Landslide susceptibility evaluation method based on improved deep learning

Abstract

The invention discloses a landslide susceptibility evaluation method based on improved deep learning, and relates to the technical field of geological disaster prediction. According to the method, the prediction accuracy is remarkably improved, deep correlation features of multi-source data are fully mined through fusion of a multi-machine learning model (RandomForest, LightGBM and AdaBoost) and an improved PeepholeLSTM model (an attention mechanism is introduced), the complex nonlinear relation between a landslide influence factor and landslide occurrence is effectively captured, and through test verification, the model AUC value can reach 0.92 or above, the accuracy rate exceeds 84%, and the recall rate exceeds 94%; the method is high in generalization ability, the overfitting risk of the model is reduced through data standardization, random disruption of a data set and dropout regularization, the model can adapt to geological environments and landslide characteristics of different areas (such as plateau mountainous areas and hilly areas), and the problem that an existing method is poor in area adaptability is solved.

Description

Technical Field

[0001] This invention relates to the field of geological disaster prediction technology, specifically to an improved deep learning-based method for landslide susceptibility assessment. Background Technology

[0002] Landslides, as typical sudden geological disasters, pose a serious threat to human life and property, the ecological environment, and infrastructure (such as roads and water conservancy projects). With the acceleration of urbanization and the intensification of human engineering activities (such as mining and road construction) on the natural environment, the frequency, scope, and degree of damage of landslide disasters are showing a significant upward trend. Therefore, achieving efficient and accurate assessment of landslide susceptibility has become a core requirement in the field of geological disaster prevention and control.

[0003] Existing landslide susceptibility assessment methods are mainly divided into two categories:

[0004] Traditional qualitative / semi-quantitative methods rely on the experience of geologists, geomorphologists, and hydrologists to make qualitative judgments by analyzing factors such as topography and geological structure (e.g., analytic hierarchy process, expert scoring). These methods suffer from high subjectivity, difficulty in quantifying evaluation results, low efficiency, and are unable to adapt to the complex and varied geological environments of different regions, exhibiting poor generalization ability.

[0005] Machine learning methods: With the development of remote sensing (RS), geographic information systems (GIS), and big data technologies, machine learning algorithms (such as random forests, AdaBoost, and LightGBM) are widely used in landslide susceptibility assessment. These methods achieve quantitative assessment through data-driven approaches, reducing subjectivity to some extent, but still have significant shortcomings:

[0006] Many models are trained using a single model and lack a multi-model collaborative fusion mechanism, making it difficult to fully explore the deep correlation features in multi-source heterogeneous data (such as terrain data, meteorological data, and geological data).

[0007] Without leveraging the advantages of deep learning technology in capturing temporal and correlational features, it is impossible to effectively establish the complex nonlinear relationship between landslide influencing factors and landslide occurrence;

[0008] Insufficient preprocessing and feature selection for multi-source data can easily lead to model overfitting, and the model has poor adaptability and limited generalization ability in different geological environments.

[0009] Therefore, there is an urgent need for a landslide susceptibility assessment method that can integrate multi-source data, combine the advantages of multiple models, and introduce improved deep learning technology to solve the problems of low accuracy, poor generalization ability, and weak adaptability in existing technologies, and provide scientific and reliable technical support for landslide disaster prevention and control. Summary of the Invention

[0010] The purpose of this invention is to provide an improved deep learning-based method for evaluating landslide susceptibility, in order to solve the technical problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: an improved deep learning-based method for assessing landslide susceptibility, comprising at least the following steps:

[0012] S1: Data preprocessing, collecting multi-source geographic information data of the landslide study area, unifying the spatial resolution and coordinate system of the multi-source geographic information data, and then performing data cleaning and standardization; the multi-source geographic information data includes topographic factors, geological factors, meteorological factors, land cover factors and human factors;

[0013] S2: Feature extraction and selection. The original values ​​of each evaluation factor in the multi-source geographic information data preprocessed in step S1 are used to construct the initial feature set. Key features are selected by combining historical landslide data and expert experience to construct the final feature dataset.

[0014] S3: Dataset Construction and Random Shuffling;

[0015] S4: Model training. The training set is imported into RandomForest, LightGBM and AdaBoost models for initial training. The output results of the three models are fused to obtain a fused feature vector. An attention mechanism is introduced into the PeepholeLSTM model. The fused feature vector is used as input. The cross-entropy loss function and Adam optimizer are used to train the improved PeepholeLSTM model. The model with the best performance is selected through the test set.

[0016] S5: Risk assessment. The landslide susceptibility probability value output by the optimal model is divided into five levels using the natural breakpoint method, and a landslide susceptibility prediction map is generated.

[0017] Furthermore, S1 includes at least the following steps:

[0018] S1.1: Data cleaning: missing values ​​are filled using interpolation, and outliers are identified and removed using the 3σ principle;

[0019] S1.2: Data standardization, using the formula:

[0020]

[0021] Map the data to the interval [0,1], where, The original data values, and These are the minimum and maximum values ​​of the original data, respectively. These are the normalized data values.

[0022] Furthermore, in S2, the key features are selected by analyzing the correlation between historical landslide data and various evaluation factors through Pearson correlation coefficient analysis, and evaluation factors with a correlation coefficient > 0.6 are selected as key features.

[0023] Furthermore, S3 includes at least the following steps:

[0024] S3.1: Data partitioning, using a random number generator to shuffle the dataset to ensure consistent distribution between the training and test sets;

[0025] S3.2: Dataset ratio, the training set and the test set are divided in a 7:3 ratio.

[0026] Furthermore, S4 includes at least the following steps:

[0027] S4.1: Training Multiple Machine Learning Models

[0028] The RandomForest model generates training samples through bootstrap sampling, determines feature split points with the goal of minimizing the Gini coefficient, and finally outputs the prediction results through majority voting.

[0029] The LightGBM model employs histogram feature discretization, one-sided gradient sampling, and mutually exclusive feature bundling techniques, and the prediction result is a weighted sum of multiple trees;

[0030] The AdaBoost model iteratively trains weak classifiers, improves accuracy by optimizing the exponential loss function, and the final prediction result is a weighted sum of the weak classifiers.

[0031] S4.2: Feature fusion, which concatenates the predicted probability values ​​from the outputs of the three machine learning models into a fused feature vector;

[0032] S4.3: Improved PeepholeLSTM model training;

[0033] The formula for the attention mechanism is:

[0034]

[0035] Where Q, K, and V are matrices generated by fusing feature vectors. The dimension is K;

[0036] Define a PeepholeLSTM model, which includes an input layer, a PeepholeLSTM layer, an attention mechanism layer, and an output layer;

[0037] Input gate: controls the degree to which new information is input;

[0038] Forget Gate: Determines whether to retain or discard information in a cell's state;

[0039] Output gate: The degree to which the control unit state outputs information about the hidden state;

[0040] Cell state update: Update the cell state based on the outputs of the input gate and the forget gate;

[0041] Hidden state update: Update the hidden state based on the output of the output gate and the updated cell state;

[0042] The calculation formulas are as follows:

[0043]

[0044] in, This is an input gate that controls the current input. The degree to which it is stored in the cellular state; This is the Sigmoid activation function, with an output between 0 and 1, representing the "permissibility" of the input value. The weight matrix corresponds to the inputs respectively. The hidden state of the previous moment The cell state at the previous moment ;

[0045]

[0046] in, This is the forget gate, which controls the current cell state. The degree to which information from the previous moment is retained; The weight matrix corresponds to the inputs respectively. The hidden state of the previous moment The cell state at the previous moment For bias terms;

[0047]

[0048] in, As an output gate, it controls the cell state. Hide the current state The degree of contribution; The weight matrix corresponds to the inputs respectively. The hidden state of the previous moment and the current cell state For bias terms;

[0049]

[0050] in, This represents the current state of the cell. The forgetting gate acts on the cell state of the previous moment, controlling the retention of memory information from the previous moment; The input gate acts on the input at the current moment, determining how the information at the current moment is added to the cell state; To represent element-wise multiplication;

[0051]

[0052] in, The current hidden state is used as the output of the network. To perform hyperbolic tangential activation on cell states, a range in... The values ​​between; The activation value of the output gate acting on the cell state determines the output of the current hidden state;

[0053] Model parameters: hidden layer size 64, number of layers 2, learning rate 0.002, training epochs 200, batch size 32, dropout regularization strength 0.5;

[0054] The expression for the cross-entropy loss function is:

[0055]

[0056] in, This represents the value of the loss function. This represents the true label of the i-th sample. This represents the probability value predicted by the model, where n represents the number of samples. Let T represent the structural parameters of the t-th tree, and T represent the number of decision trees. This represents the weight of the leaf node in the t-th tree;

[0057] S4.4: Model output. By obtaining the model with the highest performance index, the best model is saved for subsequent probability prediction.

[0058] Furthermore, S5 includes at least the following steps:

[0059] S5.1: Five-level classification standard: extremely low susceptibility (0-0.2), low susceptibility (0.2-0.4), medium susceptibility (0.4-0.6), high susceptibility (0.6-1.75), extremely high susceptibility (0.8-1.0);

[0060] S5.2: The formula for calculating the probability of landslide susceptibility is as follows: ,in This represents the probability of a landslide occurring. For feature vectors, This is the weight matrix. For bias terms, This is the Sigmoid activation function.

[0061] Furthermore, the coordinate system of the multi-source geographic information data in S1 is WGS_1984_UTM_Zone_49N;

[0062] The spatial resolution is uniformly set to 30m raster data.

[0063] Compared with the prior art, the beneficial effects of the present invention are:

[0064] 1. The prediction accuracy of this invention is significantly improved. By fusing and improving the PeepholeLSTM model (which introduces an attention mechanism) with multiple machine learning models (RandomForest, LightGBM, AdaBoost), it fully explores the deep correlation features of multi-source data and effectively captures the complex nonlinear relationship between landslide influencing factors and landslide occurrence. Tests have verified that the model's AUC value can reach over 0.92, with an accuracy of over 84% and a recall of over 94%.

[0065] 2. The present invention has strong generalization ability. By standardizing data, randomly shuffling the dataset and dropout regularization, it reduces the risk of model overfitting and enables the model to adapt to the geological environment and landslide characteristics of different regions (such as plateau mountains and hilly areas), thus solving the problem of "poor regional adaptability" of existing methods.

[0066] 3. The present invention has high data processing efficiency. The histogram discretization, GOSS and EFB technology of the LightGBM model, combined with the temporal feature processing advantages of PeepholeLSTM, greatly reduces the processing complexity of multi-source heterogeneous data and improves the efficiency of model training and prediction.

[0067] 4. The present invention is highly practical. The generated landslide susceptibility level 5 prediction map can be directly applied to geological disaster prevention and control practices, providing a scientific basis for hazard point identification, prevention and control resource allocation and emergency decision-making, and has significant engineering application value. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is the overall flowchart of the present invention;

[0070] Figure 2 This is a schematic diagram of the ROC curve of the present invention;

[0071] Figure 3 This is a schematic diagram of the four performance indicators of the model of the present invention. Detailed Implementation

[0072] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0073] Example 1:

[0074] Please see Figures 1-3 An improved deep learning-based method for assessing landslide susceptibility includes at least the following steps:

[0075] S1: Data preprocessing involves collecting multi-source geographic information data of the landslide study area, unifying the spatial resolution and coordinate system of the multi-source geographic information data, and then performing data cleaning and standardization. The multi-source geographic information data includes topographic factors, geological factors, meteorological factors, land cover factors, and human factors.

[0076] S2: Feature extraction and selection. The original values ​​of each evaluation factor in the multi-source geographic information data preprocessed in step S1 are used to construct the initial feature set. Key features are selected by combining historical landslide data and expert experience to construct the final feature dataset.

[0077] S3: Dataset Construction and Random Shuffling;

[0078] S4: Model training. The training set is imported into RandomForest, LightGBM and AdaBoost models for initial training. The output results of the three models are fused to obtain a fused feature vector. An attention mechanism is introduced into the PeepholeLSTM model. The fused feature vector is used as input. The cross-entropy loss function and Adam optimizer are used to train the improved PeepholeLSTM model. The model with the best performance is selected through the test set.

[0079] S5: Risk assessment. The natural breakpoint method is used to divide the landslide susceptibility probability values ​​output by the optimal model into five levels, and a landslide susceptibility prediction map is generated.

[0080] S1 includes at least the following steps:

[0081] S1.1: Data cleaning: missing values ​​are filled using interpolation, and outliers are identified and removed using the 3σ principle;

[0082] S1.2: Data standardization, using the formula:

[0083]

[0084] Map the data to the interval [0,1], where, The original data values, and These are the minimum and maximum values ​​of the original data, respectively. These are the normalized data values.

[0085] In S2, the key features were selected by analyzing the correlation between historical landslide data and various evaluation factors using the Pearson correlation coefficient, and evaluation factors with a correlation coefficient > 0.6 were selected as key features.

[0086] S3 includes at least the following steps:

[0087] S3.1: Data partitioning, using a random number generator to shuffle the dataset to ensure consistent distribution between the training and test sets;

[0088] S3.2: Dataset ratio, the training set and the test set are divided in a 7:3 ratio.

[0089] S4 includes at least the following steps:

[0090] S4.1: Training Multiple Machine Learning Models

[0091] The RandomForest model generates training samples through bootstrap sampling, determines feature split points with the goal of minimizing the Gini coefficient, and finally outputs the prediction results through majority voting.

[0092] The LightGBM model employs histogram feature discretization, one-sided gradient sampling, and mutually exclusive feature bundling techniques, and the prediction result is a weighted sum of multiple trees;

[0093] The AdaBoost model iteratively trains weak classifiers, improves accuracy by optimizing the exponential loss function, and the final prediction result is a weighted sum of the weak classifiers.

[0094] S4.2: Feature fusion, which concatenates the predicted probability values ​​from the outputs of the three machine learning models into a fused feature vector;

[0095] S4.3: Improved PeepholeLSTM model training;

[0096] The formula for the attention mechanism is:

[0097]

[0098] Where Q, K, and V are matrices generated by fusing feature vectors. The dimension is K;

[0099] Define the PeepholeLSTM model, which includes an input layer, a PeepholeLSTM layer, an attention mechanism layer, and an output layer;

[0100] Input gate: controls the degree to which new information is input;

[0101] Forget Gate: Determines whether to retain or discard information in a cell's state;

[0102] Output gate: The degree to which the control unit state outputs information about the hidden state;

[0103] Cell state update: Update the cell state based on the outputs of the input gate and the forget gate;

[0104] Hidden state update: Update the hidden state based on the output of the output gate and the updated cell state;

[0105] The calculation formulas are as follows:

[0106]

[0107] in, This is an input gate that controls the current input. The degree to which it is stored in the cellular state; This is the Sigmoid activation function, with an output between 0 and 1, representing the "permissibility" of the input value. The weight matrix corresponds to the inputs respectively. The hidden state of the previous moment The cell state at the previous moment ;

[0108]

[0109] in, This is the forget gate, which controls the current cell state. The degree to which information from the previous moment is retained; The weight matrix corresponds to the inputs respectively. The hidden state of the previous moment The cell state at the previous moment For bias terms;

[0110]

[0111] in, As an output gate, it controls the cell state. Hide the current state The degree of contribution; The weight matrix corresponds to the inputs respectively. The hidden state of the previous moment and the current cell state For bias terms;

[0112]

[0113] in, This represents the current state of the cell. The forgetting gate acts on the cell state of the previous moment, controlling the retention of memory information from the previous moment; The input gate acts on the input at the current moment, determining how the information at the current moment is added to the cell state; To represent element-wise multiplication;

[0114]

[0115] in, The current hidden state is used as the output of the network. To perform hyperbolic tangential activation on cell states, a range in... The values ​​between; The activation value of the output gate acting on the cell state determines the output of the current hidden state;

[0116] Model parameters: hidden layer size 64, number of layers 2, learning rate 0.002, training epochs 200, batch size 32, dropout regularization strength 0.5;

[0117] The expression for the cross-entropy loss function is:

[0118]

[0119] in, This represents the value of the loss function. This represents the true label of the i-th sample. This represents the probability value predicted by the model, where n represents the number of samples. Let T represent the structural parameters of the t-th tree, and T represent the number of decision trees. This represents the weight of the leaf node in the t-th tree;

[0120] S4.4: Model output. By obtaining the model with the highest performance index, the best model is saved for subsequent probability prediction.

[0121] S5 includes at least the following steps:

[0122] S5.1: Five-level classification standard: extremely low susceptibility (0-0.2), low susceptibility (0.2-0.4), medium susceptibility (0.4-0.6), high susceptibility (0.6-1.75), extremely high susceptibility (0.8-1.0);

[0123] S5.2: The formula for calculating the probability of landslide susceptibility is as follows: ,in This represents the probability of a landslide occurring. For feature vectors, This is the weight matrix. For bias terms, This is the Sigmoid activation function.

[0124] The coordinate system of the multi-source geographic information data in S1 is WGS_1984_UTM_Zone_49N;

[0125] The spatial resolution is uniformly set to 30m raster data.

[0126] Example 2:

[0127] This embodiment provides a specific application description based on the above embodiments, taking Zogang County as the research object. The specific steps are as follows:

[0128] Step S1: Data Preprocessing

[0129] S101 collects multi-source geospatial data of the study area, including elevation, slope, aspect, curvature, vegetation index, stratigraphic lithology, rainfall data, land use type, road distance, and waterway distance. This data originates from Geographic Information System (GIS) databases, satellite remote sensing imagery, field surveys, and publicly available datasets from relevant government departments.

[0130] S102 performs unified processing on the collected data in terms of spatial resolution and coordinate system, ensuring that all data are within the same spatial reference frame. The coordinate system of the collected data is transformed into WGS_1984_UTM_Zone_49N, and all data is resampled into 30m raster data.

[0131] S103 cleans the data, removing missing and outlier values, and performs data standardization to improve data quality. The acquired data is then normalized, mapping the values ​​to the [0,1] interval, using the following formula:

[0132] x′=

[0133] Where x is the original data value, xmin and xmax are the minimum and maximum values ​​of the data, respectively, and x′ is the normalized data value.

[0134] Step S2 Feature Extraction and Selection

[0135] S201 directly uses the original values ​​of each evaluation factor in the multi-source geographic information data to construct a dataset.

[0136] S202 selects key features based on experience, choosing factors that are highly correlated with landslide occurrence.

[0137] Step S3: Dataset Construction and Random Shuffling

[0138] S301 constructs the training and test sets in a 7:3 ratio and randomly shuffles the datasets to ensure distribution consistency.

[0139] Step S4: Model Training

[0140] S4.1 Initial Model: The dataset is imported into three machine learning models (including RandomForest, LightGBM, and AdaBoost) for training.

[0141] S4.1.1 Random Forest Model Training: The Random Forest (RF) model is trained using the training set data. The RF model consists of multiple decision trees, which improves predictive performance by reducing the risk of overfitting and increasing the model's stability and accuracy. Each decision tree is constructed based on the following formula:

[0142] Let the training set be D={(x1,y1),(x2,y2),…,(x n ,y n )}, where xi is the feature vector and yi is the corresponding label. For each decision tree, n samples are drawn with replacement from the training set to form the bootstrap sample set. At each node, m features are randomly selected from all features, and the feature that minimizes the Gini coefficient and its split point are selected for splitting. The Gini coefficient is defined as:

[0143] Gini(D)=1−

[0144] Where, p k This represents the proportion of samples in sample set D belonging to class k. For binary classification problems, a smaller Gini coefficient indicates higher dataset purity. Each decision tree is built recursively from the bootstrap sample set by splitting nodes until a preset maximum depth is reached or the number of samples in a node is less than the minimum number of split samples. Finally, the random forest model classifies the test samples using majority voting, i.e.:

[0145] y=mode(tree1(x),tree2(x),…,tree n (x))

[0146] Among them, tree i (x) represents the prediction result of the i-th decision tree for sample x, and y is the final prediction result.

[0147] S4.1.2 LightBoost Model Training: This algorithm significantly reduces the computational complexity of feature splitting through key techniques such as histogram feature discretization, one-sided gradient sampling (GOSS), and mutually exclusive feature binding (EFB).

[0148] LightGBM's prediction formula is a weighted sum of the prediction results from multiple trees:

[0149] )

[0150] in ) is the prediction result for the k-th tree, where K is the total number of trees.

[0151] S4.1.3 AdaBoost Model Training: Train the AdaBoost model using the training set data. The AdaBoost model improves prediction accuracy by optimizing the loss function, which is defined as:

[0152] L=

[0153] Among them, y i It is the true label of the sample, f(x) i () represents the label predicted by the model. The core of the AdaBoost algorithm lies in iteratively training multiple weak classifiers and combining them into a strong classifier. The prediction formula of AdaBoost is:

[0154] )

[0155] in It is the t-th weak classifier. It is its weight.

[0156] S4.2 Feature Fusion: This involves fusing the features of the output data from the three machine learning models.

[0157] S4.3 Model Improvement: An attention mechanism is introduced into the PeepholeLSTM model. Cross-entropy loss and Adam optimizer are used for model training. The feature fusion data obtained in step 4.2 is used as input data, and the parameters are adjusted and trained multiple times to obtain the improved optimal PeepholeLSTM model.

[0158] S4.3.1 Attention Mechanism: Used to enhance the representation of key features, the formula is as follows:

[0159] V

[0160] Where Q, K, and V are the query, key, and value matrices, respectively, and d k The dimension of the key.

[0161] Fully connected layer: used for final classification prediction, with ReLU activation function.

[0162] S4.3.2 defines the PeepholeLSTM model structure, including an input layer, a PeepholeLSTM layer, an attention mechanism layer, and an output layer. The computational formula for the PeepholeLSTM unit is as follows:

[0163] Input gate: controls the degree to which new information is input.

[0164] Forget Gate: Decides whether to retain or discard information in the cell state.

[0165] Output gate: The degree to which the control unit state outputs information about the hidden state.

[0166] Cell state update: Update the cell state based on the outputs of the input gate and the forget gate.

[0167] Hidden state update: Update the hidden state based on the output of the output gate and the updated cell state.

[0168] The specific mathematical expression is:

[0169] i t =σ(W ix x t +W ih h t−1 +W ci C t−1 +b i Input gate

[0170] f t =σ(W fx x t +W fh h t−1 +W cf C t−1 +b f Forgotten Gate

[0171] o t =σ(W ox x t +W oh h t−1 +W co C t +b o Output gate

[0172] C t =f t ⊙C t−1 +i t ⊙(W cx x t +W ch h t−1 +bc Cellular state

[0173] h t =o t ⊙(W hy h t +b h Hidden state

[0174] Among them, i t f t o t These represent the input gate, forget gate, and output gate, respectively; c t and h t Representing the cell state and hidden state respectively; x t The input data is defined as follows: W and b represent the weight matrix and bias term, respectively; σ is the activation function. In this invention, the parameters of the PeepholeLSTM unit are set as follows: input size is the dimension of the fused feature vector, hidden layer size is 64, number of layers is 2, output size is 2 (representing the probability of geological disasters occurring or not occurring), learning rate is 0.002, training period is 200 epochs, batch size is 32, and dropout regularization strength to prevent overfitting is 0.5.

[0175] The expression for the cross-entropy loss function in S4.3.3 is as follows:

[0176]

[0177] in, This represents the value of the loss function. This represents the true label of the i-th sample. This represents the probability value predicted by the model, where n represents the number of samples. Let T represent the structural parameters of the t-th tree, and T represent the number of decision trees. This represents the weight of the leaf node of the t-th tree.

[0178] The S4.4 model output is used to obtain the model with the highest performance index. The best model is saved for subsequent probability prediction.

[0179] S501 classifies the landslide susceptibility into five levels according to the natural breakpoint method, applies the landslide susceptibility prediction model to the landslide study area, and generates a landslide susceptibility prediction map.

[0180] S502 uses the following formula to calculate the probability of landslide susceptibility: P(slide) = σ(W) X +b) where P(slide) represents the probability of a landslide, X is the eigenvector, W is the weight matrix, b is the bias term, and σ is the activation function.

[0181] In summary:

[0182] The purpose of this invention is to overcome the shortcomings of existing landslide susceptibility assessment methods, such as strong subjectivity, low accuracy, poor generalization ability, and difficulty in handling multi-source heterogeneous data, and to provide an improved deep learning-based landslide susceptibility assessment method and system, which is achieved through the following technical objectives:

[0183] Integrate multi-source geographic information data to achieve standardized and unified data processing, thereby improving data quality;

[0184] By integrating multiple machine learning models (RandomForest, LightGBM, AdaBoost) with an improved PeepholeLSTM model, we can fully explore the deep features of the data and establish the complex relationship between landslide influencing factors and landslide occurrence.

[0185] An attention mechanism is introduced to optimize the PeepholeLSTM model, enhance the weight of key features, and improve prediction accuracy;

[0186] A five-level classification assessment of landslide susceptibility is implemented, and a visual prediction map is generated to provide decision support for landslide risk management.

[0187] This invention is applicable to landslide risk assessment, prevention and control and decision support in complex terrain areas. It can achieve accurate quantitative evaluation of landslide susceptibility by integrating multi-source geographic information data and improved deep learning models.

[0188] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An improved deep learning-based method for assessing landslide susceptibility, characterized in that: At least the following steps are included: S1: Data preprocessing, collecting multi-source geographic information data of the landslide study area, unifying the spatial resolution and coordinate system of the multi-source geographic information data, and then performing data cleaning and standardization; the multi-source geographic information data includes topographic factors, geological factors, meteorological factors, land cover factors and human factors; S2: Feature extraction and selection. The original values ​​of each evaluation factor in the multi-source geographic information data preprocessed in step S1 are used to construct the initial feature set. Key features are selected by combining historical landslide data and expert experience to construct the final feature dataset. S3: Dataset Construction and Random Shuffling; S4: Model training. The training set is imported into RandomForest, LightGBM and AdaBoost models for initial training. The output results of the three models are fused to obtain a fused feature vector. An attention mechanism is introduced into the PeepholeLSTM model. The fused feature vector is used as input. The cross-entropy loss function and Adam optimizer are used to train the improved PeepholeLSTM model. The model with the best performance is selected through the test set. S5: Risk assessment. The landslide susceptibility probability value output by the optimal model is divided into five levels using the natural breakpoint method, and a landslide susceptibility prediction map is generated.

2. The landslide susceptibility assessment method based on improved deep learning according to claim 1, characterized in that: S1 includes at least the following steps: S1.1: Data cleaning: missing values ​​are filled using interpolation, and outliers are identified and removed using the 3σ principle; S1.2: Data standardization, using the formula: Map the data to the interval [0,1], where, The original data values, and These are the minimum and maximum values ​​of the original data, respectively. These are the normalized data values.

3. The landslide susceptibility assessment method based on improved deep learning according to claim 1, characterized in that: The key features in S2 are selected by analyzing the correlation between historical landslide data and various evaluation factors using Pearson correlation coefficient, and evaluation factors with a correlation coefficient > 0.6 are selected as key features.

4. The landslide susceptibility assessment method based on improved deep learning according to claim 1, characterized in that: The S3 includes at least the following steps: S3.1: Data partitioning, using a random number generator to shuffle the dataset to ensure consistent distribution between the training and test sets; S3.2: Dataset ratio, the training set and the test set are divided in a 7:3 ratio.

5. The landslide susceptibility assessment method based on improved deep learning according to claim 1, characterized in that: The S4 includes at least the following steps: S4.1: Training Multiple Machine Learning Models The RandomForest model generates training samples through bootstrap sampling, determines feature split points with the goal of minimizing the Gini coefficient, and finally outputs the prediction results through majority voting. The LightGBM model employs histogram feature discretization, one-sided gradient sampling, and mutually exclusive feature bundling techniques, and the prediction result is a weighted sum of multiple trees; The AdaBoost model iteratively trains weak classifiers, improves accuracy by optimizing the exponential loss function, and the final prediction result is a weighted sum of the weak classifiers. S4.2: Feature fusion, which concatenates the predicted probability values ​​from the outputs of the three machine learning models into a fused feature vector; S4.3: Improved PeepholeLSTM model training; The formula for the attention mechanism is: Where Q, K, and V are matrices generated by fusing feature vectors. The dimension is K; Define a PeepholeLSTM model, which includes an input layer, a PeepholeLSTM layer, an attention mechanism layer, and an output layer; Input gate: controls the degree to which new information is input; Forget Gate: Determines whether to retain or discard information in a cell's state; Output gate: The degree to which the control unit state outputs information about the hidden state; Cell state update: Update the cell state based on the outputs of the input gate and the forget gate; Hidden state update: Update the hidden state based on the output of the output gate and the updated cell state; S4.4: Model output. By obtaining the model with the highest performance index, the best model is saved for subsequent probability prediction.

6. The landslide susceptibility assessment method based on improved deep learning according to claim 1, characterized in that: The S5 includes at least the following steps: S5.1: Five-level classification standard: extremely low susceptibility, low susceptibility, medium susceptibility, high susceptibility, extremely high susceptibility; S5.2: The formula for calculating the probability of landslide susceptibility is as follows: ,in This represents the probability of a landslide occurring. For feature vectors, This is the weight matrix. For bias terms, This is the Sigmoid activation function.

7. The landslide susceptibility assessment method based on improved deep learning according to claim 1, characterized in that: The coordinate system of the multi-source geographic information data in S1 is WGS_1984_UTM_Zone_49N; The spatial resolution is uniformly set to 30m raster data.

Images