A landslide detection method based on uncertainty perception

Abstract

This invention discloses a landslide detection method based on uncertainty perception, belonging to the field of intelligent landslide disaster identification integrating remote sensing image processing and deep learning. The method includes: acquiring multi-source remote sensing image data of the study area and corresponding landslide ground truth labels, and constructing a training sample set; constructing a deep learning model, which includes a backbone convolutional neural network, a hollow spatial pyramid pooling module, a segmentation decoding branch, and an uncertainty decoding branch; iteratively training the deep learning model using the training sample set, updating the model parameters by minimizing the total loss function; inputting the remote sensing image data of the area to be detected into the trained deep learning model, and outputting a segmentation prediction map and an uncertainty weight map of the landslide area after model processing. This application significantly improves the accuracy and robustness of landslide identification in complex environments by integrating multi-source data, automatically focusing on difficult samples using the uncertainty branch, and combining multi-scale context capture, while also achieving higher training efficiency.

Description

Technical Field

[0001] This invention belongs to the field of intelligent landslide disaster identification technology that integrates remote sensing image processing and deep learning, and specifically relates to a landslide detection method based on uncertainty perception. Background Technology

[0002] Landslides are unstable movements on the earth's surface or within rock formations caused by factors such as gravity, changes in groundwater levels, and earthquakes. They are a common and highly destructive geological hazard.

[0003] With the widespread use of high-resolution remote sensing satellites, landslide detection using remote sensing imagery has become a mainstream method. However, existing landslide identification methods are mostly based on a single data source (such as optical imagery or DEM) and traditional machine learning (such as SVM, random forest) or conventional deep learning models (U-Net, ResNet), which have significant shortcomings in complex scenarios: single data sources have limited information dimensions, making it difficult to comprehensively characterize the topographic, spectral, and temporal features of landslides; conventional deep learning models treat all pixels equally, resulting in weak identification capabilities for difficult samples such as blurred edges, vegetation cover, and shadow occlusion; and they lack systematic multi-dimensional comparative verification, making it difficult to objectively evaluate the model's advantages. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention proposes a landslide detection method based on uncertainty perception. This method is rationally designed, overcomes the shortcomings of existing technologies, and has good performance.

[0005] A landslide detection method based on uncertainty perception includes the following steps: Step S1: Obtain multi-source remote sensing image data of the study area and corresponding landslide ground truth labels, and construct a training sample set; Step S2: Construct a deep learning model, including a backbone convolutional neural network for extracting image features, a dilated spatial pyramid pooling module for capturing multi-scale contextual information, and a segmentation decoding branch and an uncertainty decoding branch respectively connected to the output of the dilated spatial pyramid pooling module; the uncertainty decoding branch is used to generate an uncertainty weight map for each pixel; Step S3: Iteratively train the deep learning model using the training sample set, and update the model parameters by minimizing the total loss function; the total loss function includes the segmentation loss of the segmentation branch and the uncertainty loss of the uncertainty branch, wherein the segmentation loss is a Dice loss weighted by the uncertainty weight graph, and the uncertainty loss is constructed based on the Dice loss of uncertainty weight and gradient separation; Step S4: Input the remote sensing image data of the area to be detected into the trained deep learning model, and output the segmentation prediction map and uncertain weight map of the landslide area after model processing. Step S5: Compare the output segmentation results with the validation dataset, and calculate the accuracy, precision, recall, average pixel accuracy, and F1 score to evaluate the model performance.

[0006] Further, step S1 includes the following steps: The multi-source remote sensing image data is truncated to remove outliers, and different bands are fused using a weighted normalization strategy to obtain a fused image. Wavelet transform is used to perform multi-scale decomposition on the fused image, and optical image feature information and topographic feature information of the landslide area are extracted from the frequency domain and spatial domain. Constructing a training sample set: Design a dual encoder structure to process the extracted optical image feature information and terrain feature information respectively, capture detailed information of the landslide area at different scales, so as to enhance the ability to identify landslide edges and easily confused ground features, and obtain the processed sample set image data. The image data in the constructed training sample set are normalized and data augmented, and then divided into training set, validation set and test set.

[0007] Furthermore, the dilated spatial pyramid pooling module includes a 1×1 convolution branch, multiple dilated convolution branches with different strides, and a global pooling branch, which are used to extract multi-scale contextual features in parallel and fuse the outputs of each branch, enabling feature extraction from different scales.

[0008] Furthermore, the segmentation decoding branch and the uncertainty decoding branch adopt the same network structure, using the feature map output by the backbone convolutional neural network as the first input feature map. The first input feature map is adjusted for the number of channels through a 1×1 convolutional layer to extract key feature maps. The feature map output by the hollow spatial pyramid module is then used as the second input feature map. An upsampling layer is then applied to upsample the resolution of the second input feature map by a factor of four to obtain the final feature map. ;Will and After the concatenation operation, a fused feature map is obtained. The fused feature map is further processed by a 3×3 convolutional layer to capture more complex features. Finally, after upsampling, the feature map is restored to the final resolution. The segmentation decoding branch, following the main network structure, outputs a feature map through the last 1×1 convolutional layer, which is then mapped to the [0,1] interval by the Sigmoid activation function to generate a landslide segmentation prediction map, i.e., the probability value of each pixel belonging to "landslide". The uncertainty decoding branch, following the main network structure, maps the feature map output from the last 1×1 convolutional layer sequentially through the Sigmoid function to the [0,1] interval, and adds an offset constant. : ; In the formula, This is the output of the Sigmoid function. The feature map output by the last 1×1 convolutional layer; Subsequently, the entire image is normalized so that the sum of the uncertainty weights of all pixels is 1, resulting in the final pixel-level uncertainty weight map, expressed as: ; In the formula, For the first Uncertain weights for each pixel , These represent the height and width of the input image, respectively.

[0009] Further, step S3 includes the following steps: S31: First, calculate the Dice coefficient: ; In the formula, This represents the set of landslide areas predicted by the model. Represents a set of real landslide labels. This represents the area of ​​the intersection region between the predicted region and the true label. , These represent the areas of the predicted region and the label region, respectively. Based on this, the pixel-wise Dice loss is obtained: ; In the formula, Indicates the first The Dice loss value corresponding to each pixel; The loss distribution of the entire image is obtained by calculating the Dice loss pixel by pixel; S32: Weighting the uncertainty Gradient separation is performed to obtain an uncertain weight map used for weighted segmentation loss. : ; in, Represents the gradient separation function; Dice loss Perform gradient separation to obtain the loss term used for training the uncertain branch. : ; S33: Multiply the uncertainty weight map after gradient separation with the pixel-wise Dice loss to construct the uncertainty-weighted segmentation loss. : ; S34: Design an uncertainty-specific loss that is opposite in direction to the segmentation loss. This enables the uncertainty branch to learn the spatial distribution of difficult samples: ; S35: Weight and fuse the uncertainty-weighted segmentation loss and the uncertainty-branch-specific loss to form the final total loss function. : ; In the formula, To balance the hyperparameters, used to control the learning intensity of the two branches; The model aims to minimize the total loss. With the goal of improving the model, the Adam optimizer is used for iterative updates. The initial learning rate is set to 0.001, and the learning rate dynamically decreases with each round until the loss no longer decreases significantly, at which point the model is considered to have converged.

[0010] Furthermore, the remote sensing image data of the area to be detected in step S4 needs to undergo the same preprocessing process as in step S1 before being input into the model, including truncation, fusion, wavelet transform, and dual encoder feature extraction.

[0011] The beneficial technical effects of this invention are as follows: (1) Overcoming the inefficiency of multi-source data fusion: By using weighted normalization fusion, wavelet transform and dual encoder structure, key features of landslides are mined from spatial domain, frequency domain and multi-scale dimensions, and the ability to distinguish edges and easily confused features is strengthened. The completeness and accuracy of feature extraction are significantly better than those of single data source or simple fusion methods. (2) Solving the problems of insufficient learning of difficult samples and background interference: Innovatively introduce uncertainty branches, focus on difficult areas and optimize them by uncertainty weighting, effectively suppress background noise, solve the problems of serious misclassification and omission in existing models and blurred boundary segmentation, and achieve higher recognition accuracy in complex terrain. (3) Construct a comprehensive performance evaluation system: A two-dimensional comparison framework of "traditional machine learning + mainstream deep learning" and five core indicators were designed to form a multi-dimensional performance verification system, providing a scientific basis for technology selection and optimization. (4) Multiple technical means are not simply superimposed, but produce a synergistic effect: multi-source fusion provides rich feature input, uncertainty branch dynamically guides the model to focus on difficult samples, and the Atrous Spatial Pyramid Pooling (ASPP) module captures multi-scale context. The synergy of the three makes the model converge more efficiently during training and accurately delineates the landslide boundary during inference. Attached Figure Description

[0012] Figure 1 This is a flowchart of a landslide detection method based on uncertainty perception in this invention; Figure 2 This is a schematic diagram of a deep learning network model structure according to the present invention; Figure 3 This is a schematic diagram of the hollow space pyramid pooling module structure in this invention; Figure 4 This is a map showing the landslide extraction results within the study area in this embodiment of the invention; Figure 5 This is an enlarged view of the extraction results of key landslide areas in an embodiment of the present invention; Figure 6 This is a comparison diagram of the method of the present invention and existing methods; Figure 7 The diagram shows the extraction results of the method of this invention and existing methods; Among them, (a) is the effect diagram using Deeplabv3+; (b) is the effect diagram using RESUnet; (c) is the effect diagram using Efficient-Net; and (d) is the effect diagram using the method of this invention. Detailed Implementation

[0013] The specific embodiments of the present invention will be further described below with reference to specific examples: A landslide detection method based on uncertainty perception can be divided into three main stages: multi-source data processing and sample library construction, deep learning uncertainty model construction and training, and multi-method comparison, verification, and evaluation, forming a complete closed loop of "data-driven - model optimization - effect verification". Figure 1 As shown, the specific steps include: Step S1: Obtain multi-source remote sensing image data of the study area and corresponding landslide ground truth labels, and construct a training sample set; specifically: In this embodiment, the multi-source remote sensing image data includes downloading Sentinel-2A remote sensing images and DEM data of the study area, and creating NDVI difference images before and after the landslide, as well as time-series information difference images before and after the landslide.

[0014] To explore the roles and complementary characteristics of different data sources in landslide identification, a multi-source data fusion method for landslide identification is proposed. This method can provide more comprehensive and multi-dimensional surface feature information, which helps to accurately describe the morphology and boundaries of landslide areas and helps the model better understand the environmental context of landslide occurrence. First, the multi-source remote sensing image data is truncated to remove outliers. Then, different bands are fused using a weighted normalization strategy to obtain a fused image. The weighted normalization strategy specifically involves calculating weights based on the signal-to-noise ratio (SNR) of each band, assigning greater weights to bands with higher SNRs. The weighted sum of all bands is then performed to obtain the fused image. The fused image is decomposed into multiple scales using wavelet transform to extract key feature information of the landslide area from the frequency and spatial domains, which helps to capture the local features and details of the image. The wavelet transform uses the 'db4' wavelet basis for three-level decomposition to extract the low-frequency approximation components and high-frequency detail components of each level as multi-scale features.

[0015] Constructing a training sample set: Design a dual encoder structure to process optical image features and terrain features respectively, capture detailed information of landslide areas at different scales, and enhance the ability to identify landslide edges and easily confused features; the dual encoder includes an optical image encoder and a terrain encoder, which adopt ResNet50 and a lightweight CNN respectively. The extracted features are input into subsequent modules after channel splicing and fusion.

[0016] The images in the constructed training sample set are normalized and data augmented, and then divided into training set, validation set and test set.

[0017] Step S2: Construct a deep learning model, such as Figure 2 As shown, it includes a backbone convolutional neural network for extracting image features, a dilated spatial pyramid pooling module for capturing multi-scale contextual information, and a segmentation decoding branch and an uncertainty decoding branch respectively connected to the output of the dilated spatial pyramid pooling module; the uncertainty decoding branch is used to generate an uncertainty weight map for each pixel; specifically: This paper proposes an innovative uncertain landslide detection model to overcome the challenges of fine-grained landslide detection. The model identifies hard-to-detect regions (uncertainties) in images through neural network analysis of landslide features and prioritizes segmentation of these regions during training, assigning them higher uncertainty weights to improve the accuracy of identifying small landslides. Importantly, the uncertainty branch module runs independently during training without incurring additional inference costs. The model also innovatively integrates multiple data sources to enhance feature understanding, enabling it to simultaneously consider terrain, landforms, and other features at different scales. The model is trained and validated using Sentinel-2A satellite imagery and DEM data, and the recognition results are evaluated using various metrics.

[0018] In this embodiment, the backbone convolutional neural network is ResNet101, a variant of deep learning that performs well and is designed for handling complex image tasks. ResNet101 is characterized by its depth, having 101 layers. It introduces the concept of residual learning, where each residual block in the architecture contains multiple convolutional layers, and residual connections between every two consecutive convolutional layers are merged. This complex structure helps the network learn the mapping relationship between input and output more seamlessly, thereby enhancing its expressive power. One type of cross-layer shortcut connection directly passes the input of the previous layer to the next, aiming to alleviate the gradient vanishing problem in deep neural networks and enhance training stability.

[0019] Following the backbone network is the Spatial Pyramid Pooling (ASPP) module, a key network module specifically designed for semantic segmentation. Its main goal is to address the challenge of retrieving semantic information across different scales. This module employs atrous convolution with varying strides, significantly expanding the receptive field and effectively capturing multi-scale information. It includes a 1×1 convolutional branch, multiple atrous convolutional branches with different strides, and a global pooling branch for parallel extraction of multi-scale contextual features, and then fuses the outputs of each branch.

[0020] like Figure 3 As shown, in this embodiment, there are 3 dilated convolution branches, each corresponding to a 3×3 convolutional layer with strides of 6, 12, and 18 respectively. The global pooling branch corresponds to a 1×1 pooling layer. The output of the 1×1 convolutional layer is concatenated with the outputs of the 1×1 convolutional branch and the 3 dilated convolutional branches after passing through the 1×1 convolutional layer and the 1×1 upsampling layer. The concatenated result is then output after passing through a 1×1 convolutional layer.

[0021] The segmentation decoding branch and the uncertainty decoding branch use the same network structure, taking the feature map output by the backbone convolutional neural network as the first input feature map. The first input feature map is then passed through a 1×1 convolutional layer to adjust the number of channels and extract key feature maps. The feature map output from the hollow space pyramid module is then used as the second input feature map. An upsampling layer is then applied to upsample the resolution of the second input feature map by a factor of four to obtain the final feature map. This makes it closer to the original input image. and After the concatenation operation, the upsampled feature map is fused with the feature map of the corresponding layer in the encoder to ensure effective combination of multi-scale information. The fused feature map is further processed through a 3×3 convolutional layer to capture more complex features. Finally, after upsampling again, the feature map is restored to the final resolution.

[0022] The segmentation and decoding branch follows the main structure of the network. It outputs a feature map through the last 1×1 convolutional layer, and then maps it to the [0,1] interval through the Sigmoid activation function to generate a landslide segmentation prediction map, which is the probability value of each pixel belonging to "landslide". The uncertainty decoding branch aims to provide an uncertainty measure for each pixel, improving the reliability of landslide predictions for areas that are difficult to identify. It generates an uncertainty weight map reflecting the intensity of uncertainty in each pixel and combines this with the prediction results from the landslide segmentation branch to generate a comprehensive loss map, providing more reliable reference information for accurate landslide segmentation. Unlike other decoder branches, the latter part of the segmentation branch decoder is integrated with the model's uncertainty branch. This integration allows for uncertainty weighting, thereby enhancing the model's performance in challenging detection scenarios, such as those involving small landslides or complex backgrounds. After processing by this part of the decoder, the model produces more accurate segmentation results.

[0023] First, the feature map output from the last 1×1 convolutional layer is sequentially mapped to the [0,1] interval using the Sigmoid function, and an offset constant is added. , The hyperparameter is greater than 0, used to improve numerical stability in the low uncertainty region. In the experiments of this application, c=0.001 is preferred, and the expression is: ; In the formula, This is the output of the Sigmoid function. The feature map output by the last 1×1 convolutional layer; Subsequently, the entire image is normalized so that the sum of the uncertainty weights of all pixels is 1, resulting in the final pixel-level uncertainty weight map, expressed as: ; In the formula, For the first Uncertain weights for each pixel , These represent the height and width of the input image, respectively. A larger weight indicates less certainty in the model's prediction for that pixel, classifying it as a difficult sample; a smaller weight indicates more certainty in the model's prediction, classifying it as a simple sample.

[0024] Deep neural networks typically begin by quickly fitting simple samples and gradually transition to fitting more complex ones. This method integrates uncertainty information from a large number of pixels into the training process. During training, pixels with higher loss are assigned greater uncertainty weights, generating an uncertainty weight map. The weighted segmentation loss is obtained by multiplying the uncertainty weight map by the Dice loss map. The optimization of this branch is driven by the uncertainty-weighted segmentation loss, thereby enhancing the model's adaptability to image data of varying complexity.

[0025] By focusing on the more uncertain aspects, the model gradually improves the accuracy of landslide identification.

[0026] Step S3: Iteratively train the deep learning model using the training sample set, and update the model parameters by minimizing the total loss function; the total loss function includes the segmentation loss of the segmentation branch and the uncertainty loss of the uncertainty branch, wherein the segmentation loss is a Dice loss weighted by the uncertainty weight graph, and the uncertainty loss is constructed based on the Dice loss of uncertainty weight and gradient separation. By constructing an uncertainty-weighted loss function, the model can automatically identify areas with blurred landslide boundaries, shaded areas, vegetated areas, small landslide areas, fragmented morphology areas, and difficult sample areas easily confused with roads, bare soil, buildings, and rivers during training. Higher learning weights are assigned to these areas, enabling the model to learn landslide characteristics more accurately within a limited number of iterations. Specifically: S31: Landslide segmentation is a typical small-object segmentation task with significant sample imbalance, making it difficult for traditional cross-entropy loss to achieve good boundary accuracy. This method uses Dice loss as the basic loss function to measure the similarity between the landslide segmentation results output by the model and the true labels.

[0027] The similarity measure between the segmentation results generated by the model and the actual segmentation labels is a set similarity metric function used to calculate the similarity between two samples, with values ​​ranging from [0, 1]. Specifically: First, calculate the Dice coefficient: In the formula, This represents the set of landslide areas predicted by the model. Represents a set of real landslide labels. This represents the area of ​​the intersection region between the predicted region and the true label. , These represent the areas of the predicted region and the labeled region, respectively.

[0028] Based on this, the pixel-wise Dice loss is obtained: In the formula, Indicates the first The Dice loss value corresponds to each pixel. The larger the loss value, the more difficult it is to classify the pixel correctly, i.e., it belongs to the hard sample category.

[0029] By calculating the Dice loss pixel by pixel, the loss distribution of the entire image can be obtained. High loss areas correspond to landslide boundaries, broken areas, occluded areas, small landslides, and easily confused ground features.

[0030] The combination of these technologies helps the model effectively learn the uncertainty information of pixels, especially when pixels present varying degrees of segmentation difficulty. This helps improve the model's accuracy across different regions in image segmentation tasks.

[0031] S32: To ensure that the split branch and the uncertain branch can learn independently and without interfering with each other, this application introduces a gradient separation mechanism. During backpropagation, gradient transmission between the uncertain branch and the split branch is blocked, so that the optimization objectives of the two branches do not affect each other.

[0032] Weight of uncertainty Gradient separation is performed to obtain the uncertainty weight map used for weighted segmentation loss: ; in, Represents the gradient separation function; Dice loss Gradient separation is performed, and the gradient flow only propagates along the uncertain branch, without backpropagating to the split branch, resulting in a loss term used to train the uncertain branch. : ; By separating gradients, the segmentation branch only accepts gradient updates from the segmentation loss, and the uncertainty branch only accepts gradient updates from the uncertainty loss. The backbone network and the ASPP module simultaneously receive gradient information from both branches, enabling multi-task joint learning.

[0033] S33: After obtaining the uncertainty weight map after gradient separation, this application multiplies it with the pixel-wise Dice loss to construct the uncertainty-weighted segmentation loss. : ; By assigning higher loss weights to high-uncertainty (difficult samples) and lower loss weights to low-uncertainty (simple samples), the model can automatically focus on difficult regions, significantly improving boundary accuracy and small landslide detection rate.

[0034] S34: Design an uncertainty-specific loss that is opposite in direction to the segmentation loss. This enables the uncertainty branch to learn the spatial distribution of difficult samples: ; Increase the uncertainty weights for high-loss pixels and decrease the uncertainty weights for low-loss pixels to gradually focus on difficult regions during training.

[0035] S35: Weight and fuse the uncertainty-weighted segmentation loss and the uncertainty-branch-specific loss to form the final total loss function. : ; In the formula, To balance the hyperparameters, used to control the learning intensity of the two branches.

[0036] The model aims to minimize the total loss. With the goal of improving the model, the Adam optimizer is used for iterative updates. The initial learning rate is set to 0.001, and the learning rate dynamically decreases with each round until the loss no longer decreases significantly, at which point the model is considered to have converged.

[0037] Step S4: Input the remote sensing image data of the area to be detected into the trained deep learning model. After model processing, the model outputs a segmentation prediction map and an uncertain weight map of the landslide area. In this embodiment, the segmentation prediction map is as follows: Figure 4 and Figure 5 As shown.

[0038] Before being input into the model, the remote sensing image data of the area to be detected needs to undergo the same preprocessing process as in step S1, including truncation, fusion, wavelet transform, and dual encoder feature extraction.

[0039] Step S5: Compare the segmentation results output in step S4 with the real labels, and calculate the accuracy, precision, recall, average pixel accuracy, and F1 score to evaluate the model performance.

[0040] In this embodiment, step S5 further includes comparing and analyzing the method of the present invention with at least one benchmark method among Support Vector Machine, Random Forest, U-Net, and ResNet on the same test set to comprehensively verify the performance advantages of the present method in the landslide identification task, specifically as follows: (1) Comparative experimental design: ① Benchmark method selection: Two types of methods were selected as comparison benchmarks: Traditional machine learning methods: Support Vector Machine (SVM), Random Forest (RF); Mainstream deep learning methods: U-Net, ResNet.

[0041] ② Uniform experimental setup: Experiments are conducted under the same dataset, hardware environment, and evaluation metrics to ensure the fairness of the comparison results.

[0042] (2) Quantitative assessment: ① Calculation of core metrics: Accuracy, Precision, Recall, Average Pixel Accuracy (MPA), and F1 score are calculated for all methods on the test set to quantify model performance from different dimensions; Efficiency analysis: Record the training time and inference speed of each method to evaluate computational efficiency.

[0043] (3) Qualitative and visual analysis: ①Result visualization: The segmentation results of different methods are visualized and compared to intuitively demonstrate the advantages of this method in identifying landslide edge details and complex backgrounds; ② Error analysis: Analyze the characteristics of false positives and false negatives, and combine with the uncertainty heatmap to further optimize the weak links of the model.

[0044] (4) Verification of conclusions: Through comprehensive analysis of quantitative indicators and qualitative results, the superiority of this method in terms of landslide identification accuracy, generalization ability and robustness is verified, providing a scientific basis for the implementation of the technology.

[0045] To verify the effectiveness of this invention, the results of the method proposed in this application were compared with the recognition results of Deeplabv3+, ResUNet, and Efficient-Net, respectively: This application focuses on three regions: Papua New Guinea, Nyingchi in Tibet, China, and Quang Nam in Vietnam. Data samples were created and models were trained. The final comparison results are as follows: Figure 6 As shown, the accuracy, precision, recall, average pixel accuracy, and F1 score of the method in this application are 98.5, 76.6, 95.8, 96.7, and 82.6, respectively. All accuracy metrics are higher than the other three methods. The detection performance of the method in this application is superior to that of Deeplabv3+, ResUNet, and Efficient-Net.

[0046] The effects of different methods on landslide extraction in the test area are as follows: Figure 7 As shown in (a) to (d), the method proposed in this application can effectively cope with the special mountain and canyon landforms and frequent precipitation, and has achieved excellent performance in various key indicators.

[0047] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A landslide detection method based on uncertainty perception, characterized in that, Includes the following steps: Step S1: Obtain multi-source remote sensing image data of the study area and corresponding landslide ground truth labels, and construct a training sample set; Step S2: Construct a deep learning model, including a backbone convolutional neural network for extracting image features, a hollow spatial pyramid pooling module for capturing multi-scale contextual information, and a segmentation decoding branch and an uncertainty decoding branch respectively connected to the output of the hollow spatial pyramid pooling module; The uncertainty decoding branch is used to generate an uncertainty weight map for each pixel; Step S3: Iteratively train the deep learning model using the training sample set, and update the model parameters by minimizing the total loss function; the total loss function includes the segmentation loss of the segmentation branch and the uncertainty loss of the uncertainty branch, wherein the segmentation loss is a Dice loss weighted by the uncertainty weight graph, and the uncertainty loss is constructed based on the Dice loss of uncertainty weight and gradient separation; Step S4: Input the remote sensing image data of the area to be detected into the trained deep learning model, and output the segmentation prediction map and uncertain weight map of the landslide area after model processing. Step S5: Compare the output segmentation results with the validation dataset, and calculate the accuracy, precision, recall, average pixel accuracy, and F1 score to evaluate the model performance.

2. The method of claim 1, wherein, Step S1 includes the following steps: The multi-source remote sensing image data is truncated to remove outliers, and different bands are fused using a weighted normalization strategy to obtain a fused image. Wavelet transform is used to perform multi-scale decomposition on the fused image, and optical image feature information and topographic feature information of the landslide area are extracted from the frequency domain and spatial domain. Constructing a training sample set: Design a dual encoder structure to process the extracted optical image feature information and terrain feature information respectively, capture detailed information of the landslide area at different scales, so as to enhance the ability to identify landslide edges and easily confused ground features, and obtain the processed sample set image data. The image data in the constructed training sample set are normalized and data augmented, and then divided into training set, validation set and test set.

3. The method of claim 1, wherein, The dilated spatial pyramid pooling module includes a 1×1 convolution branch, multiple dilated convolution branches with different strides, and a global pooling branch, which are used to extract multi-scale contextual features in parallel and fuse the outputs of each branch, enabling feature extraction from different scales.

4. The method according to claim 3, wherein, The segmentation decoding branch and the uncertainty decoding branch adopt the same network structure, using the feature map output by the backbone convolutional neural network as the first input feature map. The first input feature map is adjusted for the number of channels through a 1×1 convolutional layer to extract key feature maps. ; The feature map output by the hollow spatial pyramid module is then used as the second input feature map. The resolution of the second input feature map is upsampled by a factor of 4 through an upsampling layer to obtain the final feature map. ;Will and After the concatenation operation, a fused feature map is obtained. The fused feature map is further processed by a 3×3 convolutional layer to capture more complex features. Finally, after upsampling, the feature map is restored to the final resolution. The segmentation decoding branch, following the main network structure, outputs a feature map through the last 1×1 convolutional layer, which is then mapped to the [0,1] interval by the Sigmoid activation function to generate a landslide segmentation prediction map, i.e., the probability value of each pixel belonging to "landslide". The uncertainty decoding branch is behind the network main structure, and the feature map output by the last 1x1 convolution layer is sequentially mapped to the interval [0, 1] through a Sigmoid function and an offset constant : ； In the formula, is the output result of the Sigmoid function, is the feature map output by the last 1 × 1 convolutional layer. Subsequently, the entire image is normalized so that the sum of the uncertainty weights of all pixels is 1, resulting in the final pixel-level uncertainty weight map, expressed as: ； wherein is the uncertainty weight for the th pixel, , are the height and width of the input image, respectively.

5. The method of claim 4, wherein, Step S3 includes the following steps: S31: First, calculate the Dice coefficient: ； In the formula, This represents the set of landslide areas predicted by the model. Represents a set of real landslide labels. This represents the area of ​​the intersection region between the predicted region and the true label. , These represent the areas of the predicted region and the label region, respectively. Based on this, the pixel-wise Dice loss is obtained: ； In the formula, Indicates the first The Dice loss value corresponding to each pixel; The loss distribution of the entire image is obtained by calculating the Dice loss pixel by pixel; S32: determining an uncertainty weight performing gradient separation to obtain an uncertainty weight map for weighting the segmentation loss : ； wherein denotes a gradient separation function; Dice loss Perform gradient separation to obtain the loss term used for training the uncertain branch. : ； S33: multiply the uncertainty weight map after gradient separation with the pixel-wise Dice loss to construct the uncertainty weighted segmentation loss : ； S34: design an uncertainty branch-specific loss opposite to the segmentation loss direction enable the uncertainty branch to learn the spatial distribution of difficult samples: ； S35: weight the uncertainty-weighted segmentation loss and the uncertainty branch-specific loss to form a final total loss function : ； In the formula, is a balance hyperparameter used to control the learning strength of the two branches; The model aims to minimize the total loss. With the goal of improving the model, the Adam optimizer is used for iterative updates. The initial learning rate is set to 0.001, and the learning rate dynamically decreases with each round until the loss no longer decreases significantly, at which point the model is considered to have converged.

6. The method of claim 1, wherein, The remote sensing image data of the region to be detected in step S4 needs to undergo the same pre-processing procedures as step S1 before inputting the model, including truncation operation, fusion, wavelet transform and double-encoder feature extraction.

Images