Multi-criteria deep forecasting method of landslide risk integrating geographical and physical information
By using deep learning methods that integrate geographic and physical information, combined with high-resolution satellite imagery and a multi-criteria filter, accurate prediction of landslide risk was achieved. This solves the problem of the lack of dynamic physical mechanisms and multi-criteria evaluation in existing technologies, and improves the accuracy and spatial precision of prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING TECH & BUSINESS UNIV
- Filing Date
- 2025-09-09
- Publication Date
- 2026-06-16
Smart Images
Figure CN121170617B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a multi-criteria deep prediction method for landslide risk that integrates geographic and physical information, belonging to the field of geological disaster early warning, and is particularly suitable for multi-criteria deep prediction of landslide risk that integrates geographic and physical information. Background Technology
[0002] Landslides are among the most destructive natural disasters, severely damaging ecosystems, infrastructure, and human lives. Effective disaster management, urban design, and environmental protection all rely on accurate predictions of landslide-prone areas. Traditional prediction techniques often depend on statistical models, which struggle to account for the dynamic and complex factors influencing landslides. Data availability, the nonlinearity of topographic features, and the geographical variability of landslide triggering factors can all limit these methods. Thanks to advancements in remote sensing and deep learning, powerful tools are now available to improve landslide risk prediction, thereby increasing the accuracy and automation of the process. However, accurate landslide sensitivity mapping remains challenging due to its complex triggering mechanisms. Traditional techniques, such as heuristics and statistical models, often suffer from generalization limitations due to their reliance on specific criteria and small datasets.
[0003] With the development of remote sensing, Geographic Information Systems (GIS), and deep learning, landslide prediction models have undergone significant changes, enabling more accurate and automated sensitivity assessments. A landslide sensitivity map is a spatial distribution map that visually displays the probability of landslides occurring within a region using color gradients. Essentially, it's a "probabilistic map" that transforms complex geological, topographical, and environmental analyses into visualized risk level information. In recent years, Convolutional Neural Networks (CNNs), Deep Neural Networks (DNNs), Long Short-Term Memory Networks (LSTMs), and Recurrent Neural Networks (RNNs) have been used to extract spatial patterns from high-resolution satellite imagery for landslide sensitivity mapping. Deep learning methods have proven to achieve significantly better classification accuracy and performance, far surpassing traditional statistical methods.
[0004] However, simply combining high-resolution satellite imagery with geographic information systems results in static image segmentation and classification, lacking a characterization of dynamic physical mechanisms. Some factors contributing to landslides are well-defined and can be utilized. Furthermore, existing methods lack multi-faceted and multi-criteria comparison and evaluation mechanisms for landslide prediction results. Summary of the Invention
[0005] In view of this, the present invention aims to accurately analyze and predict landslide risks within a region based on high-resolution satellite imagery. It proposes to approach the problem from the perspectives of image segmentation of geographic information and regression analysis of physical information, combined with a deep learning network to achieve landslide sensitivity mapping. Finally, it employs multiple evaluation criteria to select the most accurate landslide sensitivity map pixel by pixel. To achieve the above objective, the present invention provides the following technical solution:
[0006] This invention provides a multi-criteria deep prediction method for landslide risk that integrates geographic and physical information, implemented based on a deep learning prediction system for landslide risk. The landslide risk deep learning prediction system consists of a high-resolution satellite, a base station containing a server, and a client computer. The satellite and base station are connected wirelessly, and the base station server and client computer are connected via a network. Its architecture is as follows: Figure 1 As shown; wherein, the client computer includes a data acquisition module, a data storage module, a data preprocessing module, and a prediction module; combined with Figure 2 The data acquisition module reads high-resolution satellite image data captured by the satellite through communication according to the sampling period. The data storage module consists of an acquisition data storage unit and a prediction data storage unit. The acquisition data storage unit is memory, connected to both the data acquisition module and the data preprocessing module, and is used to store and transmit the high-resolution satellite image data acquired by the data acquisition module. The prediction data storage unit is memory, connected to the prediction module, and is used to store and transmit the data fed back by the prediction module. The data preprocessing module is connected to the prediction module and is used to preprocess the high-resolution satellite image data. The prediction module is a processor with a deep neural network for geographic and physical information in memory. The deep neural network for geographic and physical information consists of a parallel deep neural network for geographic information and a deep neural network for physical information connected in series with a multi-criteria filter. The deep neural network for geographic information consists of a semantic segmentation network, a CNN, and a classifier connected in series. The deep neural network for physical information consists of a Digital Elevation Model (DEM). The system consists of a model (DEM), a statistical model, and a classifier connected in series to predict the landslide risk classification and confidence level corresponding to image pixels. The statistical model is a neural network of multiple regression analysis, used to statistically analyze the probabilistic relationship between topographic factors and landslide risk. The multi-criteria filter is a classifier used to evaluate and select the prediction results of the geographic information deep neural network and the physical information deep neural network.
[0007] A multi-criteria deep prediction method for landslide risk integrating geographic and physical information, the specific steps of which are as follows:
[0008] S1: The data acquisition module periodically acquires high-resolution satellite image data captured by the satellite and stores it in the acquisition data storage unit of the data storage module;
[0009] S2: The data preprocessing module scales and normalizes the acquired high-resolution satellite image data;
[0010] S3: The prediction module builds deep neural networks for geographic and physical information, and uses historical data to independently train the deep neural networks for geographic and physical information respectively.
[0011] S4: The prediction module, based on the trained geographic information deep neural network and physical information deep neural network, uses historical data again to train the multi-criteria filter.
[0012] S5: The prediction module uses a trained deep neural network of geographic and physical information to process high-resolution satellite image data to obtain a landslide sensitivity map.
[0013] Furthermore, combined with Figure 3 The specific processing flow of the aforementioned geographic information deep neural network is as follows:
[0014] (1-1) Masking the high-resolution satellite image data to obtain a masked image;
[0015] (1-2) The mask image is segmented using a semantic segmentation network to obtain a mask segmentation map;
[0016] (1-3) Use CNN to predict the mask segmentation map to obtain the sampled landslide sensitivity map; where each pixel value in the sampled landslide sensitivity map corresponds to the landslide risk probability value of that pixel.
[0017] (1-4) Using a classifier, the sampled landslide sensitivity maps are classified pixel by pixel according to the risk level to obtain the landslide sensitivity map predicted by the geographic information deep neural network and the corresponding confidence matrix.
[0018] Preferably, the semantic segmentation network is composed of DeepLabV3+ with cascaded softmax layers; the CNN is composed of an ASPP (Atrous Spatial Pyramid Pooling) module, which is a feature extractor based on ResNet and consists of multiple convolutional layers and pooling layers of different scales; and the classifier is a K-means network.
[0019] Furthermore, combined with Figure 3 The specific processing flow of the physical information deep neural network is as follows:
[0020] (2-1) Use digital elevation models to extract contour lines from high-resolution satellite image data to obtain contour maps;
[0021] (2-2) Extracting topographic factors from the contour map pixel by pixel, which consist of physical quantities such as slope, aspect, curvature, and topographic humidity index;
[0022] (2-3) Establish a statistical model for multiple regression analysis of topographic factors, and use the statistical model to calculate the probability value of landslide risk for each pixel to obtain a statistical landslide sensitivity map;
[0023] (2-4) Classifier 2 is used to classify the statistical landslide sensitivity map pixel by pixel according to the risk level, so as to obtain the landslide sensitivity map predicted by the physical information deep neural network and the corresponding confidence matrix.
[0024] Furthermore, the slope described in step (2-2) Slope aspect curvature Topographic humidity index Calculated by the raster generation tool in the digital elevation model, where, for The height corresponding to the pixel for The upstream catchment area along the contour line.
[0025] Preferably, the statistical model is one of logistic regression, random forest, or DNN.
[0026] Furthermore, the training described in step S3 specifically involves: using the cross-entropy loss function and combining it with manually labeled historical data, to fine-tune the parameters of the semantic segmentation network, CNN, and classifier one in the geographic information deep neural network, as well as the statistical model and classifier two in the physical information deep neural network.
[0027] Furthermore, combined with Figure 4 The specific working principle of the multi-criteria filter is as follows:
[0028] (3-1) The landslide sensitivity map and the corresponding confidence matrix predicted by the geographic information deep neural network, and the landslide sensitivity map and the corresponding confidence matrix predicted by the physical information deep neural network are used as inputs, and each pixel is input into the multi-criteria filter.
[0029] (3-2) When the classification results of the landslide sensitivity map predicted by the geographic information deep neural network and the landslide sensitivity map predicted by the physical information deep neural network are consistent, and there is a threshold in the corresponding confidence matrix that is greater than the manually set threshold, then the consistency result is directly adopted as the final prediction result of the pixel; otherwise, step (3-3) is executed.
[0030] (3-3) Use classifier 3 to reclassify and predict the pixel to obtain the final prediction result of the pixel.
[0031] Furthermore, the training process described in step S4 involves using the cross-entropy loss function and combining it with manually labeled historical data to optimize the parameters of classifier three in the multi-criteria filter.
[0032] Preferably, either the geographic information deep neural network or the physical information deep neural network can be used alone to draw a landslide sensitivity map.
[0033] The beneficial effects of this invention are as follows: This invention provides a multi-criteria deep prediction method for landslide risk that integrates geographic and physical information. Starting from the synergy of geographic and physical information flows, it proposes a deep neural network for geographic and physical information with a cross-validation mechanism, which ensures the accuracy of landslide risk prediction. At the same time, this network architecture, by using a deep learning-driven semantic segmentation framework, can accurately classify multiple types of land cover, significantly improve the spatial accuracy and robustness of terrain analysis, and enhance prediction capabilities by integrating terrain factors, which can effectively support the identification of landslide-prone areas in geomorphologically complex regions. Attached Figure Description
[0034] To illustrate the objectives and technical solutions of this invention, the following figures are provided:
[0035] Figure 1 This is a diagram illustrating the architecture of the deep learning prediction system for landslide risk in this invention.
[0036] Figure 2 This is a module architecture diagram of the client computer in this invention;
[0037] Figure 3 This is a flowchart of the deep neural network for geographic and physical information in this invention;
[0038] Figure 4 This is a flowchart of the multi-criteria filter in this invention;
[0039] Figure 5 This is a flowchart of Embodiment 1 of the method of the present invention;
[0040] Figure 6 This is a flowchart of the geographic information deep neural network in Embodiment 1 of the method of the present invention;
[0041] Figure 7 The comparison results of the mask segmentation map in Embodiment 1 of the method of the present invention;
[0042] Figure 8 This is a module architecture diagram of the client computer in Embodiment 2 of the method of the present invention;
[0043] Figure 9 This is a module architecture diagram of the client computer in Embodiment 3 of the method of the present invention. Detailed Implementation
[0044] To make the objectives and technical solutions of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and embodiments.
[0045] Example 1: Accurate landslide risk prediction is provided for the changing terrain based on high-resolution satellite images of Dubai from MBRSC [1]. This invention provides a multi-criteria depth prediction method for landslide risk that integrates geographic and physical information.
[0046] [1] “Semantic Segmentation Dataset Humans in the Loop.” Accessed: Mar. 16, 2025. [Online]. Available: https: / / humansintheloop.org / resources / datasets / semantic-segmentation-dataset-2 / .
[0047] Specifically, this method is implemented based on a landslide risk deep learning prediction system. The landslide risk deep learning prediction system consists of a high-resolution satellite 1, a base station 2 containing a server, and a client computer 3. Satellite 1 and base station 2 are connected wirelessly, and the base station 2 server and client computer 3 are connected via a network. Its architecture is as follows: Figure 1 As shown; wherein, the client computer 3 includes a data acquisition module 31, a data storage module 32, a data preprocessing module 33, and a prediction module 34; combined with Figure 2The data acquisition module 31 reads high-resolution satellite image data captured by satellite 1 via communication according to the sampling period; the data storage module 32 consists of an acquisition data storage unit 3201 and a prediction data storage unit 3202; the acquisition data storage unit 3201 is memory, connected to the data acquisition module 31 and the data preprocessing module 33 respectively, and is used to store and transmit the high-resolution satellite image data acquired by the data acquisition module 31; the prediction data storage unit 3202 is memory, connected to the prediction module 34, and is used to store and transmit the data fed back by the prediction module 34; the data preprocessing module 33 is connected to the prediction module 34 and is used to preprocess the high-resolution satellite image data; the prediction module 34 is a processor with a deep neural network for geographic and physical information in memory. The aforementioned geographic and physical information deep neural network consists of parallel geographic information deep neural networks and physical information deep neural networks connected in series with a multi-criteria filter 3407. The geographic information deep neural network is composed of a semantic segmentation network 3401, a CNN 3402, and a classifier 3403 connected in series. The physical information deep neural network is composed of a digital elevation model 3404, a statistical model 3405, and a classifier 3406 connected in series. It is used to predict the landslide risk classification and confidence level corresponding to image pixels. The statistical model 3405 is a neural network for multiple regression analysis, used to statistically analyze the probabilistic relationship between topographic factors and landslide risk. The multi-criteria filter 3407 is a classifier used to evaluate and select the prediction results of the geographic information deep neural network and the physical information deep neural network.
[0048] Combination Figure 5 The specific steps of this method are as follows:
[0049] S1: The data acquisition module 31 transmits commands through the base station 2 to periodically acquire high-resolution satellite image data captured by the satellite 1 and store it in the acquisition data storage unit 3201 in the data storage module 32.
[0050] S2: Data preprocessing module 33 scales the acquired high-resolution satellite image data to... Resolution is normalized. Furthermore, image enhancement can be achieved through processes including random rotation, flipping, and brightness variations.
[0051] S3: Prediction module 34 builds deep neural networks for geographic and physical information, and uses historical data to independently train the deep neural networks for geographic and physical information respectively.
[0052] S4: The prediction module 34, based on the trained geographic information deep neural network and physical information deep neural network, retrains the multi-criteria filter 3407 using historical data.
[0053] S5: Prediction module 34 uses a trained deep neural network of geographic and physical information to process high-resolution satellite image data to obtain a landslide sensitivity map.
[0054] Furthermore, combined with Figure 3 The specific processing flow of the aforementioned geographic information deep neural network is as follows:
[0055] (1-1) Masking the high-resolution satellite image data to obtain a masked image;
[0056] (1-2) The mask image is segmented using the semantic segmentation network 3401 to obtain the mask segmentation map;
[0057] (1-3) The mask segmentation map is predicted using CNN3402 to obtain the sampled landslide sensitivity map; where each pixel value in the sampled landslide sensitivity map corresponds to the landslide risk probability value of that pixel.
[0058] (1-4) The sampled landslide sensitivity map is classified into individual pixels according to the risk level using classifier 3403, and the landslide sensitivity map predicted by the geographic information deep neural network and the corresponding confidence matrix are obtained.
[0059] Preferably, the semantic segmentation network 3401 is composed of DeepLabV3+ with cascaded softmax layers; the CNN 3402 is an ASPP module based on ResNet feature extractors, consisting of four convolutional layers of different scales (one 1×1 convolutional layer, three 3×3 dilated convolutional layers with rates of 6, 12, and 18 respectively) and pooling layers, which are then concatenated and upsampled to restore the original image size. Figure 6 As shown; the classifier 3403 is a K-means network.
[0060] To demonstrate the effectiveness of image segmentation, this embodiment conducted experiments using a subset of image data, and the results are as follows: Figure 7 As shown, the masked segmentation map obtained by the semantic segmentation network 3401 is almost identical to the human-annotated real segmentation map.
[0061] Furthermore, combined with Figure 3 The specific processing flow of the physical information deep neural network is as follows:
[0062] (2-1) Use the digital elevation model 3404 to extract contour lines from high-resolution satellite image data to obtain a contour map;
[0063] (2-2) Extracting topographic factors from the contour map pixel by pixel, which consist of physical quantities such as slope, aspect, curvature, and topographic humidity index;
[0064] (2-3) Establish a statistical model 3405 for multiple regression analysis of topographic factors, and use the statistical model 3405 to calculate the probability value of landslide risk for each pixel to obtain a statistical landslide sensitivity map.
[0065] (2-4) The statistical landslide sensitivity map is classified into individual pixels according to the risk level using classifier 3406, and the landslide sensitivity map predicted by the physical information deep neural network and the corresponding confidence matrix are obtained.
[0066] Furthermore, the slope described in step (2-2) Slope aspect curvature Topographic humidity index Calculated by the raster generation tool in the digital elevation model, where, for The height corresponding to the pixel for The upstream catchment area along the contour line.
[0067] Preferably, the statistical model 3405 is logistic regression, and the corresponding probability model is... ,in, It is the sigmoid activation function. , , , , These are the hyperparameters to be trained.
[0068] Furthermore, the training described in step S3 specifically involves: using the cross-entropy loss function and combining manually labeled historical data, to fine-tune the parameters of the semantic segmentation network 3401, CNN 3402, and classifier 1 3403 in the geographic information deep neural network, as well as the statistical model 3405 and classifier 2 3406 in the physical information deep neural network.
[0069] Furthermore, combined with Figure 4 The specific working principle of the multi-criteria filter 3407 is as follows:
[0070] (3-1) The landslide sensitivity map and the corresponding confidence matrix predicted by the deep neural network of geographic information, and the landslide sensitivity map and the corresponding confidence matrix predicted by the deep neural network of physical information are used as inputs, and each pixel is input into the multi-criteria filter 3407.
[0071] (3-2) When the landslide sensitivity map predicted by the geographic information deep neural network at a pixel is consistent with the classification result of the landslide sensitivity map predicted by the physical information deep neural network, and there is a threshold in the corresponding confidence matrix that is greater than the manually set threshold, then the consistency result is directly adopted as the final prediction result of the pixel; otherwise, step (3-3) is executed.
[0072] (3-3) Use classifier 3 to reclassify and predict the pixel to obtain the final prediction result of the pixel.
[0073] Furthermore, the training process described in step S4 involves using the cross-entropy loss function and combining it with manually labeled historical data to optimize the parameters of classifier three in the multi-criteria filter 3407.
[0074] Example 2: For high-resolution satellite images of Dubai from MBRSC [1], accurate landslide risk prediction is provided for the changing terrain. Considering efficiency, and focusing solely on analysis and prediction from the perspective of Geographic Information System (GIS), this invention also provides a multi-criteria depth prediction method for landslide risk based on geographic information. Since this example is highly similar to Example 1, the parts that are the same as in Example 1 will not be repeated here.
[0075] Specifically, this method is implemented based on a landslide risk deep learning prediction system. The landslide risk deep learning prediction system consists of a high-resolution satellite 1, a base station 2 containing a server, and a client computer 3. Satellite 1 and base station 2 are connected wirelessly, and the base station 2 server and client computer 3 are connected via a network. Its architecture is as follows: Figure 1 As shown; wherein, the client computer 3 includes a data acquisition module 31, a data storage module 32, a data preprocessing module 33, and a prediction module 34; combined with Figure 2 The data acquisition module 31 reads high-resolution satellite image data captured by satellite 1 through communication according to the sampling period; the data storage module 32 consists of an acquisition data storage unit 3201 and a prediction data storage unit 3202; the acquisition data storage unit 3201 is memory, connected to the data acquisition module 31 and the data preprocessing module 33 respectively, and is used to store and transmit the high-resolution satellite image data acquired by the data acquisition module 31; the prediction data storage unit 3202 is memory, connected to the prediction module 34, and is used to store and transmit the data fed back by the prediction module 34; the data preprocessing module 33 is connected to the prediction module 34 and is used to preprocess the high-resolution satellite image data; the prediction module 34 is a processor with a geographic information deep neural network in memory, and the geographic information deep neural network is composed of a semantic segmentation network 3401, a CNN 3402, and a classifier 3403 connected in series.
[0076] The specific steps of this method are as follows:
[0077] S1: The data acquisition module 31 transmits commands through the base station 2 to periodically acquire high-resolution satellite image data captured by the satellite 1 and store it in the acquisition data storage unit 3201 in the data storage module 32.
[0078] S2: Data preprocessing module 33 scales the acquired high-resolution satellite image data to... Resolution is normalized. Furthermore, image enhancement can be achieved through processes including random rotation, flipping, and brightness variations.
[0079] S3: Prediction module 34 builds a geographic information deep neural network and uses historical data to complete the training of the geographic information deep neural network.
[0080] S4: Prediction module 34 processes high-resolution satellite image data based on a trained geographic information deep neural network to obtain a landslide sensitivity map.
[0081] Example 3: For high-resolution satellite images of Dubai from MBRSC [1], accurate landslide risk prediction is provided for the constantly changing terrain. Considering efficiency, and focusing solely on analysis and prediction from the perspective of physical information, this invention also provides a multi-criteria depth prediction method for landslide risk based on physical information. Since this example is highly similar to Example 1, the parts that are the same as in Example 1 will not be repeated here.
[0082] Specifically, this method is implemented based on a landslide risk deep learning prediction system. The landslide risk deep learning prediction system consists of a high-resolution satellite 1, a base station 2 containing a server, and a client computer 3. Satellite 1 and base station 2 are connected wirelessly, and the base station 2 server and client computer 3 are connected via a network. Its architecture is as follows: Figure 1 As shown; wherein, the client computer 3 includes a data acquisition module 31, a data storage module 32, a data preprocessing module 33, and a prediction module 34; combined with Figure 2The data acquisition module 31 reads high-resolution satellite image data captured by satellite 1 via communication according to the sampling period; the data storage module 32 consists of an acquisition data storage unit 3201 and a prediction data storage unit 3202; the acquisition data storage unit 3201 is memory, connected to the data acquisition module 31 and the data preprocessing module 33 respectively, and is used to store and transmit the high-resolution satellite image data acquired by the data acquisition module 31; the prediction data storage unit 3202 is memory, connected to the prediction module 34, and is used to store and transmit the data fed back by the prediction module 34; the data preprocessing module 33 is connected to the prediction module 34 and is used to preprocess the high-resolution satellite image data; the prediction module 34 is a processor with a physical information deep neural network in memory, the physical information deep neural network is composed of a digital elevation model 3404, a statistical model 3405, and a classifier 3406 connected in series, and is used to predict the landslide risk classification and confidence level corresponding to the image pixels; the statistical model 3405 is a neural network for multiple regression analysis, used to statistically analyze the probabilistic relationship between topographic factors and landslide risk.
[0083] The specific steps of this method are as follows:
[0084] S1: The data acquisition module 31 transmits commands through the base station 2 to periodically acquire high-resolution satellite image data captured by the satellite 1 and store it in the acquisition data storage unit 3201 in the data storage module 32.
[0085] S2: Data preprocessing module 33 scales the acquired high-resolution satellite image data to... Resolution is normalized. Furthermore, image enhancement can be achieved through processes including random rotation, flipping, and brightness variations.
[0086] S3: Prediction module 34 builds a physical information deep neural network and uses historical data to complete the training of the physical information deep neural network.
[0087] S4: Prediction module 34 processes high-resolution satellite image data based on a trained physical information deep neural network to obtain a landslide sensitivity map.
[0088] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.
Claims
1. A multi-criteria depth prediction method for landslide risk integrating geographic and physical information, characterized in that, Includes the following steps: S1: The data acquisition module (31) periodically acquires high-resolution satellite image data captured by the satellite (1) and stores it in the acquisition data storage unit (3201) in the data storage module (32). S2: The data preprocessing module (33) scales and normalizes the acquired high-resolution satellite image data; S3: The prediction module (34) builds a deep neural network for geographic and physical information, and uses historical data to complete the independent training of the deep neural network for geographic information and the deep neural network for physical information respectively; S4: The prediction module (34) uses historical data to train the multi-criteria filter (3407) again based on the trained geographic information deep neural network and physical information deep neural network. S5: The prediction module (34) uses a trained deep neural network of geographic and physical information to process high-resolution satellite image data to obtain a landslide sensitivity map; The processing flow of the geographic information deep neural network is as follows: (1-1) Mask encoding is performed on the high-resolution satellite image data to obtain a mask image; (1-2) The mask image is segmented using a semantic segmentation network (3401) to obtain a mask segmentation map; (1-3) The mask segmentation map is predicted using a CNN (3402) to obtain a sampled landslide sensitivity map; (1-4) Each pixel value in the sampled landslide sensitivity map is defined as the corresponding landslide risk probability value using classifier one (3403), and the risk level is classified pixel by pixel to output the landslide sensitivity map and the confidence matrix corresponding to the landslide sensitivity map. The processing flow of the physical information deep neural network is as follows: (2-1) Use the digital elevation model (3404) to extract contour lines from high-resolution satellite image data to obtain a contour map; (2-2) Extract the slope, aspect, curvature, and topographic humidity index of the contour map pixel by pixel; (2-3) Based on the extracted topographic factors, use the statistical model (3405) to calculate the landslide risk probability pixel by pixel to obtain a statistical landslide sensitivity map; (2-4) Use classifier two (3406) to classify each pixel value in the statistical landslide sensitivity map according to the risk level pixel by pixel, and output the landslide sensitivity map and the confidence matrix corresponding to the landslide sensitivity map; The specific working principle of the multi-criteria filter (3407) is as follows: (3-1) The landslide sensitivity map predicted by the geographic information deep neural network and the corresponding confidence matrix, the landslide sensitivity map predicted by the physical information deep neural network and the corresponding confidence matrix are used as inputs, and each pixel is input into the multi-criteria filter (3407); (3-2) When the classification results of the landslide sensitivity map predicted by the geographic information deep neural network and the landslide sensitivity map predicted by the physical information deep neural network are consistent, and there is a threshold in the corresponding confidence matrix that is greater than the manually set threshold, then the consistency result is directly adopted as the final prediction result of the pixel; otherwise, step (3-3) is executed; (3-3) The pixel is reclassified and predicted using classifier three to obtain the final prediction result of the pixel.
2. The multi-criteria deep prediction method for landslide risk that integrates geographic and physical information according to claim 1, characterized in that the semantic segmentation network (3401) is composed of DeepLabV3+ with cascaded softmax layers; the CNN (3402) is composed of a ResNet-based feature extractor cascaded with an ASPP module consisting of multiple convolutional layers and pooling layers of different scales; and the classifier (3403) is a K-means network.
3. The multi-criteria depth prediction method for landslide risk integrating geographic and physical information according to claim 1, characterized in that, The slope described in step (2-2) Slope aspect curvature Topographic humidity index Calculated by the raster generation tool in the digital elevation model, where, for The height corresponding to the pixel for The upstream catchment area along the contour line.
4. The multi-criteria depth prediction method for landslide risk integrating geographic and physical information according to claim 1, characterized in that, The statistical model (3405) is one of logistic regression, random forest or DNN.
5. The multi-criteria depth prediction method for landslide risk integrating geographic and physical information according to claim 1, characterized in that, The training described in step S3 specifically involves: using the cross-entropy loss function and combining it with manually labeled historical data, to fine-tune the parameters of the semantic segmentation network (3401), CNN (3402), classifier one (3403) in the geographic information deep neural network, and the statistical model (3405) and classifier two (3406) in the physical information deep neural network; the training process described in step S4 involves using the cross-entropy loss function and combining it with manually labeled historical data to fine-tune the parameters of classifier three in the multi-criteria filter (3407).
6. The multi-criteria depth prediction method for landslide risk integrating geographic and physical information according to claim 1, characterized in that, Both the geographic information deep neural network and the physical information deep neural network can be used individually to draw landslide sensitivity maps.
7. A landslide risk deep learning prediction system applied to the multi-criteria deep prediction method for landslide risk that integrates geographic and physical information as described in any one of claims 1 to 6, characterized in that, The system consists of a high-resolution satellite (1), a base station (2) containing a server, and a client (3) computer. The satellite (1) and the base station (2) are connected via wireless communication, and the base station (2) server and the client (3) computer are connected via a network. The client (3) computer includes a data acquisition module (31), a data storage module (32), a data preprocessing module (33), and a prediction module (34). The data acquisition module (31) reads high-resolution satellite image data captured by the satellite (1) through communication according to the sampling period. The data storage module (32) consists of an acquisition data storage unit (3201) and a prediction data storage unit (3202). The acquisition data storage unit (3201) is memory, which is connected to the data acquisition module (31) and the data preprocessing module (33) respectively, and is used to store and transmit the high-resolution satellite image data acquired by the data acquisition module (31). The prediction data storage unit (3202) is memory, which is connected to the prediction module (34) and is used to store and transmit the data fed back by the prediction module (34). The data preprocessing module (33) is connected to the prediction module (34) and is used to preprocess high-resolution satellite image data. The prediction module (34) is a processor with a deep neural network for geographic and physical information in memory. The deep neural network for geographic and physical information consists of a multi-criteria filter (3407) connected in parallel between the deep neural network for geographic information and the deep neural network for physical information. The deep neural network for geographic information consists of a semantic segmentation network (3401), a CNN (3402), and a classifier (3403) connected in series. The deep neural network for physical information consists of a digital elevation model (3404), a statistical model (3405), and a classifier (3406) connected in series. It is used to predict the landslide risk classification and confidence level corresponding to the image pixels. The statistical model (3405) is a neural network for multiple regression analysis and is used to statistically analyze the probability relationship between topographic factors and landslide risk. The multi-criteria filter (3407) is a classifier used to evaluate and select the prediction results of the deep neural network for geographic information and the deep neural network for physical information.