Landslide identification method based on mixed attention mechanism of channel and space

Abstract

The invention relates to a landslide identification method based on a channel and space mixed attention mechanism. The method comprises the following steps: acquiring remote sensing image data, and constructing a landslide identification model comprising an encoder, an improved jump connection and feature fusion module, a mixed attention module, a decoder and an output layer. The encoder performs layer-by-layer convolution down-sampling on the remote sensing image and outputs features of each layer; the improved module carries out convolution normalization, splicing and weighted fusion on the features to obtain fusion features; in the fusion stage, a mixed attention module is embedded, EMA generates channel weights and enhances features, PPA generates space weights and enhances features, and the EMA and the PPA are spliced and subjected to convolution activation to obtain final enhanced features. After receiving, the decoder performs up-sampling fusion, and an output layer obtains a landslide segmentation result through an activation function; and finally, the model is trained by using a preset loss function, landslide identification is realized, a landslide area can be accurately extracted, and the problem of class imbalance is relieved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of landslide identification, in particular to a landslide identification method based on a channel and space hybrid attention mechanism. BACKGROUND

[0002] Landslides are a kind of serious geological disasters, which have the characteristics of strong suddenness, great destructiveness and wide influence range, and pose a significant threat to transportation, infrastructure and personnel safety. In order to realize intelligent monitoring and rapid warning of landslide disasters, it is necessary to accurately extract the landslide area from remote sensing images. However, in actual remote sensing data, the landslide area is often affected by factors such as vegetation coverage, shadow, terrain undulation and illumination condition changes, and its performance characteristics in the image are complex and fuzzy, which makes landslide segmentation a challenging task.

[0003] Traditional machine learning methods rely on manual feature design and are difficult to achieve ideal results in complex environments. In recent years, deep learning methods, especially semantic segmentation models based on convolutional neural networks (CNN), have been widely used in landslide identification. However, existing methods still have the following shortcomings, including: insufficient multi-scale feature extraction of landslide areas, leading to insufficient recognition accuracy of the model in complex terrain. The boundary between the landslide area and the background is fuzzy, and the prediction result is prone to adhesion or fracture. The proportion of landslide pixels in remote sensing images is much smaller than that of background pixels, and the class imbalance is serious, and the model training is prone to deviate to the non-landslide area. SUMMARY

[0004] Therefore, it is necessary to provide a landslide identification method based on a channel and space hybrid attention mechanism to improve the accuracy of landslide identification.

[0005] A landslide identification method based on a channel and space hybrid attention mechanism, the method comprising: acquiring remote sensing image data; constructing a landslide identification model; the landslide identification model comprises an encoder, an improved jump connection and feature fusion module, a hybrid attention module, a decoder and an output layer; the hybrid attention module is composed of an EMA module and a PPA module in parallel; using the encoder to perform layer-by-layer convolution and down-sampling operations on the remote sensing image to obtain the output features of each layer of the encoder; using the improved jump connection and feature fusion module to first perform convolution and batch normalization operations on the output features to obtain normalized features, then splicing the normalized features to obtain spliced features, and performing convolution and weighted fusion on the spliced features to obtain fusion features; In the fusion stage, a hybrid attention module is embedded. Based on the EMA module, global average pooling and multi-scale one-dimensional convolution operations are performed on the fused features to generate channel attention weights. The channel attention weights are multiplied with the fused features channel by channel to obtain the EMA enhanced features. Based on the PPA module, convolution and activation operations are performed on the fused features to generate spatial attention weights. The spatial attention weights are multiplied with the fused features element by element to obtain the PPA enhanced features. The EMA enhanced features and PPA enhanced features are concatenated along the channel dimension to obtain the concatenated attention features. The concatenated attention features are then convolved and activated to obtain the final enhanced features. The final enhanced features are received by the decoder, and the spatial resolution is restored by progressive upsampling. The enhanced features are then fused with the final enhanced features to obtain the outputs of each layer of the decoder. Based on the output of each layer of the decoder received by the output layer, the pixel-level class probability distribution is obtained through the activation function, which is the landslide segmentation result; The landslide identification model is trained according to a pre-set loss function, and the trained landslide identification model is used to identify landslides.

[0006] The aforementioned landslide identification method based on a hybrid attention mechanism of channels and space addresses the problem of insufficient multi-scale feature extraction in traditional models by using layer-by-layer convolutional downsampling in the encoder to balance shallow fine-grained features with deep global semantic features. Improved skip-connection convolutional normalization and multi-feature splicing fusion then solve the same problem. The embedded EMA-PPA parallel attention module captures channel dependence through multi-scale one-dimensional convolution, highlighting key landslide channel features, while the latter generates enhanced boundary and target area responses through spatial weights. This joint modeling of channel and spatial features significantly improves the model's ability to represent landslide areas in complex terrain, accurately capturing landslide features at different scales even under interference from vegetation cover and shadows. The hybrid attention module effectively optimizes boundary segmentation, avoiding prediction results from sticking or breaking. The fusion of the decoder and enhanced features restores spatial resolution while preserving details, and the pixel-level probability distribution of the output layer further ensures accurate classification, significantly improving the accuracy of landslide area segmentation. A loss function is set to balance the contribution of positive and negative samples, mitigating model bias caused by the low proportion of landslide pixels in remote sensing images, reducing the missed detection rate of landslide areas, and ensuring the reliability of early warnings. This application can be efficiently applied to landslide disaster monitoring and early warning, taking into account both accuracy and anti-interference capabilities, and provides strong technical support for the safety protection of transportation, infrastructure and personnel. It has broad practical value and prospects for promotion. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating a landslide identification method based on a hybrid attention mechanism of channels and space in one embodiment. Figure 2 This is an internal structural diagram of a computer device in one embodiment. DETAILED DESCRIPTION

[0008] In order to make the purposes, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0009] In one embodiment, as shown in Figure 1 A landslide identification method based on a channel and space mixed attention mechanism is provided, comprising the following steps: Step 102, acquiring remote sensing image data; constructing a landslide identification model; the landslide identification model comprises an encoder, an improved skip connection and feature fusion module, a mixed attention module, a decoder and an output layer; the mixed attention module is composed of an EMA module and a PPA module in parallel.

[0010] Step 104, using the encoder to perform layer-by-layer convolution and down-sampling operation on the remote sensing image to obtain the output features of each layer of the encoder.

[0011] The encoder is used to perform layer-by-layer convolution and down-sampling on the input remote sensing image, extract features of different scales, perform layer-by-layer convolution and down-sampling on the input remote sensing image, extract fine-grained features such as texture, boundary and color at the shallow layer, and extract global semantic information and abstract features at the deep layer. The output features of each layer are outputted. Through the hierarchical progressive manner, the model can consider both local information and global context information, and provide multi-scale feature support for subsequent segmentation. Breaking the limitation of traditional models that can only extract single-scale features, the hierarchical progressive feature extraction covers local details and global context information at the same time, provides multi-scale and multi-dimensional feature materials for subsequent feature fusion, and solves the problem of low recognition accuracy caused by insufficient feature dimensions in complex terrain.

[0012] Step 106, using the improved skip connection and feature fusion module to first perform convolution and batch normalization operation on the output features of each layer to obtain normalized features, then splicing the normalized features to obtain spliced features, and performing convolution and weighted fusion on the spliced features to obtain fused features.

[0013] An important improvement of the landslide identification model of the application is the use of dense skip connection structure, which makes the information transmission between the encoder and the decoder more sufficient, and avoids the deficiency of single level feature transmission. The application is further optimized on this basis. In the skip connection path, convolution and normalization operations are added to ensure the consistency of the features during transmission. Specifically, it includes: first, the output features of each layer of the encoder are convolved and batch normalized, then spliced to generate, and finally convolved and weighted fused to obtain the fused features. Convolution and batch normalization operations ensure consistent distribution when cross-layer feature transmission, avoid feature offset or distortion during transmission, accelerate model training, alleviate gradient vanishing, and improve feature transmission stability; multi-feature splicing and weighted fusion integrate the advantages of different levels of features, avoid the lack of single level feature information, strengthen the integrity and representativeness of the features, provide a high-quality feature basis for subsequent attention enhancement, and dense skip connection performs multi-level transmission of features between different layers of the encoder and the decoder, enhancing feature mobility.

[0014] Step 108, in the fusion stage, embed the hybrid attention module, perform global average pooling, multi-scale one-dimensional convolution operation on the fused features according to the EMA module, generate channel attention weights, multiply the channel attention weights with the fused features channel by channel to obtain EMA enhanced features; perform convolution and activation operations on the fused features according to the PPA module to generate spatial attention weights, multiply the spatial attention weights with the fused features element by element to obtain PPA enhanced features; splice the EMA enhanced features and the PPA enhanced features in the channel dimension to obtain spliced attention features, and perform convolution and activation on the spliced attention features to obtain the final enhanced features.

[0015] To solve the problems of insufficient feature expression and fuzzy boundary in the existing model in the landslide region segmentation, the application sets a hybrid attention module in the fusion stage, uses the EMA module to highlight the significant differences between different channels, efficiently captures the key channel features of the landslide region under multi-scale, improves the sensitivity of the model to complex environments, vegetation coverage and landslide under shadow, and solves the problem of insufficient multi-scale feature extraction; the PPA module is used to accurately locate the boundary pixels of the landslide and the background, suppress the background noise, effectively improve the prediction result adhesion or fracture caused by the fuzzy boundary, and improve the boundary segmentation accuracy; the parallel structure realizes the joint modeling of channel and spatial features, so that the final enhanced features have channel significance and spatial sensitivity at the same time, which greatly enhances the representation ability of the model to the landslide region.

[0016] Step 110, using the decoder to receive the final enhanced feature, recovering the spatial resolution by stepwise upsampling, and fusing with the final enhanced feature to obtain the output of each layer of the decoder.

[0017] At the same time of restoring the spatial resolution, the key features after attention enhancement are integrated into the decoding process to avoid losing the details of the landslide area during the resolution restoration, ensuring that the decoder output retains both the global structure and clear local details, providing a guarantee for the generation of accurate segmentation results.

[0018] Step 112, according to the output layer receiving decoder each layer output, through the activation function to get the pixel level class probability distribution, that is, the landslide segmentation result.

[0019] The decoder output is converted into intuitive pixel-level probability results, which can directly distinguish between landslide and non-landslide areas, and the probability value can quantify the confidence of the pixel belonging to the landslide, providing accurate and interpretable judgment basis for subsequent result post-processing and disaster warning.

[0020] Step 114, according to the pre-set loss function, the mountain landslide identification model is trained, and the trained mountain landslide identification model is used to realize the mountain landslide identification.

[0021] Setting the loss function can effectively alleviate the class imbalance problem that the number of landslide pixels in remote sensing images is much less than that of background pixels. BCE optimizes the intra-class loss, and Dice is more sensitive to landslide pixel loss. After the weighted balance of the two, the model training is avoided to be biased towards the non-landslide area, the landslide area missing rate is reduced, and the reliability and generalization ability of the model identification are improved.

[0022] The above mountain landslide identification method based on the mixed attention mechanism of channels and space, the present application through the encoder layer by layer convolution down sampling, taking into account the shallow layer fine-grained features and deep layer global semantic features, then through the improved jump connection convolution normalization, multi-feature splicing fusion, solve the problem of insufficient multi-scale feature extraction of traditional model; The embedded EMA-PPA parallel attention module, the former captures the channel dependence through multi-scale one-dimensional convolution, highlights the key channel features of the landslide, and the latter strengthens the boundary and target area response through spatial weight generation. The two jointly model the channel and spatial features, greatly improving the model's representation ability for landslide areas under complex terrain. Even with interference such as vegetation coverage and shadow, it can still accurately capture landslide features of different scales. The mixed attention module effectively optimizes the boundary segmentation, avoiding the sticking or breaking of the prediction results; the decoder and the enhanced features are fused to restore the spatial resolution while retaining the details, and the pixel-level probability distribution of the output layer further ensures the classification accuracy, significantly improving the accuracy of landslide area segmentation. Setting the loss function balances the loss contribution of positive and negative samples, alleviates the model bias problem caused by the low proportion of landslide pixels in remote sensing images, reduces the missing rate of landslide area, and ensures the reliability of the warning. The present application can be efficiently applied to landslide disaster monitoring and warning, taking into account the accuracy and anti-interference, providing strong technical support for traffic, infrastructure and personnel safety protection, and has wide practical value and popularization prospect.

[0023] In one of the embodiments, the remote sensing image is subjected to layer-by-layer convolution and down-sampling operation by the encoder to obtain the output features of each layer of the encoder, including: The input remote sensing image is subjected to layer-by-layer convolution and down-sampling operation by the encoder to extract feature representations of different scales, acquire texture, boundary and color fine-grained features in the shallow stage, and extract global semantic information and abstract feature representations in the deep stage, thereby obtaining the output features of each layer of the encoder.

[0024] In one of the embodiments, the output of the first layer of the encoder is represented as:

[0025] wherein, represents the feature map of the i-th layer, , , are the convolution kernel and the bias, respectively, represents the convolution operation, represents the nonlinear activation function.

[0026] In one of the embodiments, the output features of each layer are subjected to convolution and batch normalization operation by the improved skip connection and feature fusion module to obtain normalized features, the normalized features are spliced to obtain spliced features, and the spliced features are subjected to convolution and weighted fusion to obtain fusion features, including: The convolution and normalization operation are added in the skip connection path to make the features consistent during transmission, and the normalized features are

[0027] wherein, represents the batch normalization, is the convolution kernel for adjusting the channel; The normalized features transmitted by different layers are spliced to obtain the spliced features

[0028] wherein, represents the splicing operation, is the number of fused feature layers; The spliced features are subjected to convolution and weighted fusion to obtain the fusion features

[0029] wherein, , are the convolution kernel and the bias, respectively, represents the convolution operation, represents the nonlinear activation function.

[0030] ​In one embodiment, the EMA module performs global average pooling, multi-scale one-dimensional convolution operation on the fusion feature, generates channel attention weight, multiplies the channel attention weight with the fusion feature channel by channel to obtain EMA enhanced feature, including: The input feature is globally averaged pooled to obtain a channel descriptor:

[0031] wherein, H represents the height of the fusion feature, W represents the width of the fusion feature, represents the feature value of the fusion feature at channel c, pixel . Then use multi-scale one-dimensional convolution to model channel dependence:

[0032] The generated channel attention weight is:

[0033] The enhanced feature is represented as:

[0034] wherein, represents channel-by-channel multiplication.

[0035] In a specific embodiment, the channel descriptor is obtained by global average pooling, and multi-scale one-dimensional convolution is used to efficiently capture the local dependence between channels. The output weight vector is multiplied with the input feature channel by channel, thereby highlighting the significant differences between different channels. It can effectively improve the sensitivity of the model to landslide areas at multiple scales.

[0036] In one embodiment, the PPA module performs convolution and activation operations on the fusion feature to generate spatial attention weight, multiplies the spatial attention weight with the fusion feature element by element to obtain PPA enhanced feature, including: The PPA module performs convolution and activation operations on the fusion feature to generate spatial attention weight:

[0037] The spatial attention weight is multiplied with the fusion feature element by element to obtain the PPA enhanced feature: .

[0038] ​In specific embodiments, the long-range dependencies between pixels are captured by the spatial attention mechanism, which can highlight the key areas and suppress the background noise. It is particularly suitable for processing cases where the landslide and non-landslide boundaries are blurred and the texture differences are weak. The output spatial weight map is multiplied element by element with the input features, making the model pay more attention to the target boundary.

[0039] In one of the embodiments, the EMA enhanced features and the PPA enhanced features are spliced in the channel dimension to obtain spliced attention features, and the spliced attention features are convolved and activated to obtain the final enhanced features, including: The EMA enhanced features and the PPA enhanced features are spliced in the channel dimension to obtain spliced attention features:

[0040] The spliced attention features are convolved and activated to obtain the final enhanced features:

[0041] wherein, represents a nonlinear activation function, b is a bias.

[0042] In specific embodiments, the output features of the fused EMA and PPA are spliced in the channel dimension, and then fused through convolution and Sigmoid activation. The final output enhanced features have both channel saliency and spatial sensitivity, which can effectively highlight the landslide area and suppress irrelevant background.

[0043] In one of the embodiments, the decoder receives the final enhanced features, restores the spatial resolution through stepwise upsampling, and fuses with the final enhanced features to obtain the output of each layer of the decoder, including:

[0044] wherein, represents an upsampling operation, represents the final enhanced features, l represents the number of layers of the decoder.

[0045] In one of the embodiments, the output layer receives the output of each layer of the decoder, and obtains the pixel-level class probability distribution, i.e., the landslide segmentation result, through an activation function, including:

[0046] wherein, represents the probability that a pixel belongs to a landslide, represents a nonlinear activation function, is a bias term corresponding to the output layer, To output the weight parameters of the convolutional layer, Indicates the decoder's first Layers at pixels Output characteristics at the location.

[0047] In a specific embodiment, during the decoding stage, spatial resolution is restored by progressive upsampling and fused with attention-enhanced skip connection features. The improved decoder can recover both the global structure and preserve local details, thus exhibiting clear slippage boundaries in the segmentation results. Finally, the output layer obtains the pixel-level class probability distribution through a Softmax or Sigmoid activation function.

[0048] In one embodiment, the pre-set loss function is:

[0049]

[0050]

[0051] in, Indicates the true label, Indicates the predicted probability. Number of pixels and This indicates different weights.

[0052] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0053] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 2As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium, an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is used to communicate with the external terminal through the network connection. The computer program is executed by the processor to implement a landslide identification method based on a channel and space mixed attention mechanism. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0054] Those skilled in the art can understand that, Figure 2 The skilled in the art can understand that,

[0055] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM) and the like.

[0056] Any technical features in the above embodiments can be combined, and for the sake of brevity, not all possible combinations are described above, however, as long as the combinations of technical features do not contradict each other, they shall be considered within the scope of the present disclosure.

[0057] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it shall not be understood as a limitation on the scope of the present application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these shall be within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A method for landslide identification based on a channel and spatial mixed attention mechanism, characterized in that, The method comprises: acquiring remote sensing image data; constructing a landslide identification model; the landslide identification model comprises an encoder, an improved skip connection and feature fusion module, a hybrid attention module, a decoder and an output layer; the hybrid attention module is composed of an EMA module and a PPA module in parallel; using the encoder to perform layer-by-layer convolution and downsampling operations on the remote sensing image to obtain the output features of each layer of the encoder; using the improved skip connection and feature fusion module to first perform convolution and batch normalization operations on the output features of each layer to obtain normalized features, then splicing the normalized features to obtain spliced features, and performing convolution and weighted fusion on the spliced features to obtain fused features; embedding the hybrid attention module in the fusion stage, performing global average pooling and multi-scale one-dimensional convolution operations on the fused features according to the EMA module to generate channel attention weights, multiplying the channel attention weights and the fused features channel by channel to obtain EMA enhanced features, performing convolution and activation operations on the fused features according to the PPA module to generate spatial attention weights, and multiplying the spatial attention weights and the fused features element by element to obtain PPA enhanced features; splicing the EMA enhanced features and the PPA enhanced features in the channel dimension to obtain spliced attention features, and performing convolution and activation on the spliced attention features to obtain final enhanced features; using the decoder to receive the final enhanced features, restoring the spatial resolution by stepwise upsampling, and fusing the final enhanced features to obtain the output of each layer of the decoder; receiving the output of each layer of the decoder according to the output layer, and obtaining the class probability distribution at the pixel level, i.e., the landslide segmentation result, through an activation function; training the landslide identification model according to a pre-set loss function, and realizing landslide identification using the trained landslide identification model.

2. The method of claim 1, wherein, using the encoder to perform layer-by-layer convolution and downsampling operations on the remote sensing image to obtain the output features of each layer of the encoder, comprising: using the encoder to perform layer-by-layer convolution and downsampling operations on the input remote sensing image to extract feature representations of different scales, obtaining texture, boundary and color fine-grained features in the shallow stage, and extracting global semantic information and abstract feature representations in the deep stage to obtain the output features of each layer of the encoder.

3. The method of claim 2, wherein, The method further comprises: The encoder first The output of the layer is represented as: wherein, denotes the layer feature map, , are the convolution kernel and bias, respectively, denotes the convolution operation, denotes the nonlinear activation function.

4. The method of claim 1, wherein, using the improved skip connection and feature fusion module to first perform convolution and batch normalization operations on the output features of each layer to obtain normalized features, splicing the normalized features to obtain spliced features, and performing convolution and weighted fusion on the spliced features to obtain fused features, comprising: adding convolution and normalization operations in the skip connection path to maintain consistency of the features during transmission, obtaining normalized features as wherein, denotes batch normalization, is an adjusted convolution kernel for the channel; splicing the normalized features transmitted by different levels to obtain spliced features as wherein, denotes a concatenation operation, is the number of fused feature layers; performing convolution and weighted fusion on the spliced features to obtain fused features as: wherein, , are a convolution kernel and a bias, respectively, denotes a convolution operation, denotes a non-linear activation function.

5. The method of claim 1, wherein, performing global average pooling and multi-scale one-dimensional convolution operations on the fused features according to the EMA module to generate channel attention weights, multiplying the channel attention weights and the fused features channel by channel to obtain EMA enhanced features, comprising: performing global average pooling on the input features to obtain a channel descriptor: wherein, H represents the height of the fused feature, W represents the width of the fused feature, represents the feature value of the fused feature at channel c, pixel place; Then the channel dependency is modeled using multi-scale one-dimensional convolution as: The generated channel attention weight is: The enhanced feature representation is: wherein denotes a channel-wise multiplication.

6. The method of claim 5, wherein, According to the PPA module, the fused feature is convolved and activated to generate a spatial attention weight, and the spatial attention weight and the fused feature are multiplied element by element to obtain a PPA enhanced feature, including: According to the PPA module, the fused feature is convolved and activated to generate a spatial attention weight, and the spatial attention weight and the fused feature are multiplied element by element to obtain a PPA enhanced feature, including: The spatial attention weight and the fused feature are multiplied element by element to obtain a PPA enhanced feature, including: 。 7. The method of claim 6, wherein, The EMA enhanced feature and the PPA enhanced feature are spliced in the channel dimension to obtain a spliced attention feature, and the spliced attention feature is convolved and activated to obtain a final enhanced feature, including: The EMA enhanced feature and the PPA enhanced feature are spliced in the channel dimension to obtain a spliced attention feature, and the spliced attention feature is convolved and activated to obtain a final enhanced feature, including: The EMA enhanced feature and the PPA enhanced feature are spliced in the channel dimension to obtain a spliced attention feature, and the spliced attention feature is convolved and activated to obtain a final enhanced feature, including: wherein, denotes a non-linear activation function, b is a bias.

8. The method of claim 1, wherein, The decoder receives the final enhanced feature, restores the spatial resolution through step-by-step upsampling, and fuses with the final enhanced feature to obtain the output of each layer of the decoder, including: wherein, denotes an up-sampling operation, denotes the final enhanced feature, l denotes the number of layers of the decoder.

9. The method of claim 1, wherein, According to the output layer, the output of each layer of the decoder is received, and the class probability distribution of the pixel level is obtained through the activation function, that is, the landslide segmentation result, including: wherein, represents a pixel belongs to a landslide, represents a nonlinear activation function, is an output layer corresponding bias term, is a weight parameter of the output convolution layer, represents the output feature of the decoder's layer at pixel point .

10. The method of claim 1, wherein, The pre-set loss function is wherein, denotes the true label, denotes the predicted probability, is the number of pixels, and denotes a different weight.

Images