i.e., Stage 1, Stage 2, Stage 3, Stage 4), wherein Stage 1 uses the standard ResNet bottleneck block (ResBlock); Stage 2, Stage 3, and Stage 4 all use... Figure 3 The multi-scale dilated attention block (MDA block) shown is used to implement steps S10241 to S10247. That is, in this embodiment, the multi-scale dilated attention block replaces the standard BasicBlock in stages 2, 3, and 4 of the HRNet encoder.

[0024] Table 1. HRnet Branch Architecture Diagram Furthermore, the residual block is preset to ResBlock. Features are extracted from the initial feature map using the residual block to determine the fifth feature map, completing Stage 1. Simultaneously, the fifth feature map list is initialized using the fifth feature map; at this point, the fifth feature map list only contains the fifth feature map at a resolution scale of H / 4×W / 4×C1. Then, the number of multi-resolution feature extractions (3 times, i.e., Stage 2, Stage 3, and Stage 4) and the corresponding resolution scale for each multi-resolution feature extraction are determined by the number of resolution scales (4 in this embodiment). (Stage 2 corresponds to H / 8×W / 8×C2, Stage 3 corresponds to H / 16×W / 16×C3, and Stage 4 corresponds to H / 32×W / 32×C4). Then, through... Figure 3 The multi-scale convolutional hybrid attention block shown implements operations on Stage 2, Stage 3, and Stage 4. First, in each iteration of multi-resolution feature extraction, the fifth feature map is downsampled based on the resolution scale corresponding to the current multi-resolution feature extraction to form the fifth feature map at the resolution scale corresponding to the current multi-resolution feature extraction. Then, the fifth feature map is added to the fifth feature map list, thereby updating the fifth feature map list. Taking Stage 2 as an example, a fifth feature map with a resolution scale of H / 8×W / 8×C2 is generated at this time, thereby updating the fifth feature map list. like Figure 3As shown, due to the limited receptive field of the standard convolutional layers in the original residual blocks of HRNet, it is difficult to fully capture the contextual information required for objects with significant scale differences. Although the multi-resolution parallel branches of HRNet alleviate this problem to some extent, the contextual modeling and information extraction capabilities within each branch are equally crucial. Therefore, a serial multi-dilation convolution module (MDCM) is introduced. Dilated convolution is widely used to increase the receptive field without increasing the size of the convolutional kernel or the computational cost. The serial multi-dilation convolution module consists of three consecutive 3×3 dilated convolutional layers (Conv3×3), with the dilation rate increasing progressively (d=1, d=2, d=3). Specifically, steps S10242 and S10243 are as follows: The fifth feature map is input into a dilated convolutional layer with a dilation rate of 1, which focuses on extracting local detail features; then it undergoes normalization (Brom) and ReLU activation; then it is input into a dilated convolutional layer with a dilation rate of 2, whose input is already the feature extracted by the previous layer and rich in local information, thus further incorporating medium-range contextual information on the basis of local information; then it undergoes normalization and ReLU activation; then it is input into a dilated convolutional layer with a dilation rate of 3, which can capture global, semantic-level contextual information on the basis of the features constructed by the first two layers, and finally, after normalization and ReLU activation, the sixth feature map is obtained; the fifth feature map undergoes a progressive receptive field expansion mechanism of hierarchical abstraction and enrichment process of local texture extraction, medium-range structure extraction, and global context extraction, ensuring that the network can efficiently and adaptively capture multi-scale information; at the same time, in order to ensure the stability of training, normalization and ReLU activation functions are set. Then, by performing an identity residual connection between the fifth and sixth feature maps, the seventh feature map corresponding to each fifth feature map is determined. This not only alleviates the gradient vanishing problem in deep networks but also effectively preserves spatial detail information from the original input, ensuring that while introducing powerful contextual modeling capabilities, the low-level features crucial for dense prediction tasks are not lost. Steps S10244 to S10246 are then implemented using a gated fusion attention (GFA) mechanism. Specifically, the GFA mechanism includes an SE attention mechanism (channel attention mechanism) and an Edge attention mechanism (edge ​​attention mechanism). The core purpose of the SE attention mechanism is to adaptively recalibrate the importance weights of each feature channel. First, for the input seventh feature map... pass Perform global average pooling, and then pool the values ​​of each channel. Spatial information is compressed into a scalar value to obtain a channel descriptor. The formula for this operation is: ; The seventh feature map One channel; For channel descriptors; then Through incentive operations The generation of channel dependencies, i.e., the attention weights for each feature channel in the seventh feature map, is achieved through a bottleneck structure consisting of two fully connected layers and a sigmoid activation function. The formula is as follows: ; and These are the weights of two fully connected layers. It is the Sigmoid activation function. The ReLU activation function is used; then the attention weights for each feature channel in the seventh feature map are applied. With the corresponding fifth feature map Perform channel-by-channel multiplication to obtain the eighth feature map corresponding to each fifth feature map. This operation is performed through a function. The ultimate effect is to enhance the response of task-related channels, suppress the response of irrelevant or noisy channels, and achieve adaptive recalibration of channel features. Next, regarding the Edge Attention mechanism, it is explicitly designed to enhance high-frequency edge and boundary information in the feature map. First, a preliminary edge response map is generated through a 3×3 convolutional layer with one output channel. edge response map Spatial attention weight map is generated by normalization using the Sigmoid activation function. In this spatial attention weight map, regions with values ​​close to 1 correspond to strong edges or boundaries, while regions with values ​​close to 0 correspond to flat regions. Then, by element-wise multiplying the spatial attention weight map of each seventh feature map with the corresponding fifth feature map and concatenating the residuals with the fifth feature map, the ninth feature map corresponding to the fifth feature map is obtained. , This operation ensures that edge information is enhanced without losing the original complete features, and also avoids feature destruction caused by overemphasizing edges. Finally, the gating parameters can be set to be trainable, which is... The tenth feature map is obtained by weighting and fusing the eighth and ninth feature maps corresponding to each fifth feature map using gating parameters. The implementation process is as follows: Finally, the tenth feature map corresponding to each fifth feature map is used as the new fifth feature map to update each fifth feature map in the fifth feature map list. Ultimately, the first feature map list is... Their resolution scales correspond to H / 4×W / 4×C1, H / 8×W / 8×C2, H / 16×W / 16×C3 and H / 32×W / 32×C4, respectively.

[0025] Please refer to Figure 4 , Figure 4 This is a comparative schematic diagram of multi-resolution feature extraction provided in an embodiment of the present invention; as shown below. Figure 4 As shown, Figure 4 (a) is a schematic diagram of a remote sensing image provided in an embodiment of the present invention. Figure 4 (b) A schematic diagram of a truth label for a remote sensing image provided in an embodiment of the present invention; Figure 4 (c) A schematic diagram of a conventional feature extraction method provided in an embodiment of the present invention; Figure 4 (d) is a schematic diagram of the multi-scale convolutional hybrid attention block output provided in an embodiment of the present invention; Figure 4 (e) is a schematic diagram of gated fusion attention output provided in an embodiment of the present invention; wherein, Figure 4 (b) The truth label diagram illustrates the correct feature extraction from the remote sensing image; while the comparison Figure 4 (c) Figure 4 (d) and Figure 4 (e) It can be concluded that the multi-scale convolutional hybrid attention block and gated fusion attention provided in this embodiment can extract more accurate features compared with traditional techniques.

[0026] In this embodiment, global contextual features are extracted from the initial feature map based on a preset state-space algorithm to obtain the second feature map of the remote sensing image at each resolution scale, thereby determining the list of second feature maps of the remote sensing image. This includes: determining the number of global contextual feature extractions based on the preset resolution scale; determining the eleventh feature map based on the initial feature map; iterating the eleventh feature map based on the preset state-space algorithm until the number of global contextual feature extractions is reached, completing the global contextual feature extraction of the eleventh feature map, and determining the global contextual feature extraction result; in each iteration, obtaining the eleventh feature map of the current iteration, performing global contextual feature extraction on the initial feature map of the current iteration, determining the second feature map at the resolution scale corresponding to the current iteration; and using the second feature map at the resolution scale corresponding to the current iteration as the eleventh feature map of the next iteration; based on the global contextual feature extraction result, obtaining the second feature map of the remote sensing image at each resolution scale, thereby determining the list of second feature maps of the remote sensing image.

[0027] This embodiment introduces a state-space algorithm to iterate the eleventh feature map, which not only enables efficient extraction of global context features, but also obtains a second feature map at each resolution scale through iteration. This further enhances the correlation and comprehensiveness of the extraction of global context features in the second feature maps at different resolution scales, thereby improving the accuracy and comprehensiveness of landslide identification.

[0028] In this embodiment, in each iteration, the eleventh feature map of the current iteration is obtained, global context feature extraction is performed on the initial feature map of the current iteration to determine the second feature map at the resolution scale corresponding to the current iteration; and the second feature map at the resolution scale corresponding to the current iteration is used as the eleventh feature map of the next iteration, including: in each iteration, obtaining the eleventh feature map of the current iteration; performing layer normalization on the eleventh feature map to determine the twelfth feature map corresponding to the eleventh feature map; performing linear transformation and channel segmentation on the twelfth feature map to generate the thirteenth and fourteenth feature maps corresponding to the twelfth feature map; performing linear transformation, depthwise separable convolution and SiLu activation on the thirteenth feature map in sequence to obtain the fifteenth feature map; based on a preset four-way... The scanning strategy serializes the fifteenth feature map to determine the corresponding feature sequence, and calculates the feature sequence based on a preset parallel scanning algorithm to determine the sixteenth feature map; it performs linear transformation and SiLu activation on the fourteenth feature map to obtain the seventeenth feature map; it performs element-wise multiplication and linear transformation on the sixteenth and seventeenth feature maps to obtain the eighteenth feature map; it performs SiLu activation and regularization on the eighteenth feature map to obtain the nineteenth feature map; it performs residual concatenation between the nineteenth and eleventh feature maps to obtain the twentieth feature map; it uses the twentieth feature map as the second feature map at the resolution scale corresponding to the current iteration; and it uses the second feature map at the resolution scale corresponding to the current iteration as the eleventh feature map for the next iteration.

[0029] This embodiment obtains two feature maps, the thirteenth and fourteenth, for different processing methods through layer normalization, linear transformation, and channel segmentation. In one processing method, linear transformation, depthwise separable convolution, SiLu activation, and layer normalization of the thirteenth feature map provide spatially detailed input for subsequent four-way scanning, ensuring that these key local clues are not lost during global context feature extraction. Then, the fifteenth feature map is serialized using a four-way scanning strategy, effectively capturing long-range dependencies from multiple directions and integrating information from different spatial locations, improving the perception of global context features. The parallel scanning algorithm for feature sequence computation not only avoids the sequential computation bottleneck of traditional recurrent neural networks such as RNNs but also maintains linear complexity, allowing the sixteenth feature map to focus on long-range context information related to the current landslide pixel while ignoring irrelevant background noise. In another processing method, linear transformation and SiLu activation allow the seventeenth feature map to retain and slightly transform the original feature information. Next, the sixteenth and seventeenth feature maps with two different processing methods are multiplied element-wise and linearly transformed. Then, through SiLu activation, regularization and residual connection, the twentieth feature map can contain rich global context features, avoiding the problems of incomplete segmentation and missed detection caused by local observation in traditional technology, thereby improving the accuracy of landslide identification.

[0030] In an alternative embodiment, Figure 5 A schematic diagram of a visual basic block provided in an embodiment of the present invention; Figure 6 A schematic diagram of a selective scanning module provided in an embodiment of the present invention; Figure 7 This embodiment of the invention provides a schematic diagram of parallel scanning computation using a four-directional scanning method. To compensate for the shortcomings of the HRNet branch in long-range global dependency modeling, this embodiment constructs a parallel Mamba branch. This branch receives the same input as the HRNet backbone and aims to build a multi-scale feature pyramid rich in global context information. Considering the characteristics of landslides, such as variable morphology, blurred boundaries, and complex relationships with the background environment, this embodiment introduces a VisionSelective State Block (VSSBlock) based on the Selective State Space Model (Mamba) as the core building block. This module injects local inductive biases through deep convolution to capture texture details, employs a four-directional scanning strategy to eliminate directional deviations, and fully models the long-range contextual dependencies of landslides with arbitrary orientations. Its core selective mechanism dynamically focuses on key information related to landslide identification. Finally, features are fused through gating units, significantly improving the segmentation accuracy and robustness of the model in complex scenes. The formula for VSSBlock is: Furthermore, based on the resolution scale, the number of global context feature extraction operations is determined to be 3; that is, the second feature maps at resolution scales of H / 8×W / 8×C2, H / 16×W / 16×C3, and H / 32×W / 32×C4 are finally extracted; these correspond to the three visual base blocks in the Mamba branch proposed in this embodiment; wherein, the first visual base block is used to receive the eleventh feature map and output the second feature map at the H / 8×W / 8×C2 resolution scale; the input of the second visual base block is the output of the first visual base block, and its output is the second feature map at the H / 16×W / 16×C3 resolution scale; the input of the third visual base block is the output of the second visual base block, and its output is the second feature map at the H / 32×W / 32×C4 resolution scale. Then, the initial feature map is used as the eleventh feature map, as follows... Figures 5 to 7 As shown, in each visual base block (i.e., in each iteration of the state-space algorithm on the eleventh feature map), the eleventh feature map of the current iteration is obtained; then, layer normalization (NormLayer) is performed on the eleventh feature map to determine the twelfth feature map corresponding to the eleventh feature map; then, the selective scan module (SS2D) is used to process the twelfth feature map; specifically, in the selective scan module (SS2D), the twelfth feature map is first linearly transformed and channel segmented to generate the thirteenth feature map corresponding to the twelfth feature map (i.e., , used for Figure 6 The first branch in the graph) and the fourteenth feature map (i.e. , used for Figure 6 The second branch in the formula is implemented as shown in the formula. ;for Figure 6 Regarding the first branch, its implementation process is similar to the formula. As shown; specifically, the thirteenth feature map is sequentially subjected to a linear transformation, depthwise separable convolution (DWConv), and SiLu activation to obtain the fifteenth feature map. Then, as... Figure 7As shown, a four-way scanning strategy is used to serialize the fifteenth feature map, determining the corresponding feature sequence. Then, the parallel scanning algorithm of the SSM module (i.e., the S6 module) is used to calculate the feature sequence, determining the sixteenth feature map. Since the landslide morphology is extremely irregular and direction-independent, this embodiment employs the core selective scanning mechanism of the Mamba architecture, namely the four-way scanning strategy (including horizontal and vertical bidirectional scanning). This unfolds the two-dimensional fifteenth feature map into a sequence along the spatial dimension for operation, effectively capturing long-range dependencies from multiple directions and integrating information from different spatial locations, improving the model's ability to perceive global information. The feature sequence is then calculated using the parallel scanning algorithm of the SSM module (i.e., the S6 module). The SSM module (i.e., the S6 module) originates from an existing Mamba paper and represents a variant of the structured state-space model. Its core lies in achieving selectivity through input-dependent parameters. Unlike the fixed system matrix in traditional state-space models, the parameters in the S6 module are dynamically calculated and generated from the input sequence. This allows the model to dynamically decide whether to retain or ignore historical information based on the current input, thereby focusing on the long-range context most relevant to the current pixel. The S6 module employs a parallel scanning algorithm to efficiently compute the entire sequence. This algorithm includes forward horizontal scanning, reverse horizontal scanning, forward vertical scanning, and reverse vertical scanning; avoiding the sequential computation bottleneck of traditional recurrent neural networks such as RNNs, achieving Transformer-like parallel training efficiency while maintaining linear complexity. In this embodiment, the S6 module enables the model to dynamically focus on long-range context information related to the current landslide pixel while ignoring irrelevant background noise, thus achieving accurate identification of landslides in complex terrain environments and generating the final output sixteenth feature map through a merging operation. Figure 6 Regarding the second branch, its implementation process is similar to the formula. As shown; specifically, the fourteenth feature map is linearly transformed and SiLu activated to obtain the seventeenth feature map.

[0031] Then as Figure 6As shown, the sixteenth feature map obtained from the first branch and the seventeenth feature map obtained from the second branch are multiplied element-wise and linearly transformed to obtain the eighteenth feature map. Then, the eighteenth feature map is subjected to SiLu activation and dropout to obtain the nineteenth feature map. Finally, the nineteenth feature map is residually concatenated with the eleventh feature map to obtain the twentieth feature map. At this point, the twentieth feature map is the output of the current visual feature block and also the input of the next visual feature block; simultaneously, it is also the second feature map at the resolution scale corresponding to the current iteration. Therefore, through global context feature extraction of three visual feature blocks (i.e., the state-space algorithm iterating over the eleventh feature map), a list of second feature maps can be obtained as follows: Their resolution scales correspond to H / 8×W / 8×C2, H / 16×W / 16×C3, and H / 32×W / 32×C4, respectively.

[0032] In this embodiment, based on a preset bidirectional cross-attention fusion algorithm and resolution scale, feature map fusion is performed on the first feature map list and the second feature map list to determine the third feature map list of the remote sensing image at each resolution scale. This includes: pairing the first feature map of the first feature map list and the second feature map of the second feature map list based on the resolution scale to determine feature map pairs at each resolution scale; wherein each feature map pair includes: a first feature map and a second feature map; in the feature map pair at each resolution scale, the first feature map is divided into a first query tensor, a first key tensor, and a first value tensor; the second feature map is divided into a second query tensor, a second key tensor, and a second value tensor; the first query tensor and the second key tensor are multiplied to obtain a first tensor, and based on So... The ftmax function compresses the first tensor to determine the second tensor; the second tensor is multiplied by the first value tensor to obtain the third tensor; the third tensor is residually concatenated with the first feature map to determine the twenty-first feature map; the second query tensor is multiplied by the first key tensor to obtain the fourth tensor, and the fourth tensor is compressed using the Softmax function to determine the fifth tensor; the fifth tensor is multiplied by the second value tensor to obtain the sixth tensor; the sixth tensor is residually concatenated with the second feature map to determine the twenty-second feature map; the twenty-first and twenty-second feature maps are fused using the sigmoid function to determine the third feature map at each resolution scale; based on the third feature maps at each resolution scale, a list of third feature maps at each resolution scale is determined.

[0033] In one alternative embodiment, please refer to Figure 8 , Figure 8 This is a schematic diagram of a bidirectional cross-attention fusion module provided in an embodiment of the present invention; it is used to implement step S104. Specifically, the first feature map list is as follows: Their resolution scales correspond to H / 4×W / 4×C1, H / 8×W / 8×C2, H / 16×W / 16×C3, and H / 32×W / 32×C4, respectively. The second feature map list is as follows: Their resolution scales correspond to H / 8×W / 8×C2, H / 16×W / 16×C3, and H / 32×W / 32×C4, respectively. Therefore, to fully utilize the rich local detail features extracted by the HRnet branch and the powerful global context features captured by the Mamba branch, this embodiment implements a bidirectional cross-attention fusion algorithm through a bidirectional cross-attention fusion module, thereby obtaining a third feature map list at each resolution scale. Specifically, since the first feature map in the first feature map list... Its resolution scale is H / 8×W / 8×C2, and there is no corresponding resolution scale in the second feature map list. Therefore, this embodiment only performs bidirectional cross-attention fusion on the latter three resolution scales; for example Figure 8 As shown, firstly, the first feature maps of the first feature map list and the second feature maps of the second feature map list, which have the same resolution scale, are paired to generate feature map pairs at different resolution scales. Then, for each feature map pair, based on the currently common attention mechanism, the first feature map is divided into a first query tensor. First key tensor and the first value tensor The second feature map is divided into the second query tensor. Second key tensor Second-valued tensor ; after that, as Figure 8 As shown in the "Hrnet branch", the first query tensor With the second bond tensor Multiply the two tensors to obtain the first tensor, then compress the first tensor using the Softmax function to determine the second tensor; then combine the second tensor with the first tensor. Multiplying them yields a third tensor; then, residual concatenating the third tensor with the first feature map determines the twenty-first feature map; this process can... Seen as towards Based on the rich spatial details I possess, I inquire which global contextual information is most relevant to them, and and By providing a complete set of long-range dependencies and global semantic representations, it selectively incorporates the most relevant global context from Mamba branches, reducing false positives in locally similar but globally unrelated regions. And like... Figure 8 As shown in the "Mamba branch", the second query tensor With the first key tensor Multiply the results to obtain the fourth tensor, then compress the fourth tensor using the Softmax function to determine the fifth tensor; finally, combine the fifth tensor with the second-value tensor. Multiplying them yields the sixth tensor; then, residual concatenating the sixth tensor with the second feature map determines the twenty-second feature map; this process can... Seen as towards Based on the rich spatial details I possess, I inquire which global contextual information is most relevant to them, and and By providing a comprehensive set of long-range dependencies and global semantic representations, the most relevant global context from the Mamba branch is selectively incorporated, reducing false positives in locally similar but globally irrelevant regions. Finally, the 21st and 22nd feature maps are fused using the Sigmoid function to determine the third feature map at each resolution scale. Therefore, the list of third feature maps can be obtained as follows: The corresponding resolution scales are H / 8×W / 8×C2, H / 16×W / 16×C3 and H / 32×W / 32×C4, respectively.

[0034] This embodiment employs a bidirectional cross-attention fusion algorithm to fuse a first feature map rich in local details and a second feature map with strong global context features. This results in a third feature map possessing both rich local details and strong global context features at each resolution scale. Specifically, in one direction, the first query tensor is multiplied by the second key tensor to obtain the first tensor, which is then compressed using the Softmax function to obtain the second tensor. Further multiplication and residual connection operations inject local details into their corresponding global context features. In the other direction, the second query tensor is multiplied by the first key tensor, and then the Softmax function, multiplication, and residual connection operations are used to inject global context features into their corresponding local details. Finally, the twenty-first and twenty-second feature maps from both directions are fused, achieving bidirectional cross-attention fusion. This results in a third feature map possessing both rich local details and strong global context features, thereby improving the accuracy and comprehensiveness of landslide identification.

[0035] In this embodiment, based on a preset bidirectional calibration fusion algorithm and resolution scale, bidirectional calibration fusion is performed sequentially on every two adjacent third feature maps in the third feature map list to update every two adjacent third feature maps in the third feature map list, thereby obtaining a fourth feature map list of the remote sensing image. This includes: based on the resolution scale, performing feature alignment and stitching on every two adjacent third feature maps in the third feature map list to generate initial joint features corresponding to each pair of third feature maps; performing convolution, batch normalization, and SiLu activation operations on the initial joint features to obtain a first joint feature corresponding to the initial joint features; and performing convolution operations on the first joint features. The process involves convolution and batch normalization to obtain the second joint feature corresponding to the initial joint feature; convolution and batch normalization are then applied to the initial joint feature to obtain the third joint feature corresponding to the initial joint feature; residual concatenation and SiLu activation are performed on the second and third joint features to obtain the fourth joint feature corresponding to the initial joint feature; the fourth joint feature is then concatenated with two third feature maps from the initial joint feature to update the two third feature maps in the initial joint feature; after updating every two adjacent third feature maps in the third feature map list, the fourth feature map list of the remote sensing image is obtained based on the updated third feature map list.

[0036] In an alternative embodiment, Figure 9 This is a schematic diagram of a bidirectional calibration fusion module provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of a residual block branch provided in an embodiment of the present invention; firstly, by adjusting the resolution scale, every two adjacent third feature maps in the third feature map list are feature-aligned and stitched together to generate initial joint features corresponding to every two third feature maps; specifically, the third feature map list... In and Perform feature alignment and concatenation to generate the corresponding initial joint features; list the third feature maps. In and Feature alignment and concatenation are performed to generate corresponding initial joint features. For ease of subsequent representation, this embodiment defines the high-resolution scale third feature map as follows for every two adjacent third feature maps. The third feature map at the low-resolution scale is After that Downsampling to bilinear interpolation The spatial dimensions are obtained. To achieve feature alignment, then as follows Figure 9 As shown, and The features are concatenated along the channel dimension to form an initial joint feature that incorporates multi-resolution scale context. Both branches take as input the concatenated joint features, which simultaneously contain high-resolution spatial details and low-resolution semantic information. The goal of each branch is to filter out complementary information useful for calibrating the target features from this mixed information. The initial joint features are then fed into... Figure 10 The residual block (RB) branch shown consists of an uplink branch and a downlink branch. The downlink branch learns the detail information offset Δdown from the high-resolution features to correct the boundary blurring and detail loss of the low-resolution features, while the uplink branch learns the semantic information offset Δup from the low-resolution features to correct the semantic discontinuity of the high-resolution features. The workflow of the uplink and downlink branches is consistent: In the residual block branch, the initial joint features are convolutionally processed (3×3 Conv), batch normalized (BatchMorm), and SiLu activated to obtain the first joint feature corresponding to the initial joint features; then, the first joint feature is convolutionally processed and batch normalized to obtain the second joint feature corresponding to the initial joint features; then, the initial joint features are convolutionally processed and batch normalized to obtain the third joint feature corresponding to the initial joint features; finally, the second and third joint features are residually concatenated and SiLu activated to obtain the fourth joint feature corresponding to the initial joint features, which is the fourth joint feature output by the residual block branch; then, the fourth joint feature output by the uplink branch is combined with the third feature map at a high-resolution scale. By splicing the images together, a new feature map is obtained. ; New feature map As a new third feature map This updates the third feature map at a high-resolution scale. Simultaneously, the fourth joint feature output from the downlink branch is combined with the third feature map at a low-resolution scale. By splicing the images together, a new feature map is obtained. ; New feature map As a new third feature map This updates the third feature map at a high-resolution scale. After updating every two adjacent third feature maps in the third feature map list based on the above steps, a fourth feature map list of the remote sensing image is obtained based on the updated third feature map list; the fourth feature map list is represented as follows. The corresponding resolution scales are H / 8×W / 8×C2, H / 16×W / 16×C3 and H / 32×W / 32×C4, respectively.

[0037] This embodiment employs a bidirectional calibration fusion algorithm to perform bidirectional calibration fusion on every two adjacent third feature maps in the third feature map list, resulting in a fourth feature map list that can fuse both spatial details at high-resolution scales and strong semantic features at low-resolution scales. Specifically, feature alignment and concatenation yield an initial joint feature containing high-resolution spatial details and low-resolution semantic information. Then, convolution, batch normalization, and SiLu activation operations on the initial joint feature ensure that the fourth joint feature contains information on how features at the current two resolution scales should complement and adjust each other. Finally, a concatenation operation ensures that the updated two third feature maps not only correct the boundary blurring and detail loss issues of low-resolution features but also the semantic discontinuity issues of high-resolution features. This avoids the loss of crucial landslide information and the degradation of segmentation performance, improving the accuracy and comprehensiveness of subsequent landslide identification.

[0038] In this embodiment, image segmentation is performed on the fourth feature map list based on the attention mechanism and the object context representation mechanism to determine the landslide identification result of the mountain. This includes: performing bilinear upsampling, concatenation, normalization, and ReLU activation on the fourth feature map in the fourth feature map list to obtain the twenty-third feature map; extracting features from the twenty-third feature map based on the attention mechanism to obtain the auxiliary segmentation prediction map corresponding to the twenty-third feature map; performing a softmax operation on the auxiliary segmentation prediction map to obtain the object context feature map corresponding to the twenty-third feature map; adding the object context feature map and the auxiliary segmentation prediction map to obtain the landslide identification feature map; and performing a classification convolution on the landslide identification feature map to obtain the landslide identification result of the mountain.

[0039] In one alternative embodiment, please refer to Figure 11 , Figure 11 This is a schematic diagram of an object context representation module based on SimAM enhancement provided in an embodiment of the present invention; as shown below. Figure 11 As shown, it introduces the SimAM Attention Mechanism (SimAM) and object context representation to implement step S106. First, as... Figure 11As shown, for the fourth feature map in the fourth feature map list, bilinear sampling is first performed to convert all resolution scales to the highest resolution scale. Then, through stitching, normalization, and ReLU activation, the twenty-third feature map is obtained. Next, the SimAM attention mechanism is used to extract features from the twenty-third feature map. Specifically, the SimAM attention mechanism adaptively enhances salient features and suppresses redundant information using energy function theory, generating a coarse auxiliary segmentation prediction map aux. Then, a Softmax operation is performed on the auxiliary segmentation prediction map aux to obtain the probability distribution of each pixel belonging to each category, thus obtaining soft object regions, i.e., the object context feature map. Soft object regions define the probability distribution of different potential object regions. Furthermore, the core idea of ​​the SimAM attention mechanism originates from energy function theory, which evaluates the importance of each spatial location in the feature map through a unified metric that requires no learnable parameters. For a given feature map, the mean of the feature map in each channel is calculated. and variance According to the energy function theory, each position is defined as follows: Energy function This function measures the linear separability of a given location from other locations. It satisfies the formula: The lower the energy, the more pronounced the landslide characteristics of the pixel, and the more attention that pixel should receive. It is a tiny constant, defaulting to 1× This is used for numerical stability. The above formula means that the lower the energy, the better the stability of the neuron. The greater the difference from surrounding neurons, the higher the importance. The energy function is inverted and normalized using the Sigmoid activation function and compared with the original image. Perform element-wise multiplication to obtain the attention weight map. In this operation, the attention weights act as a spatial selection filter: for salient regions with high weights, the original feature values ​​are preserved or even enhanced; for insignificant regions with low weights, the feature values ​​are suppressed. Therefore, the attention weight map itself is a direct representation of the salient regions of the landslide, and the aforementioned multiplication operation achieves targeted enhancement of these salient regions. and The probability features of each category are calculated to obtain an auxiliary segmentation prediction map. Then, after... Perform auxiliary segmentation prediction on the map Operation, then through The object-contextual (OC) representation of each object region is calculated, where Indicates pixel position, Indicates category, Indicates the location in the auxiliary segmentation prediction map This belongs to the category The original prediction score, Indicates the process After normalization, the position in the auxiliary segmentation prediction map Category The probability of the object context is calculated. This representation aggregates the features of all pixels belonging to the same semantic region, capturing the global context information of each category and effectively representing the typical feature patterns of that category. Next, the object context representation is broadcast back to the original spatial dimensions to generate a spatially consistent object context feature map (Context). Then, the object context feature map (Context) is added to the auxiliary segmentation prediction map (aux) to achieve deep fusion of context information and original features. Finally, a classification convolutional layer (CLS head) is used to generate a refined final prediction map, which is the landslide identification result for the mountain.

[0040] This embodiment obtains a 23rd feature map by performing bilinear upsampling, convolution, and concatenation operations on the fourth feature map, restoring it to its original resolution and fusing multi-scale information, while enriching it with details and semantics. Next, an attention mechanism is used to extract features from the 23rd feature map, resulting in an auxiliary segmentation prediction map that focuses on the feature regions most relevant to landslides. Then, a softmax operation is used to introduce an object context representation mechanism, obtaining the corresponding object context feature map for the 23rd feature map. This allows the object context feature map to encode the global distribution prior and inter-regional relationships of different semantic categories in the image. Finally, by adding the object context feature map to the auxiliary segmentation prediction map and using classification convolution, the landslide recognition result is obtained, significantly improving the accuracy and spatial integrity of the landslide recognition result and avoiding the false detection and false negative problems caused by ignoring the global context in traditional methods.

[0041] In an optional embodiment, to verify the effectiveness of the bidirectional cross-attention calibration fusion landslide identification method of this embodiment, i.e., to validate the performance of the two-branch collaborative landslide identification model (hereinafter referred to as DMHRnet), this embodiment provides two datasets. The first dataset is the Bijie Landslide Dataset, a public benchmark widely used in landslide detection and segmentation tasks. This dataset uses 0.8-meter high-resolution optical images taken by the TripleSat satellite, covering a large number of landslide events. The dataset contains 770 landslide samples and 2003 non-landslide samples, and its landslide areas are usually large and have obvious features, providing high-quality labeled data for the model to learn typical landslide morphologies. The second dataset is the GVLM dataset, which is a large-scale, open-source, and globally covered ultra-high-resolution landslide mapping dataset, mainly used to test the generalization performance of the model. This dataset contains 17 subsets from different geographical environments, with a total coverage area of ​​163.77 square kilometers. Each subset contains a pair of two-phase images with a spatial resolution of 0.59 meters and corresponding pixel-level labeled masks. Its global sample distribution provides a rigorous benchmark for validating the model's adaptability in unknown regions. Three evaluation metrics were then established: precision, recall, and intersection over union (IoU). These metrics measure model performance from different dimensions, making them particularly suitable for tasks like landslide detection, which exhibit significant class imbalance and require precise boundary localization. The definitions of these metrics are as follows: 、 and True positives (TP) are landslide pixels correctly identified by the model, false positives (FP) are non-landslide pixels incorrectly classified as landslide pixels, true negatives (TN) represent correctly identified non-landslide pixels, and false negatives (FN) represent missed landslide pixels. Precision reflects the model's ability to minimize FP, thus verifying the model's reliability; while recall reflects the model's ability to minimize FN, reflecting the model's recall capability. IOU measures the overlap between the predicted output and the ground truth (GT), and is a core indicator for semantic segmentation tasks, sensitively reflecting the accuracy of boundary segmentation. Furthermore, based on the above dataset and evaluation metrics, please refer to Table 2, which is a performance comparison table of different landslide identification methods provided in the embodiments of this invention. As shown in Table 2, "DMHRnet" represents the landslide recognition method with bidirectional cross-attention calibration fusion provided in this embodiment; FCN represents Fully Convolutional Network; Unet represents U-shaped network; PSP-net represents Pyramid Scene Parsing Network; UpperNet represents Unified Perceptual Parsing Network; DeeplabV3+ represents DeepLab third version enhancement; HRnet represents High-Resolution Network; HRFormer represents High-Resolution Visual Transformer; ConvNeXt refers to A ConvNet for the 2020s; TransUNet refers to Hybrid Transformer-U-shaped network; HRTrans refers to a model that uses HRFormer to expand the feature receptive field and utilizes a bidirectional short connection fusion model (such as BiFormer) for multi-resolution feature aggregation; Swin-Unet refers to the structure of Swin Transformer encoder and U-Net decoder; Vmamba refers to the visual base model of the state space model.

[0042] Table 2 Performance Comparison of Different Landslide Identification Methods DMHRNet outperforms traditional convolutional networks on the Bijie dataset with 94.78% precision, 94.37% recall, and 82.08% IoU. Particularly noteworthy is IoU, a core metric for segmentation, where the proposed method achieves a 1.76 percentage point improvement over the currently best-performing ConvNeXt (80.32%), demonstrating its superior pixel-level localization accuracy. On the GVLM dataset, DMHRNet also achieves the best overall performance, with an IoU of 70.44%, validating its stable generalization ability across different geographical distributions and data characteristics. Compared to traditional convolutional networks, DMHRNet improves IoU by 7.4% and 7.57% over FCN and U-Net respectively on the Bijie dataset, fully demonstrating the effectiveness of introducing a state-space model for global context modeling. Compared to Transformer-based models, our method achieves an IoU advantage of 5.8% and 2.83% over TransUNet and Swin-UNet respectively on the GVLM dataset, while maintaining a higher recall rate. This indicates that it can more effectively avoid false negatives and alleviate the overfitting problem of the Transformer architecture on small datasets. Compared to ConvNeXt, which also uses a modern architecture, DMHRnet maintains a fairly high recall rate while consistently outperforming it in both precision and IoU on both datasets. This confirms that the Mamba-HRnet dual-branch design adopted in this invention can more balancedly preserve details and global inference capabilities compared to pure convolutional or pure Transformer architectures.

[0043] For further details, please refer to Figure 12 , Figure 12 This is a visual identification diagram illustrating different landslide identification methods provided in embodiments of the present invention; wherein, Figure 12 (a) A schematic diagram of a remote sensing image provided in an embodiment of the present invention; Figure 12 (b) A truth map of a remote sensing image provided in an embodiment of the present invention; Figure 12 (c) A landslide identification result diagram of an existing FCN model provided in an embodiment of the present invention; Figure 12 (d) A landslide identification result diagram of an existing Unet provided in an embodiment of the present invention; Figure 12 (e) A landslide identification result diagram of an existing PSP-net model provided in an embodiment of the present invention; Figure 12 (f) A landslide identification result diagram of an existing UPerNet model provided in an embodiment of the present invention; Figure 12 (g) A landslide identification result diagram of an existing DeeplabV3+ model provided in an embodiment of the present invention; Figure 12 (h) is a landslide identification result diagram of an existing HRnet model provided in an embodiment of the present invention; Figure 12 (i) A landslide identification result diagram of an existing HRFormer model provided in an embodiment of the present invention; Figure 12 (j) is a landslide identification result diagram of an existing HRTrans model provided in an embodiment of the present invention; Figure 12 (k) is a landslide identification result diagram of an existing ConvNeXt model provided in an embodiment of the present invention; Figure 12 (l) A landslide identification result diagram of an existing TransUNet model provided in an embodiment of the present invention; Figure 12 (m) is a landslide identification result diagram of an existing Vmamba model provided in an embodiment of the present invention; Figure 12 (n) is a landslide identification result diagram of an existing Swin-Unet model provided in an embodiment of the present invention; Figure 12 (o) is a landslide identification result diagram of the dual-branch collaborative landslide identification model provided in the embodiment of the present invention; as shown. Figure 12 As shown, in data (I), (II), (III), and (IV) of the Bijie dataset, and data (I), (II), (III), and (IV) of the GVLM dataset, green represents the landslide identification results. The landslide identification results of the dual-branch collaborative landslide identification model provided in this embodiment are closer to the ground truth map, which illustrates the effectiveness of the bidirectional cross-attention calibration fusion landslide identification method in this embodiment and also verifies the performance of the dual-branch collaborative landslide identification model.

[0044] Example 2 Please refer to Figure 13 , Figure 13 A schematic diagram of a landslide identification system based on bidirectional cross-attention calibration fusion provided in an embodiment of the present invention includes: an initial feature map acquisition module 201, a multi-resolution local feature extraction module 202, a global context feature extraction module 203, a bidirectional cross-attention fusion module 204, a bidirectional calibration fusion module 205, and an attention object context representation module 206. The initial feature map acquisition module 201 is used to acquire remote sensing images of the mountain, extract features from the remote sensing images, and acquire an initial feature map; the multi-resolution local feature extraction module 202 is used to perform multi-resolution local feature extraction on the initial feature map based on a preset multi-scale convolutional hybrid attention algorithm, to obtain the first feature map of the remote sensing image at each preset resolution scale, and to determine the first feature map list of the remote sensing image; the global context feature extraction module 203 is used to perform global context feature extraction on the initial feature map based on a preset state space algorithm, to obtain the second feature map of the remote sensing image at each resolution scale, and to determine the second feature map list of the remote sensing image; the bidirectional cross-attention fusion module 204 uses... Based on a preset bidirectional cross-attention fusion algorithm and resolution scale, feature map fusion is performed on the first feature map list and the second feature map list to determine the third feature map list of the remote sensing image at each resolution scale; the bidirectional calibration fusion module 205 is used to perform bidirectional calibration fusion on every two adjacent third feature maps in the third feature map list in turn, based on a preset bidirectional calibration fusion algorithm and resolution scale, to update every two adjacent third feature maps in the third feature map list, and obtain the fourth feature map list of the remote sensing image; the attention object context representation module 206 is used to perform image segmentation on the fourth feature map list based on the attention mechanism and the object context representation mechanism to determine the landslide identification result of the mountain.

[0045] In this embodiment, the multi-resolution local feature extraction module 202 includes: a multi-resolution local feature extraction unit; the multi-resolution local feature extraction unit is used to extract features from an initial feature map based on a preset residual block to determine a fifth feature map; and initialize a fifth feature map list based on the fifth feature map; determine the number of multi-resolution feature extractions and the resolution scale corresponding to each multi-resolution feature extraction based on the number of preset resolution scales; iteratively extract multi-resolution features from the fifth feature map list based on a preset multi-scale convolutional hybrid attention algorithm and the resolution scale corresponding to each multi-resolution feature extraction until the number of multi-resolution feature extractions is reached, thereby updating the fifth feature map list; based on the updated fifth feature map list, obtain the first feature map of the remote sensing image at each preset resolution scale, and then determine the first feature map list of the remote sensing image; in each iteration of multi-resolution feature extraction, based on the resolution scale corresponding to the current multi-resolution feature extraction, add a fifth feature map to the fifth feature map list, combine the preset multi-scale convolutional hybrid attention algorithm to extract features from each fifth feature map in the fifth feature map list, and update each fifth feature map in the fifth feature map list based on the feature extraction results.

[0046] In this embodiment, the multi-resolution local feature extraction unit includes: a multi-resolution local feature extraction subunit; the multi-resolution local feature extraction subunit is used to perform downsampling operation on the fifth feature map based on the resolution scale corresponding to the current multi-resolution feature extraction in each iteration of multi-resolution feature extraction, so as to form a fifth feature map at the resolution scale corresponding to the current multi-resolution feature extraction, and then update the fifth feature map list; perform continuous dilated convolution on each fifth feature map, and perform normalization and ReLU activation after each dilated convolution to determine the sixth feature map corresponding to each fifth feature map; perform identity residual connection between the fifth feature map and the sixth feature map to determine the seventh feature map corresponding to each fifth feature map; and calibrate each feature channel in each seventh feature map based on the channel attention mechanism. The process involves generating attention weights for each feature channel in each seventh feature map, multiplying the attention weights of each feature channel in each seventh feature map by the corresponding fifth feature map channel-by-channel to obtain the eighth feature map corresponding to each fifth feature map, generating a spatial attention weight map for each seventh feature map based on an edge attention mechanism, and multiplying the spatial attention weight map of each seventh feature map by the corresponding fifth feature map element-by-element and performing residual concatenation to obtain the ninth feature map corresponding to each fifth feature map, weighting and fusing the eighth and ninth feature maps corresponding to each fifth feature map based on preset gating parameters to obtain the tenth feature map corresponding to each fifth feature map, and using the tenth feature map corresponding to each fifth feature map as the new fifth feature map to update each fifth feature map in the fifth feature map list.

[0047] In this embodiment, the global context feature extraction module 203 includes: a global context feature extraction unit; the global context feature extraction unit is used to determine the number of global context feature extractions based on a preset resolution scale; determine the eleventh feature map based on the initial feature map; iterate the eleventh feature map based on a preset state space algorithm until the number of global context feature extractions is reached, complete the global context feature extraction of the eleventh feature map, and determine the global context feature extraction result; in each iteration, obtain the eleventh feature map of the current iteration, perform global context feature extraction on the initial feature map of the current iteration, determine the second feature map at the resolution scale corresponding to the current iteration; and use the second feature map at the resolution scale corresponding to the current iteration as the eleventh feature map of the next iteration; based on the global context feature extraction result, obtain the second feature map of the remote sensing image at each resolution scale, and then determine the list of second feature maps of the remote sensing image.

[0048] In this embodiment, the global context feature extraction unit includes: a global context feature extraction subunit; the global context feature extraction subunit is used to obtain the eleventh feature map of the current iteration in each iteration; perform layer normalization on the eleventh feature map to determine the twelfth feature map corresponding to the eleventh feature map; perform linear transformation and channel segmentation on the twelfth feature map to generate the thirteenth and fourteenth feature maps corresponding to the twelfth feature map; perform linear transformation, depthwise separable convolution and SiLu activation on the thirteenth feature map in sequence to obtain the fifteenth feature map; and serialize the fifteenth feature map based on a preset four-way scanning strategy to determine the feature order corresponding to the fifteenth feature map. The algorithm calculates the feature sequence based on a preset parallel scanning algorithm to determine the sixteenth feature map; it performs linear transformation and SiLu activation on the fourteenth feature map to obtain the seventeenth feature map; it performs element-wise multiplication and linear transformation on the sixteenth and seventeenth feature maps to obtain the eighteenth feature map; it performs SiLu activation and regularization on the eighteenth feature map to obtain the nineteenth feature map; it performs residual concatenation on the nineteenth and eleventh feature maps to obtain the twentieth feature map; it uses the twentieth feature map as the second feature map at the resolution scale corresponding to the current iteration; and it uses the second feature map at the resolution scale corresponding to the current iteration as the eleventh feature map for the next iteration.

[0049] In this embodiment, the bidirectional cross-attention fusion module 204 includes: a bidirectional cross-attention fusion unit; the bidirectional cross-attention fusion unit is used to pair the first feature map of the first feature map list and the second feature map of the second feature map list based on the resolution scale to determine the feature map pair at each resolution scale; wherein, each feature map pair includes: a first feature map and a second feature map; in the feature map pair at each resolution scale, the first feature map is divided into a first query tensor, a first key tensor, and a first value tensor; the second feature map is divided into a second query tensor, a second key tensor, and a second value tensor; the first query tensor and the second key tensor are multiplied to obtain a first tensor, and the first tensor is compressed based on the Softmax function. The process involves: determining the second tensor; multiplying the second tensor with the first value tensor to obtain the third tensor; performing a residual concatenation between the third tensor and the first feature map to determine the twenty-first feature map; multiplying the second query tensor with the first key tensor to obtain the fourth tensor, and compressing the fourth tensor using the Softmax function to determine the fifth tensor; multiplying the fifth tensor with the second value tensor to obtain the sixth tensor; performing a residual concatenation between the sixth tensor and the second feature map to determine the twenty-second feature map; fusing the twenty-first and twenty-second feature maps using the Sigmoid function to determine the third feature map at each resolution scale; and determining a list of third feature maps at each resolution scale based on the third feature maps at each resolution scale.

[0050] In this embodiment, the bidirectional calibration fusion module 205 includes: a bidirectional calibration fusion unit; The bidirectional calibration fusion unit is used to perform feature alignment and stitching on every two adjacent third feature maps in the third feature map list based on the resolution scale, generating an initial joint feature corresponding to each pair of third feature maps; performing convolution, batch normalization, and SiLu activation operations on the initial joint feature to obtain the first joint feature corresponding to the initial joint feature; performing convolution and batch normalization on the first joint feature to obtain the second joint feature corresponding to the initial joint feature; performing convolution and batch normalization on the initial joint feature to obtain the third joint feature corresponding to the initial joint feature; performing residual connection and SiLu activation operations on the second and third joint features to obtain the fourth joint feature corresponding to the initial joint feature; stitching the fourth joint feature with each of the two third feature maps in the initial joint feature to update the two third feature maps in the initial joint feature; after updating every two adjacent third feature maps in the third feature map list, the fourth feature map list of the remote sensing image is obtained based on the updated third feature map list.

[0051] In this embodiment, the attention object context representation module 206 includes: an attention object context representation unit; the attention object context representation unit is used to perform bilinear upsampling, concatenation, normalization and ReLU activation on the fourth feature map in the fourth feature map list to obtain the twenty-third feature map; to extract features from the twenty-third feature map based on the attention mechanism to obtain the auxiliary segmentation prediction map corresponding to the twenty-third feature map; to perform a Softmax operation on the auxiliary segmentation prediction map to obtain the object context feature map corresponding to the twenty-third feature map; to add the object context feature map and the auxiliary segmentation prediction map to obtain the landslide recognition feature map; and to perform classification convolution on the landslide recognition feature map to obtain the landslide recognition result of the mountain.

[0052] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A landslide identification method based on bidirectional cross-attention calibration fusion, characterized in that, include: Acquire remote sensing images of the mountain, extract features from the remote sensing images, and obtain an initial feature map; Based on a preset multi-scale convolutional hybrid attention algorithm, multi-resolution local feature extraction is performed on the initial feature map to obtain the first feature map of the remote sensing image at each preset resolution scale, and the first feature map list of the remote sensing image is determined. Global contextual features are extracted from the initial feature map based on a preset state space algorithm to obtain the second feature map of the remote sensing image at each resolution scale, thereby determining the list of second feature maps of the remote sensing image; Based on a preset bidirectional cross-attention fusion algorithm and the resolution scale, feature map fusion is performed on the first feature map list and the second feature map list to determine the third feature map list of the remote sensing image at each resolution scale. Based on the preset bidirectional calibration fusion algorithm and the resolution scale, bidirectional calibration fusion is performed sequentially on every two adjacent third feature maps in the third feature map list to update every two adjacent third feature maps in the third feature map list, thereby obtaining the fourth feature map list of the remote sensing image. Based on the attention mechanism and the object context representation mechanism, image segmentation is performed on the fourth feature map list to determine the landslide identification result of the mountain.

2. The landslide identification method based on bidirectional cross-attention calibration fusion as described in claim 1, characterized in that, The initial feature map is subjected to multi-resolution local feature extraction based on a preset multi-scale convolutional hybrid attention algorithm to obtain the first feature map of the remote sensing image at each preset resolution scale, and the list of the first feature maps of the remote sensing image is determined, including: Based on the preset residual block, feature extraction is performed on the initial feature map to determine the fifth feature map; and the fifth feature map list is initialized based on the fifth feature map. Based on the number of preset resolution scales, determine the number of multi-resolution feature extractions and the resolution scale corresponding to each multi-resolution feature extraction; Based on a preset multi-scale convolutional hybrid attention algorithm and the resolution scale corresponding to each multi-resolution feature extraction, the fifth feature map list is iteratively multi-resolution feature extraction is performed until the number of multi-resolution feature extractions is reached, thus completing the update of the fifth feature map list; based on the updated fifth feature map list, the first feature map of the remote sensing image at each preset resolution scale is obtained, thereby determining the first feature map list of the remote sensing image. In each iteration of multi-resolution feature extraction, based on the resolution scale corresponding to the current multi-resolution feature extraction, the fifth feature map is added to the fifth feature map list. The preset multi-scale convolutional hybrid attention algorithm is used to extract features from each of the fifth feature maps in the fifth feature map list, and the fifth feature map in the fifth feature map list is updated based on the feature extraction results.

3. The landslide identification method based on bidirectional cross-attention calibration fusion as described in claim 2, characterized in that, In each iteration of multi-resolution feature extraction, based on the resolution scale corresponding to the current multi-resolution feature extraction, a fifth feature map is added to the fifth feature map list. A preset multi-scale convolutional hybrid attention algorithm is used to extract features from each fifth feature map in the fifth feature map list. Based on the feature extraction results, each fifth feature map in the fifth feature map list is updated. This includes: In each iteration of multi-resolution feature extraction, the fifth feature map is downsampled based on the resolution scale corresponding to the current multi-resolution feature extraction to form the fifth feature map at the resolution scale corresponding to the current multi-resolution feature extraction, thereby updating the list of fifth feature maps; Perform continuous dilated convolutions on each of the fifth feature maps, and perform normalization and ReLU activation after each dilated convolution to determine the sixth feature map corresponding to each of the fifth feature maps; The fifth feature map is connected to the sixth feature map by an identity residual connection to determine the seventh feature map corresponding to each fifth feature map; Based on the channel attention mechanism, each feature channel in each of the seventh feature maps is calibrated to generate the attention weight of each feature channel in each of the seventh feature maps; and the attention weight of each feature channel in each of the seventh feature maps is multiplied channel by channel with the corresponding fifth feature map to obtain the eighth feature map corresponding to each of the fifth feature maps. A spatial attention weight map for each of the seventh feature maps is generated based on the edge attention mechanism, and the spatial attention weight map for each of the seventh feature maps is multiplied element-wise and residually connected with the corresponding fifth feature map to obtain the ninth feature map corresponding to each of the fifth feature maps. Based on preset gating parameters, the eighth and ninth feature maps corresponding to each fifth feature map are weighted and fused to obtain the tenth feature map corresponding to each fifth feature map; The tenth feature map corresponding to each of the fifth feature maps is used as the new fifth feature map to update each fifth feature map in the list of fifth feature maps.

4. The landslide identification method based on bidirectional cross-attention calibration fusion as described in claim 1, characterized in that, The process involves extracting global contextual features from the initial feature map using a preset state-space algorithm to obtain a second feature map of the remote sensing image at each resolution scale, thereby determining a list of second feature maps for the remote sensing image, including: The number of times global context features are extracted is determined based on a preset resolution scale; Based on the initial feature map, the eleventh feature map is determined; The eleventh feature map is iterated based on a preset state space algorithm until the number of global context feature extractions is reached, and the global context feature extraction of the eleventh feature map is completed, and the global context feature extraction result is determined. In each iteration, the eleventh feature map of the current iteration is obtained, global context features are extracted from the initial feature map of the current iteration, and the second feature map at the resolution scale corresponding to the current iteration is determined; and the second feature map at the resolution scale corresponding to the current iteration is used as the eleventh feature map of the next iteration. Based on the global context feature extraction results, a second feature map of the remote sensing image at each resolution scale is obtained, thereby determining a list of second feature maps of the remote sensing image.

5. The landslide identification method based on bidirectional cross-attention calibration fusion as described in claim 4, characterized in that, In each iteration, the eleventh feature map of the current iteration is obtained, global context features are extracted from the initial feature map of the current iteration, and the second feature map at the resolution scale corresponding to the current iteration is determined. The second feature map at the resolution scale corresponding to the current iteration is used as the eleventh feature map for the next iteration, including: In each iteration, obtain the eleventh feature map of the current iteration; Perform layer normalization on the eleventh feature map to determine the twelfth feature map corresponding to the eleventh feature map; The twelfth feature map is subjected to linear transformation and channel segmentation to generate the thirteenth and fourteenth feature maps corresponding to the twelfth feature map; The thirteenth feature map is sequentially subjected to linear transformation, depthwise separable convolution, and SiLu activation to obtain the fifteenth feature map; The fifteenth feature map is serialized based on a preset four-way scanning strategy to determine the feature sequence corresponding to the fifteenth feature map, and the sixteenth feature map is determined by calculating the feature sequence based on a preset parallel scanning algorithm. The fourteenth feature map is linearly transformed and SiLu activated to obtain the seventeenth feature map; The sixteenth and seventeenth feature maps are multiplied element-wise and linearly transformed to obtain the eighteenth feature map; The eighteenth feature map is subjected to SiLu activation and regularization to obtain the nineteenth feature map; The 19th feature map is residually connected to the 11th feature map to obtain the 20th feature map; The twentieth feature map is used as the second feature map at the resolution scale corresponding to the current iteration; and the second feature map at the resolution scale corresponding to the current iteration is used as the eleventh feature map for the next iteration.

6. The landslide identification method based on bidirectional cross-attention calibration fusion as described in claim 1, characterized in that, The step of fusing the first and second feature map lists based on a preset bidirectional cross-attention fusion algorithm and the resolution scale to determine a third feature map list of the remote sensing image at each resolution scale includes: Based on the resolution scale, the first feature map of the first feature map list and the second feature map of the second feature map list are paired to determine the feature map pair at each resolution scale; wherein, each feature map pair includes: a first feature map and a second feature map; In each feature map pair at the resolution scale, the first feature map is divided into a first query tensor, a first key tensor, and a first value tensor; the second feature map is divided into a second query tensor, a second key tensor, and a second value tensor. Multiply the first query tensor and the second key tensor to obtain the first tensor, and compress the first tensor based on the Softmax function to determine the second tensor; Multiply the second tensor by the first value tensor to obtain the third tensor; then perform a residual concatenation between the third tensor and the first feature map to determine the twenty-first feature map; The second query tensor is multiplied by the first key tensor to obtain the fourth tensor, and the fourth tensor is compressed based on the Softmax function to determine the fifth tensor; The fifth tensor is multiplied by the second value tensor to obtain the sixth tensor; and the sixth tensor is residually concatenated with the second feature map to determine the twenty-second feature map; The 21st and 22nd feature maps are fused based on the Sigmoid function to determine the third feature map at each resolution scale; Based on the third feature map at each resolution scale, a list of third feature maps at each resolution scale is determined.

7. The landslide identification method based on bidirectional cross-attention calibration fusion as described in claim 1, characterized in that, Based on a preset bidirectional calibration fusion algorithm and the resolution scale, bidirectional calibration fusion is sequentially performed on every two adjacent third feature maps in the third feature map list to update every two adjacent third feature maps in the third feature map list, thereby obtaining a fourth feature map list of the remote sensing image, including: Based on the resolution scale, each pair of adjacent third feature maps in the third feature map list is aligned and stitched together to generate an initial joint feature corresponding to each pair of third feature maps; The initial joint features are subjected to convolution, batch normalization and SiLu activation operations to obtain the first joint features corresponding to the initial joint features. Perform convolution and batch normalization on the first joint feature to obtain the second joint feature corresponding to the initial joint feature; Perform convolution and batch normalization on the initial joint features to obtain the third joint features corresponding to the initial joint features; The second joint feature and the third joint feature are subjected to residual connection and SiLu activation operation to obtain the fourth joint feature corresponding to the initial joint feature; The fourth joint feature is concatenated with the two third feature maps in the initial joint feature to update the two third feature maps in the initial joint feature; After updating every two adjacent third feature maps in the third feature map list, a fourth feature map list of the remote sensing image is obtained based on the updated third feature map list.

8. The landslide identification method based on bidirectional cross-attention calibration fusion as described in claim 1, characterized in that, The process of performing image segmentation on the fourth feature map list based on attention and object context representation mechanisms to determine the landslide identification result of the mountain includes: The fourth feature map in the fourth feature map list is subjected to bilinear upsampling, concatenation, normalization and ReLU activation to obtain the twenty-third feature map. Based on the attention mechanism, feature extraction is performed on the 23rd feature map to obtain the auxiliary segmentation prediction map corresponding to the 23rd feature map; Perform a Softmax operation on the auxiliary segmentation prediction map to obtain the object context feature map corresponding to the 23rd feature map; The object context feature map is added to the auxiliary segmentation prediction map to obtain the landslide identification feature map; The landslide identification feature map is subjected to classification convolution to obtain the landslide identification result of the mountain.

9. A landslide identification system based on bidirectional cross-attention calibration fusion, characterized in that, include: The module includes an initial feature map acquisition module, a multi-resolution local feature extraction module, a global context feature extraction module, a bidirectional cross-attention fusion module, a bidirectional calibration fusion module, and an attention object context representation module. The initial feature map acquisition module is used to acquire remote sensing images of the mountain, extract features from the remote sensing images, and acquire an initial feature map. The multi-resolution local feature extraction module is used to extract multi-resolution local features from the initial feature map based on a preset multi-scale convolutional hybrid attention algorithm, to obtain the first feature map of the remote sensing image at each preset resolution scale, and to determine the first feature map list of the remote sensing image; The global context feature extraction module is used to extract global context features from the initial feature map based on a preset state space algorithm to obtain the second feature map of the remote sensing image at each resolution scale, and then determine the list of second feature maps of the remote sensing image. The bidirectional cross-attention fusion module is used to perform feature map fusion on the first feature map list and the second feature map list based on a preset bidirectional cross-attention fusion algorithm and the resolution scale, to determine the third feature map list of the remote sensing image at each resolution scale. The bidirectional calibration fusion module is used to perform bidirectional calibration fusion on every two adjacent third feature maps in the third feature map list in sequence based on a preset bidirectional calibration fusion algorithm and the resolution scale, so as to update every two adjacent third feature maps in the third feature map list and obtain the fourth feature map list of the remote sensing image. The attention object context representation module is used to perform image segmentation on the fourth feature map list based on the attention mechanism and the object context representation mechanism to determine the landslide identification result of the mountain.

10. A landslide identification system based on bidirectional cross-attention calibration fusion as described in claim 9, characterized in that, The bidirectional cross-attention fusion module includes: a bidirectional cross-attention fusion unit; The bidirectional cross-attention fusion unit is used to pair the first feature map of the first feature map list and the second feature map of the second feature map list based on the resolution scale to determine the feature map pair at each resolution scale; wherein, each feature map pair includes: a first feature map and a second feature map; In each feature map pair at the resolution scale, the first feature map is divided into a first query tensor, a first key tensor, and a first value tensor; the second feature map is divided into a second query tensor, a second key tensor, and a second value tensor. Multiply the first query tensor and the second key tensor to obtain the first tensor, and compress the first tensor based on the Softmax function to determine the second tensor; Multiply the second tensor by the first value tensor to obtain the third tensor; then perform a residual concatenation between the third tensor and the first feature map to determine the twenty-first feature map; The second query tensor is multiplied by the first key tensor to obtain the fourth tensor, and the fourth tensor is compressed based on the Softmax function to determine the fifth tensor; The fifth tensor is multiplied by the second value tensor to obtain the sixth tensor; and the sixth tensor is residually concatenated with the second feature map to determine the twenty-second feature map; The 21st and 22nd feature maps are fused based on the Sigmoid function to determine the third feature map at each resolution scale; Based on the third feature map at each resolution scale, a list of third feature maps at each resolution scale is determined.

Images