Landslide susceptibility prediction method under future scene

Abstract

The invention relates to the field of geological disaster prediction, and discloses a landslide susceptibility prediction method in a future scenario, and the method comprises the steps: selecting a landslide influence factor according to a disaster-pregnant environment and disaster-causing factors of a landslide in a research area; after the landslide influence factors are preprocessed, a data set is constructed; constructing a landslide prediction model, and training and verifying the landslide prediction model by using the data set to obtain a model with optimal parameters; replacing an LULC factor and a rainfall factor in the landslide influence factors with an LULC factor simulated by a PLUS model and a rainfall factor simulated by an atmospheric circulation model respectively; and inputting other factors in the simulated LULC factor, the simulated rainfall factor and the landslide influence factor into the model with the optimal parameters to obtain a future scene landslide susceptibility prediction result. The method has the advantages that the defect that landslide susceptibility prediction is conducted only by simulating a single landslide influence factor is overcome, and the prediction result is more reliable.

Description

Technical Field

[0001] The present invention relates to the field of geological disaster prediction, and in particular to a method for predicting landslide susceptibility in future scenarios. Background Art

[0002] Landslide susceptibility refers to the spatial probability of a landslide occurring in a certain area under specific geological conditions. This concept is based on the assumption that landslides tend to occur in areas with similar conditions to those in which historical landslides occurred. Predicting landslide susceptibility in future scenarios to understand future landslide-prone areas is an effective means to reduce landslide risks. Land use and land cover (LULC) and rainfall are extremely critical landslide influencing factors in landslide susceptibility, representing the impact of human engineering activities and regional climate on landslides, respectively. Changes in the two factors will significantly affect the spatial pattern of landslide susceptibility. Against the background of increasing human engineering activities and increasing extreme climate events, simulating LULC and rainfall factors in future scenarios, and inputting the simulated factors and other landslide influencing factors into the landslide susceptibility model, and then predicting landslide susceptibility in future scenarios can provide a scientific basis for future regional land planning and disaster prevention strategy formulation.

[0003] Under the idea of ​​predicting landslide susceptibility based on landslide influencing factors in future scenarios, models such as artificial neural networks and cellular automata are used to simulate the changes in LULC in future scenarios, and to predict landslide susceptibility based on LULC factors in future scenarios; and the multi-scenario rainfall simulation data in the International Coupled Model Intercomparison Project 6 (CMIP6) are widely used to represent the spatial distribution of future rainfall, so as to predict the landslide susceptibility distribution considering rainfall changes in future scenarios. Although the above ideas for predicting landslide susceptibility have been successfully applied in many landslide-prone areas around the world, the two factors of LULC and rainfall factors are usually studied independently in current studies, and the research on predicting landslide susceptibility by simulating the changes of these two factors in future scenarios is relatively limited.

[0004] In addition, the feature extraction ability of the landslide susceptibility model has an important impact on the accuracy of landslide susceptibility and directly determines the credibility of the landslide susceptibility prediction results. At present, physical models, statistical models and machine learning models have been used for regional landslide susceptibility prediction, but advanced deep learning models with relatively stronger feature extraction capabilities have not yet been applied in this field. The ViT (Vision Transformer) model is a model based on the Transformer structure. It uses the self-attention mechanism to capture the dependencies and contextual information of different positions in the image sequence, obtains the degree of correlation between each position in the sequence and other positions, and realizes the modeling of global information in the image, so as to better realize landslide feature extraction. The ViT model has been applied in landslide susceptibility tasks and has been proven to have a stronger landslide feature extraction ability than traditional machine learning models, and the accuracy of the obtained landslide susceptibility results is more trustworthy.

[0005] There is a landslide susceptibility assessment method that introduces dynamic changes in land use, patent application number: CN202310427898.X, which uses a logistic regression model to compare and analyze the model prediction effects before and after the introduction of land use change factors. The introduction of land use change factors can improve the accuracy of the model susceptibility assessment results. However, it often only simulates the changes of one landslide influencing factor, and the feature extraction capability of the landslide susceptibility model used is relatively insufficient. Therefore, in order to obtain more reliable landslide susceptibility prediction results in future scenarios, it is necessary to improve the original method of predicting landslide susceptibility by simulating only a single landslide influencing factor and using a model with relatively weak feature extraction capabilities. Summary of the invention

[0006] The purpose of the present invention is to propose a method for predicting landslide susceptibility in future scenarios, so as to solve the technical problem that the existing landslide susceptibility prediction only considers a single influencing factor, resulting in inaccurate landslide prediction for future scenarios.

[0007] A method for predicting landslide susceptibility in future scenarios comprises the following steps:

[0008] S1: Select landslide influencing factors according to the landslide disaster-prone environment and disaster-causing factors in the study area;

[0009] S2: After preprocessing the landslide influencing factors, the data set is constructed;

[0010] S3: constructing a landslide prediction model, using the data set to train and verify the landslide prediction model, and obtaining a model with optimal parameters;

[0011] S4: Replace the LULC factor and rainfall factor in the landslide influencing factor with the LULC factor simulated by the PLUS model and the rainfall factor simulated by the atmospheric circulation model, respectively;

[0012] S5. Input the simulated LULC factor, simulated rainfall factor, and other factors in the landslide influencing factor into the model with the best parameters to obtain the prediction results of landslide susceptibility in future scenarios.

[0013] A storage medium stores instructions and data for implementing a method for predicting the susceptibility of landslides in future scenarios.

[0014] A future scenario landslide susceptibility prediction device comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a future scenario landslide susceptibility prediction method.

[0015] The beneficial effect provided by the present invention is: a method for predicting landslide susceptibility under future scenarios is proposed, which simultaneously considers the impact of changes in LULC and rainfall factors on future landslide susceptibility, making up for the deficiency of predicting landslide susceptibility by only simulating a single landslide influencing factor.

[0016] At the same time, this method uses the ViT model with stronger feature extraction capability to predict landslide susceptibility. The prediction results obtained by this method are more credible than those of physical models, statistical models and machine learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic flow chart of the method of the present invention;

[0018] Figure 2 It is the flow chart of landslide sample and non-landslide sample generation;

[0019] Figure 3 It is a schematic diagram of the ViT model structure;

[0020] Figure 4 It is a detailed process schematic diagram of the method of the present invention;

[0021] Figure 5 It is a working schematic diagram of the hardware device of the present invention. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0023] Before formally describing the present invention, the scheme of the present invention is first generally described for easy understanding.

[0024] Please refer to Figure 1 , Figure 1 It is a schematic flow diagram of the method of the present invention;

[0025] The present invention provides a method for predicting the susceptibility of landslides in future scenarios, comprising the following steps:

[0026] S1: Select landslide influencing factors according to the landslide disaster-prone environment and disaster-causing factors in the study area;

[0027] It should be noted that the landslide influencing factors include: elevation, slope, aspect, lithology, distance from fault, distance from river, normalized difference vegetation index, rainfall factor and LULC factor.

[0028] As an example, firstly, according to the disaster-prone environment and disaster-causing factors of landslides in the study area, elevation, slope, aspect, lithology, distance from faults, distance from rivers, normalized difference vegetation index, rainfall and LULC factor are selected as landslide influencing factors.

[0029] Then, the remote sensing and geographic information source data corresponding to each factor are collected, including digital elevation models, regional geological maps, remote sensing images, meteorological data, and surface cover data.

[0030] Next, ArcMap 10.6 and ENVI 5.3 platforms were used to extract the influencing factors of various landslides. The elevation, slope and aspect factors were obtained based on the digital elevation model; the lithology and distance to fault factors were extracted based on the regional geological map; the normalized difference vegetation index factor was extracted based on the remote sensing image; the rainfall factor was obtained based on the meteorological data; and the distance to the river and LULC factors were extracted based on the surface cover data.

[0031] S2: After preprocessing the landslide influencing factors, the data set is constructed;

[0032] It should be noted that the preprocessing method includes the following operations performed in sequence: coordinate system one, outlier removal, grid resampling and numerical normalization.

[0033] As an embodiment, the present invention performs coordinate system one, outlier removal, grid resampling and value normalization operations on each factor to ensure the unity of the spatial resolution and spatial coordinate system of each landslide influencing factor.

[0034] Please refer to Figure 2 , Figure 2 It is a flow chart of generating landslide samples and non-landslide samples. The process of constructing the data set in the present invention is as follows:

[0035] First, the nine landslide influencing factors are stacked into a three-dimensional matrix (H, W, C), where H represents the height of each factor grid image, W represents the width of each factor grid image, and C represents the number of landslide influencing factors, which is 9 here.

[0036] Then, a landslide pixel is randomly selected in the entire area as the center and label of the landslide sample block, and the sample block is three-dimensionally expanded in the H, W, and C directions to form a landslide sample block with a size of 72×72×9. At the same time, the same number of non-landslide sample blocks are generated according to the same process. The two are combined to form a landslide dataset for training and testing of the landslide susceptibility model.

[0037] Next, on the premise that the ratio of landslide and non-landslide samples in each part remains one to one, the landslide dataset is randomly divided into training set, validation set and test set in a ratio of 6:2:2, that is, 60% of the data is used to train the model and extract landslide features, 20% of the data is used to verify the performance of the model, and 20% of the data is used to test the performance of the model.

[0038] S3: constructing a landslide prediction model, using the data set to train and verify the landslide prediction model, and obtaining a model with optimal parameters;

[0039] It should be noted that the landslide prediction model in the present invention adopts the ViT model. Figure 3 , Figure 3 It is a structural diagram of the landslide prediction model.

[0040] ViT is a deep learning model for natural image classification. Compared with the traditional computer vision model based on convolutional neural network structure, ViT adopts the Transformer architecture.

[0041] ViT uses the self-attention mechanism to capture dependencies and contextual information in image sequences, so that the model can obtain the degree of association between each position in the sequence and other positions, and achieve modeling of global information in the image. ViT does not use the decoder part in the original Transformer, but only uses the encoder part for feature extraction. Unlike traditional convolution-based computer vision models, ViT directly treats the input two-dimensional image data as a one-dimensional sequence for feature extraction.

[0042] Specifically, ViT first cuts the input image into a set of non-overlapping patches of fixed size, each patch is regarded as a token, and the pixels of each patch are flattened into a one-dimensional vector; then, the one-dimensional vector of each patch becomes a higher-dimensional embedding vector through linear mapping (Linear Embedding). At the same time, in order to enable the model to learn the relative position information of each patch in the image, positional encoding (Positional Embedding) is added to represent and retain the position information. In addition, the embedding vector will also add a class token for the final classification, which can capture the global semantic information of the entire image, so that the model can understand the context of the entire image; next, the position-encoded embedding vector sequence is input into a multi-layer Transformer encoder. Each encoder consists of a multi-head self-attention mechanism, a feedforward neural network, layer normalization, and residual connection. The multi-head self-attention mechanism allows the model to assign different weights to image patches at different positions to capture the global features of the image;

[0043] Finally, the output sequence of the last Transformer encoder is passed to a classifier consisting of fully connected layers, which can map the comprehensive information of the image to the prediction of the spatial probability of landslide occurrence based on the class tokens in the output sequence.

[0044] In the present invention, the model hyperparameters are continuously debugged and optimized, the accuracy and loss value of the validation set are continuously monitored during the training process to select the best performing ViT model, and the best ViT model is saved through the ModelCheckpoint function to obtain the model with the best parameters.

[0045] S4: Replace the LULC factor and rainfall factor in the landslide influencing factor with the LULC factor simulated by the PLUS model and the rainfall factor simulated by the atmospheric circulation model, respectively;

[0046] It should be noted that the present invention constructs a PLUS model based on historical LULC data and land use change driving factors to simulate future LULC factors.

[0047] Using three atmospheric circulation models in the CMIP6 program, rainfall simulation data under multiple scenarios in the future were obtained.

[0048] As an example, the process of obtaining the simulated LULC factor in step S4 is as follows:

[0049] S411. Obtain historical LULC factor data and land use change driving factors at different phases;

[0050] Specifically, the present invention extracts historical LULC data of the study area, selects elevation, slope, aspect, soil type, average annual precipitation, average annual temperature, distance to rivers, distance to roads, population density and GDP as driving factors of land use change, and performs data preprocessing operations such as data projection and raster resampling on each driving factor.

[0051] S412, based on the historical LULC factor data of two phases, the expansion areas of various land use types were extracted through the PLUS model toolkit;

[0052] Specifically, based on the LULC data of the two temporal phases, the expansion areas of various land use types were extracted using the Exact Land Expansion tool of the PLUS model;

[0053] S413, importing the extracted grids of expansion areas of various land use types into the toolkit of the PLUS model, and importing the driving factors of land use change in step S411, and outputting the development probability of each land use type and the contribution of each driving factor to the expansion of each land use type;

[0054] Specifically, the extracted expansion area grids of each land type are imported into the LEAS module of the PLUS model, and all the land use change driving factor grid data are imported at the same time, and the development probability of each land type and the contribution of each driving factor to the expansion of each land type are output;

[0055] S414. Set different parameters of the toolkit in the PLUS model, run the set PLUS model, and obtain LULC simulation data.

[0056] Specifically, based on the output of the LEAS module, set the CARS module of the PLUS model

[0057] The neighborhood weights of land use types, land use transfer rule matrix and future demand of land use types are used to run the PLUS model to obtain LULC simulation data;

[0058] According to the overall accuracy of the LULC simulation data compared with the real LULC data, the above three parameters are continuously adjusted to finally obtain the spatial simulation results of LULC distribution under the future scenario with the best accuracy.

[0059] Specifically, in the present invention, the detailed functions and setting rules of the three parameters are as follows.

[0060] (1) Neighborhood weight of land use type

[0061] The neighborhood weight is used to quantitatively characterize the differences in the expansion capacity of each land type under the influence of driving factors. Its range is between 0 and 1. The higher the value, the stronger the expansion capacity of the land type, and the less likely it is to be converted into other land types. The neighborhood weight of each land type is determined by calculating the proportion of the expansion area of ​​each land type to the total expansion area. The calculation formula is as follows:

[0062]

[0063] Where W i is the neighborhood weight of land type i, LU i is the expansion area of ​​land type i, LU min is the minimum expansion area among all land types, LU max It is the largest expansion area among all land types.

[0064] (2) Land use transfer rule matrix

[0065] The land use transfer rule matrix represents the conversion restrictions between various land use types. The matrix value is 0 when conversion is not allowed and 1 when conversion is allowed.

[0066] (3) Future demand for land use types

[0067] Before conducting land use simulation, it is necessary to predict the future land use type demand. The Markov chain model is a prediction method that predicts the future land use change based on the current land use situation. The model can accurately fit the nonlinear relationship between data and has strong applicability to various types of data. Therefore, the Markov chain model is often used to simulate the land use type demand in future scenarios.

[0068] As an example, the process of obtaining the simulated rainfall factor in step S4 is as follows:

[0069] S421, select different atmospheric circulation models and the required shared socio-economic pathways;

[0070] First, the required atmospheric circulation model is selected, which is a core component of the CMIP6 program and simulates climate change on time scales from minutes to decades.

[0071] According to previous research on atmospheric circulation models in China, three atmospheric circulation models, EC-Earth3, GFDL-ESM4 and MRI-ESM2-0, were selected for experiments.

[0072] Then, the required shared socioeconomic pathways (SSPs) are selected. It is a key component of the atmospheric circulation model, which aims to provide different socioeconomic development scenarios for climate change models in order to assess the impact of these scenarios on future climate. Three carbon emission development scenarios, SSP1-1.9, SSP2-4.5, and SSP5-8.5, were selected. Each development scenario represents a different future social development trend, that is, three trends in the carbon emission concentration of human society in the future.

[0073] S422. Based on the required shared socioeconomic pathway, rainfall factors obtained from different atmospheric circulation models are averaged to obtain simulated rainfall factors.

[0074] In the present invention, taking into account the inherent uncertainty of the atmospheric circulation model, based on the three adopted future carbon emission scenarios, the average rainfall of the three atmospheric circulation models in the same scenario is selected as the simulated rainfall factor.

[0075] Finally, the simulated rainfall factors were subjected to preprocessing operations such as clipping, projection, and resampling to achieve spatial unification of rainfall simulation factors and other landslide influencing factor data.

[0076] S5. Input the simulated LULC factor, simulated rainfall factor, and other factors in the landslide influencing factor into the model with the best parameters to obtain the prediction results of landslide susceptibility in future scenarios.

[0077] As an embodiment, the present invention uses simulated LULC and rainfall factors to replace the original two landslide impact factors, and constructs a new landslide simulation data set under future scenarios with other seven landslide impact factors.

[0078] Then, the landslide simulation dataset is input into the previously saved ViT model with the best training accuracy, and the powerful landslide feature extraction capability of the ViT model is used to obtain more reliable landslide susceptibility prediction results under future scenarios. Figure 4 , Figure 4 It is a detailed process schematic diagram of the above method of the present invention.

[0079] See also Figure 5 , Figure 5 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a future scenario landslide susceptibility prediction device 401, a processor 402 and a storage medium 403.

[0080] A future scenario landslide susceptibility prediction device 401: The electronic digital marketing device 401 implements the future scenario landslide susceptibility prediction method.

[0081] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the method for predicting landslide susceptibility in future scenarios.

[0082] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the method for predicting landslide susceptibility in future scenarios.

[0083] The beneficial effects of the present invention are as follows: a landslide susceptibility prediction method under future scenarios is proposed, which simultaneously considers the impact of changes in LULC and rainfall factors on future landslide susceptibility, making up for the deficiency of often only predicting landslide susceptibility by simulating a single landslide influencing factor.

[0084] At the same time, this method uses the ViT model with stronger feature extraction capability to predict landslide susceptibility. The prediction results obtained by this method are more credible than those of physical models, statistical models and machine learning models.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting landslide susceptibility in future scenarios, characterized by: Includes the following: S1: Select landslide influencing factors according to the landslide disaster-prone environment and disaster-causing factors in the study area; S2: After preprocessing the landslide influencing factors, the data set is constructed; S3: constructing a landslide prediction model, using the data set to train and verify the landslide prediction model, and obtaining a model with optimal parameters; S4: Replace the LULC factor and rainfall factor in the landslide influencing factor with the LULC factor simulated by the PLUS model and the rainfall factor simulated by the atmospheric circulation model, respectively; S5. Input the simulated LULC factor, simulated rainfall factor, and other factors in the landslide influencing factor into the model with the best parameters to obtain the prediction results of landslide susceptibility in future scenarios.

2. A method for predicting landslide susceptibility in future scenarios as claimed in claim 1, characterized in that: The landslide influencing factors include: elevation, slope, slope aspect, lithology, distance from fault, distance from river, normalized difference vegetation index, rainfall factor and LULC factor.

3. A method for predicting landslide susceptibility in future scenarios as claimed in claim 1, characterized in that: The preprocessing method includes the following operations: coordinate system one, outlier removal, grid resampling and numerical normalization performed in sequence.

4. A method for predicting landslide susceptibility in future scenarios as claimed in claim 1, characterized in that: The data set is divided into a training set, a validation set and a test set according to a preset ratio.

5. A method for predicting landslide susceptibility in future scenarios as claimed in claim 1, characterized in that: The landslide prediction model adopts the ViT model.

6. A method for predicting landslide susceptibility in future scenarios as claimed in claim 1, characterized in that: The process of obtaining the simulated LULC factor in step S4 is as follows: S411. Obtain historical LULC factor data and land use change driving factors at different phases; S412, based on the historical LULC factor data of two phases, the expansion areas of various land use types were extracted through the PLUS model toolkit; S413, importing the extracted grids of expansion areas of various land use types into the toolkit of the PLUS model, and importing the driving factors of land use change in step S411, and outputting the development probability of each land use type and the contribution of each driving factor to the expansion of each land use type; S414. Set different parameters of the toolkit in the PLUS model, run the set PLUS model, and obtain LULC simulation data.

7. A method for predicting landslide susceptibility in future scenarios as claimed in claim 6, characterized in that: The different parameters in step S414 include: land use type neighborhood weight parameter, land use transfer rule matrix and land use type future demand parameter; The neighborhood weight of land use type is as follows: Where W i is the neighborhood weight of land type i, LU i is the expansion area of ​​land type i, LU min is the minimum expansion area among all land types, LU max It is the largest expansion area among all land types; The land use transfer rule matrix represents the conversion restrictions between various land use types. The matrix value is 0 when the conversion is not allowed, and the matrix value is 1 when the conversion is allowed. The future demand parameters of land use types are obtained using the Markov chain model.

8. A method for predicting landslide susceptibility in future scenarios as claimed in claim 1, characterized in that: The process of obtaining the simulated rainfall factor in step S4 is as follows: S421, select different atmospheric circulation models and the required shared socio-economic pathways; S422. Based on the required shared socioeconomic pathway, rainfall factors obtained from different atmospheric circulation models are averaged to obtain simulated rainfall factors.

9. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a future scenario landslide susceptibility prediction method as described in any one of claims 1 to 8.

10. A device for predicting the susceptibility of landslides in future scenarios, characterized by: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a future scenario landslide susceptibility prediction method as described in any one of claims 1 to 8.

Images