Landslide prediction and early warning method based on deep learning

Abstract

The invention belongs to the field of landslide early warning and risk management and control research, and particularly discloses a landslide prediction and early warning method based on deep learning, and the method comprises the following steps: dividing rainfall events, marking positive events and negative events, and dividing historical landslide data into an initial landslide set and a newly added landslide set; generating an initial landslide susceptibility map; updating the initial landslide susceptibility map based on the newly added landslide set, and generating an updated landslide susceptibility map; constructing a deep learning sample set; using a deep learning model to train and predict the comprehensive sample set; wherein the deep learning model comprises a convolutional neural network and a long short-term memory network, the convolutional neural network is used for extracting deep disaster-pregnant environment features in multi-dimensional static space data, the long short-term memory network is used for extracting dynamic time sequence features, and after the two features are fused, the landslide probability is output through the classifier. According to the method, high-precision and dynamically-updated rainfall type landslide probability prediction can be realized, and direct decision support is provided for landslide early warning.

Description

Technical Field

[0001] This application belongs to the research field of landslide early warning and risk management, and more specifically, it relates to a landslide prediction and early warning method based on deep learning. Background Technology

[0002] In the field of landslide early warning and risk management research, rainfall is a major factor triggering landslides; therefore, establishing accurate landslide early warning models is crucial for disaster prevention and mitigation. With the development of artificial intelligence technology, more and more researchers are beginning to use deep learning methods for landslide prediction and early warning. In existing technologies, some researchers have used soil moisture, pore water pressure, and rainfall monitoring data collected from landslide-prone slopes to build deep learning models. These models can make accurate landslide predictions based on monitoring data and can capture the moisture response to heavy rainfall events. Other researchers have proposed a deep learning system for short-term prediction of rainfall-induced shallow landslides. This system is based on multiple neural network models and uses a voting scheme to obtain a binary landslide incidence prediction from each lag period and its set. These research findings have played an important role in large-scale disaster prediction and testing.

[0003] However, existing deep learning methods still have some shortcomings in landslide early warning. Most models rely on static landslide susceptibility maps, which cannot be dynamically updated according to newly occurring disasters, resulting in poor adaptability. Secondly, the models do not adequately integrate the spatial characteristics of the geological environment with the temporal characteristics of rainfall processes, making it difficult to accurately predict landslides.

[0004] Therefore, how to achieve high-precision, dynamically updated probability prediction of rainfall-induced landslides is an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a landslide prediction and early warning method based on deep learning, which can achieve high-precision, dynamically updated probability prediction of rainfall-induced landslides, and provide direct decision support for landslide early warning.

[0006] To achieve the above objectives, firstly, this application provides a landslide prediction method based on deep learning, comprising the following steps: S10, acquire multi-source data of the study area, including geological and topographical spatial data, land cover classification data, historical landslide data, and rainfall data; S20, based on the rainfall data and historical landslide data, rainfall events are divided and labeled as positive and negative events. At the same time, the historical landslide data is divided into an initial landslide set and a newly added landslide set. Among them, positive events are rainfall events that induce landslides, and negative events are rainfall events that do not induce landslides. S30, Based on the geological and topographical spatial data and the initial landslide set, an initial landslide susceptibility map is generated; Based on the newly added landslide set, the initial landslide susceptibility map is updated by calculating the Euclidean distance between each grid cell and the newly added landslide point, and an updated landslide susceptibility map is generated. S40. Based on the updated landslide susceptibility map and rainfall events, a deep learning sample set is constructed. The regions with susceptibility values ​​higher than the first preset threshold and all historical landslide locations in the updated landslide susceptibility map are defined as the first positive samples, and the regions with susceptibility values ​​lower than the second preset threshold are defined as the first negative samples. The first positive and negative samples and positive and negative events are combined to form a comprehensive sample set. S50, a deep learning model is used to train and predict the comprehensive sample set; wherein, the deep learning model includes a convolutional neural network branch and a long short-term memory network branch. The convolutional neural network branch is used to process the geological and topographical spatial data to extract deep disaster-prone environmental features from its multidimensional static spatial data. The long short-term memory network branch is used to process the rainfall event sequence to capture dynamic temporal features containing cumulative effects and intensity changes during the rainfall process. After fusing these two features, the landslide probability is output through a classifier.

[0007] The landslide prediction method based on deep learning provided in this application has the following effects: By dynamically updating the landslide susceptibility map, the susceptibility base map can be adaptively adjusted according to newly occurring disaster data, overcoming the deficiency of static base maps that cannot be dynamically updated, and improving the model's adaptability to changes in the geological environment; at the same time, a dual-branch architecture of convolutional neural networks and long short-term memory networks is adopted to deeply extract spatial static disaster-causing factors and temporal dynamic triggering factors, and effectively fuse them, thereby accurately capturing the complex causal mechanism of rainfall-induced landslides determined by both spatial characteristics and temporal processes; this synergistic effect of dynamic updating and spatiotemporal feature fusion can achieve high-precision, dynamically updatable probability prediction of rainfall-induced landslides, providing direct decision support for landslide early warning.

[0008] As a further preferred embodiment, in step S10, the multi-source data includes digital elevation models, slope, aspect, curvature, lithology distribution, geological structure, sedimentary layer thickness, water system, road network, land cover classification data, historical landslide logging data, and rainfall data.

[0009] As a further preferred option, step S20 specifically includes: The shortest drought period method is used to objectively divide long-term rainfall data, identify and define independent rainfall events; The segmented rainfall events are labeled as follows: rainfall events that match historical landslide data in terms of time are defined as positive events, while rainfall events that do not induce landslides but have certain disaster potential are defined as negative events by random sampling and excluding all known landslide occurrence periods; Historical landslide data is split along the time axis into an initial landslide set and a newly added landslide set.

[0010] As a further preferred option, step S30 specifically includes: Using the raster as the basic evaluation unit, statistical methods such as variance inflation factor are first used to perform collinearity diagnosis on the geological and topographical spatial data in order to screen and retain a set of mutually independent evaluation factors. The initial landslide set was used as the second positive sample, and the second negative sample was generated by sampling non-landslide areas in the study area according to a preset ratio. The first positive event and the first negative event were combined to form a dataset. Machine learning algorithms were used to train and predict the dataset, thereby generating an initial landslide susceptibility map covering the entire study area. After obtaining the newly added landslide set, the normalized spatial distance index method is introduced to update the initial susceptibility map. Based on the spatial clustering characteristics of landslides, this method quantifies the impact of new disaster events on the susceptibility of the surrounding environment by calculating the Euclidean distance between any grid cell in the study area and all newly added landslide points, thereby generating an updated landslide susceptibility map that reflects the latest disaster information.

[0011] As a further preferred embodiment, the formula for calculating the normalized spatial distance index is:

[0012]

[0013] In the formula, It is the normalized distance exponent; It's a new landslide; It is an evaluation unit; It is the Euclidean distance between the new landslide and the evaluation unit; i For indexing newly added landslide points; j This is the index for the evaluation unit.

[0014] As a further preferred option, in step S40, the first positive and negative samples and positive and negative events are combined in a Cartesian fashion to form a comprehensive sample set, wherein the combination of positive events and the first positive sample is the final positive sample, and other combinations are negative samples.

[0015] As a further preferred embodiment, in step S50, the deep disaster-prone environment features and dynamic temporal features are fused through splicing and a fully connected layer.

[0016] As a further preferred option, in step S50, the classifier is a Softmax classifier.

[0017] As a further preferred embodiment, in step S50, the landslide probability is used to generate a spatial distribution map of the landslide probability.

[0018] Secondly, this application provides a landslide early warning method based on deep learning, used to provide landslide early warning based on the landslide probability predicted by the landslide prediction method based on deep learning as described above.

[0019] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0020] Figure 1 This is a flowchart of the deep learning-based landslide prediction method provided in this application; Figure 2 This is a technical framework diagram of the landslide prediction method based on deep learning provided in the embodiments of this application; Figure 3 These are comparison diagrams of landslide susceptibility maps provided in the embodiments of this application; wherein, (a) is the initial landslide susceptibility map, (b) is the landslide susceptibility map after susceptibility update, (c) is a detail comparison diagram, and (d) is a diagram showing the changes before and after susceptibility update. Figure 4 This is a comparison chart of landslide prediction results provided in the embodiments of this application; wherein, (a) is the landslide prediction result for the event on June 15, 2011, (b) is the result of the rainfall event on June 29, 2021 using the susceptibility update mechanism, and (c) is the result of the rainfall event on June 29, 2021 without using the susceptibility update mechanism. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] like Figure 1 As shown, this application provides a landslide prediction method based on deep learning, including steps S10 to S50, which are detailed below: Step S10 involves acquiring multi-source data for the study area, including geological and topographical spatial data, land cover classification data, historical landslide data, and rainfall data. This step provides a comprehensive data foundation for landslide prediction, ensuring the accuracy and completeness of the data.

[0023] Step S20: Based on rainfall data and historical landslide data, rainfall events are classified and labeled as positive and negative events. At the same time, historical landslide data are divided into an initial landslide set and a newly added landslide set. Among them, positive events are rainfall events that induce landslides, and negative events are rainfall events that do not induce landslides.

[0024] The step S20 provided in this application divides rainfall events and labels them as positive and negative based on rainfall data and historical landslide data. At the same time, it divides historical landslide data into an initial landslide set and a newly added landslide set, which can provide accurate sample data for model training and updating, enabling the model to better learn the spatiotemporal characteristics of landslide occurrence.

[0025] Step S30: Based on geological and topographical spatial data and the initial landslide set, generate an initial landslide susceptibility map; based on the newly added landslide set, update the initial landslide susceptibility map by calculating the Euclidean distance between each grid cell and the newly added landslide point, and generate an updated landslide susceptibility map.

[0026] Step S30 provided in this application, which generates an initial landslide susceptibility map and updates the initial landslide susceptibility map by adding a new landslide set, enables the model to dynamically adapt to changes in the geological environment, thereby improving the model's adaptability and prediction accuracy.

[0027] Step S40: Based on the updated landslide susceptibility map and rainfall events, a deep learning sample set is constructed. The regions with susceptibility values ​​higher than a first preset threshold and all historical landslide locations in the updated landslide susceptibility map are defined as first positive samples, and the regions with susceptibility values ​​lower than a second preset threshold are defined as first negative samples. The first positive and negative samples and positive and negative events are combined to form a comprehensive sample set.

[0028] Step S50: Use a deep learning model to train and predict the comprehensive sample set.

[0029] The deep learning model includes a convolutional neural network branch and a long short-term memory network branch. The convolutional neural network branch is used to process geological and topographical spatial data to extract deep disaster-prone environmental features from its multidimensional static spatial data. The long short-term memory network branch is used to process rainfall event sequences to capture dynamic temporal features containing cumulative effects and intensity changes during rainfall. After fusing these two features, the landslide probability is output through a classifier.

[0030] The landslide prediction method based on deep learning provided in this application has the following effects: By dynamically updating the landslide susceptibility map, the susceptibility base map can be adaptively adjusted according to newly occurring disaster data, overcoming the deficiency of static base maps that cannot be dynamically updated, and improving the model's adaptability to changes in the geological environment; at the same time, a dual-branch architecture of convolutional neural networks and long short-term memory networks is adopted to deeply extract spatial static disaster-causing factors and temporal dynamic triggering factors, and effectively fuse them, thereby accurately capturing the complex causal mechanism of rainfall-induced landslides determined by both spatial characteristics and temporal processes; this synergistic effect of dynamic updating and spatiotemporal feature fusion can achieve high-precision, dynamically updatable probability prediction of rainfall-induced landslides, providing direct decision support for landslide early warning.

[0031] In one embodiment, the technical solution to achieve the above objective can be as follows: This embodiment proposes a method for predicting the probability of rainfall-induced landslides by integrating dynamic susceptibility updates and a spatiotemporal deep learning model. This method does not require high-precision geotechnical parameters and can self-optimize based on the latest disaster data to achieve large-area, high-resolution dynamic probability prediction.

[0032] Specifically, the framework of the landslide early warning method provided in this embodiment is as follows: Figure 2 As shown, the framework is mainly divided into 5 modules: Module I is data preparation; Module II is the division of rainfall events and landslide data; Module III is the initial landslide susceptibility assessment; Module IV is the landslide susceptibility update and deep learning sample set construction; Module V is the spatiotemporal probability prediction of landslides based on deep learning.

[0033] Module I: Data Preparation This module aims to systematically collect and preprocess multi-source heterogeneous data required for the research. First, it acquires basic geological and topographical spatial data covering the study area, specifically including digital elevation models (DEMs) and their derived slope, aspect, and curvature data, as well as vector or raster data on lithology distribution, geological structures (such as faults), sedimentary thickness, drainage systems, and road networks. Second, it obtains land cover classification data for the study area through the interpretation of remote sensing imagery. Simultaneously, it organizes historical disaster data to establish a landslide logging database containing crucial information such as the precise geographical location and occurrence time of landslides. Finally, it collects long-term precipitation data recorded by meteorological stations within and around the study area, with a temporal accuracy down to the hour or day.

[0034] Module II: Rainfall Events and Landslide Data Classification The core task of this module is to construct an event and sample database for model training and updates. First, the shortest drought period method is applied to objectively segment long-term rainfall records, thereby identifying and defining independent rainfall events. Second, the segmented rainfall events are labeled: rainfall events that coincide with historical landslide records in the time dimension are defined as positive events (i.e., disaster-inducing rainfall), while rainfall events that do not induce landslides but have a certain disaster-causing potential are defined as negative events by randomly sampling and excluding all known landslide occurrence periods. Finally, to achieve dynamic model updates, the historical landslide catalog is split along the time axis, forming at least two independent subsets, such as an initial landslide set for building the baseline model and a new landslide set for subsequent model iterations and optimizations.

[0035] Module III: Initial Landslide Susceptibility Assessment and Landslide Susceptibility Update This module aims to generate a spatial distribution map of basic landslide susceptibility in the study area. The process uses a raster as the basic evaluation unit and first employs statistical methods such as variance inflation factor to diagnose collinearity of the basic geological and topographical factors prepared in step one, screening and retaining a set of independent evaluation factors to ensure model robustness. Subsequently, the initial landslide set defined in step two is used as positive samples, and negative samples are generated by sampling non-landslide areas within the study area at a predetermined ratio (e.g., 1:1), jointly constructing a dataset for model training. Finally, random forest or other comparable machine learning algorithms are used to train and predict on this dataset, thereby generating an initial landslide susceptibility map covering the entire study area. This module is the core component for achieving dynamic adaptation of the model. After obtaining a new landslide set, the normalized spatial distance index (Equation 1) is introduced to update the initial susceptibility map. Based on the spatial clustering characteristics of landslides, this method quantifies the impact of new disaster events on the susceptibility of the surrounding environment by calculating the Euclidean distance between any grid cell in the study area and all newly added landslide points. The closer the distance, the greater the increase in susceptibility, thereby generating an updated landslide susceptibility map that reflects the latest disaster information.

[0036]

[0037]

[0038] in, It is the normalized distance exponent; It's a new landslide; It is an evaluation unit; It is the Euclidean distance between the new landslide and the evaluation unit; i For indexing newly added landslide points; j This is the index for the evaluation unit.

[0039] Module IV: Construction of Deep Learning Sample Sets First, a sample set required for the deep learning model is constructed based on the updated susceptibility map. Specifically, "extremely high" susceptibility areas and all historical landslide locations are jointly defined as positive landslide samples, and "extremely low" susceptibility areas are defined as negative landslide samples. Finally, these spatial samples are combined with the rainfall event samples (positive / negative events) defined in step two using a Cartesian combination to form a comprehensive spatiotemporal dataset containing four types (TP, FP, TN, FN). Only the combination of high-susceptibility areas under induced rainfall (TP) is used as the final positive sample for model training, and all other combinations are used as negative samples.

[0040] Module V: Deep Learning-Based Landslide Probability Prediction This step aims to achieve dynamic prediction of landslide spatiotemporal probability using a specially designed deep learning model. The model employs an innovative dual-branch parallel architecture, integrating a convolutional neural network (CNN) and a long short-term memory (LSTM) network. A one-dimensional CNN branch is specifically designed to process and extract deep-seated disaster-prone environmental features from multi-dimensional static spatial data, including topography and geology. Simultaneously, a stacked LSTM network branch is designed to process rainfall event sequence data, accurately capturing dynamic temporal features such as cumulative effects and intensity changes during rainfall. The feature vectors extracted by the two branches are then fed into a spatiotemporal feature fusion module, deeply integrated through concatenation and fully connected layers. Finally, the model uses a Softmax classifier to calculate the instability probability of each grid cell under a specific rainfall event, outputting a probability value between 0 and 1. This model can be deployed on servers or computer systems, receiving real-time weather forecast data as input, and can dynamically generate and output a high-resolution (e.g., 15-meter) spatial distribution map of landslide probability covering the entire study area within a specific future time period (e.g., 24 hours), providing direct decision support for landslide early warning.

[0041] The key technical point of this embodiment is: 1. Susceptibility-based dynamic update mechanism: The susceptibility-based update mechanism enables the model's base map to evolve with the occurrence of new disaster events, effectively improving the model's adaptability to changes in the geological environment and the accuracy of long-term predictions.

[0042] 2. Spatiotemporal decoupling deep learning architecture: An original CNN-LSTM dual-branch model structure was designed. This structure can extract the spatial static disaster-causing factors (processed by CNN) and the temporal dynamic triggering factors (processed by LSTM) of landslide occurrence separately and deeply, and then effectively fuse them to accurately capture the complex causal mechanism of rainfall-induced landslides determined by both spatial features and temporal processes.

[0043] 3. Data-driven, efficient prediction paradigm: This embodiment eliminates the reliance on high-precision, difficult-to-obtain geotechnical parameters required by traditional physical models, and is entirely data-driven. By fusing multi-source heterogeneous data, it achieves large-scale, high-resolution, and high-precision dynamic landslide probability prediction, demonstrating strong practicality and scalability.

[0044] The following is a specific implementation example of this application: This embodiment was verified in Kecheng District, Quzhou City, Zhejiang Province. Through retrospective analysis of historical heavy rainfall events and comparison with models that did not implement the key technologies of this embodiment, the results show that this embodiment has achieved significant technical effects and superior early warning performance.

[0045] First, this embodiment demonstrates high-precision and high-generalization early warning performance. This was verified by analyzing two independent heavy rainfall events ten years apart (the event on June 15, 2011, and the event on June 29, 2021). Figure 3 As shown, the prediction accuracy (ACC) of the method in this embodiment exceeded 83% in both events, and the mean absolute error (MAE) was less than 0.16. This result fully demonstrates that the method has the stability and robustness for long-term application and can serve as a reliable landslide disaster early warning tool.

[0046] Secondly, comparative analysis confirms the decisive role of the key technologies in this embodiment. Taking the rainfall event of June 29, 2021 as an example, a performance comparison is made between models implementing the key technologies of this embodiment and those not implementing these technologies. Figure 4 As shown in the figure, the model without the update strategy achieved an accuracy of only 17%, while the accuracy increased significantly to 83.33% after implementing the strategy of this embodiment. For the six landslides that actually occurred in this event, the predicted probability given by the method of this embodiment was on average 0.16 higher than that of the unupdated model. This comparative result strongly demonstrates that the susceptibility dynamic update mechanism of this embodiment is the core of achieving high-precision prediction, effectively correcting risk assessment, and significantly improving the performance of the early warning model.

[0047] Finally, this embodiment achieves true dynamic spatiotemporal prediction. The model can generate corresponding landslide probability maps with significantly different spatial distributions based on rainfall event inputs with different spatiotemporal characteristics. The distribution of high-risk areas is highly coupled with the rainfall event, proving that this method is a dynamic hazard assessment model that is highly sensitive to triggering conditions, rather than a disguised static susceptibility model.

[0048] In summary, this embodiment overcomes many shortcomings of the prior art and provides a new paradigm for high-precision, dynamically updated rainfall-induced landslide probability prediction. It provides strong technical support for regional geological disaster meteorological early warning and risk management, and has significant socio-economic benefits and broad application prospects.

[0049] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A landslide prediction method based on deep learning, characterized in that, Includes the following steps: S10, acquire multi-source data of the study area, including geological and topographical spatial data, land cover classification data, historical landslide data, and rainfall data; S20, based on the rainfall data and historical landslide data, rainfall events are divided and labeled as positive and negative events. At the same time, the historical landslide data is divided into an initial landslide set and a newly added landslide set. Among them, positive events are rainfall events that induce landslides, and negative events are rainfall events that do not induce landslides. S30, Based on the geological and topographical spatial data and the initial landslide set, an initial landslide susceptibility map is generated; Based on the newly added landslide set, the initial landslide susceptibility map is updated by calculating the Euclidean distance between each grid cell and the newly added landslide point, and an updated landslide susceptibility map is generated. S40. Based on the updated landslide susceptibility map and rainfall events, a deep learning sample set is constructed. The regions with susceptibility values ​​higher than the first preset threshold and all historical landslide locations in the updated landslide susceptibility map are defined as the first positive samples, and the regions with susceptibility values ​​lower than the second preset threshold are defined as the first negative samples. The first positive and negative samples and positive and negative events are combined to form a comprehensive sample set. S50, a deep learning model is used to train and predict the comprehensive sample set; wherein, the deep learning model includes a convolutional neural network branch and a long short-term memory network branch. The convolutional neural network branch is used to process the geological and topographical spatial data to extract deep disaster-prone environmental features from its multidimensional static spatial data. The long short-term memory network branch is used to process the rainfall event sequence to capture dynamic temporal features containing cumulative effects and intensity changes during the rainfall process. After fusing these two features, the landslide probability is output through a classifier.

2. The landslide prediction method based on deep learning as described in claim 1, characterized in that, In step S10, the multi-source data includes digital elevation model, slope, aspect, curvature, lithology distribution, geological structure, sedimentary layer thickness, water system, road network, land cover classification data, historical landslide logging data, and rainfall data.

3. The landslide prediction method based on deep learning as described in claim 1, characterized in that, Step S20 is as follows: The shortest drought period method is used to objectively divide long-term rainfall data, identify and define independent rainfall events; The segmented rainfall events are labeled as follows: rainfall events that match historical landslide data in terms of time are defined as positive events, while rainfall events that do not induce landslides but have certain disaster potential are defined as negative events by random sampling and excluding all known landslide occurrence periods; Historical landslide data is split along the time axis into an initial landslide set and a newly added landslide set.

4. The landslide prediction method based on deep learning as described in claim 1, characterized in that, Step S30 is as follows: Using the raster as the basic evaluation unit, statistical methods such as variance inflation factor are first used to perform collinearity diagnosis on the geological and topographical spatial data in order to screen and retain a set of mutually independent evaluation factors. The initial landslide set was used as the second positive sample, and the second negative sample was generated by sampling non-landslide areas in the study area according to a preset ratio. The first positive event and the first negative event were combined to form a dataset. Machine learning algorithms were used to train and predict the dataset, thereby generating an initial landslide susceptibility map covering the entire study area. After obtaining the newly added landslide set, the normalized spatial distance index method is introduced to update the initial susceptibility map. Based on the spatial clustering characteristics of landslides, this method quantifies the impact of new disaster events on the susceptibility of the surrounding environment by calculating the Euclidean distance between any grid cell in the study area and all newly added landslide points, thereby generating an updated landslide susceptibility map that reflects the latest disaster information.

5. The landslide prediction method based on deep learning as described in claim 4, characterized in that, The formula for calculating the normalized spatial distance index is as follows: In the formula, It is the normalized distance exponent; It's a new landslide; It is an evaluation unit; It is the Euclidean distance between the new landslide and the evaluation unit; i For indexing newly added landslide points; j This is the index for the evaluation unit.

6. The landslide prediction method based on deep learning as described in claim 1, characterized in that, In step S40, the first positive and negative samples and positive and negative events are combined in a Cartesian fashion to form a comprehensive sample set. The combination of a positive event and the first positive sample is the final positive sample, and the other combinations are negative samples.

7. The landslide prediction method based on deep learning as described in claim 1, characterized in that, In step S50, the deep disaster-prone environment features and dynamic temporal features are fused through splicing and a fully connected layer.

8. The landslide prediction method based on deep learning as described in claim 1, characterized in that, In step S50, the classifier is a Softmax classifier.

9. The landslide prediction method based on deep learning as described in claim 1, characterized in that, In step S50, the landslide probability is used to generate a spatial distribution map of the landslide probability.

10. A landslide early warning method based on deep learning, characterized in that, The landslide probability predicted by the deep learning-based landslide prediction method according to any one of claims 1 to 9 is used for landslide early warning.

Images