Method for predicting Qinghai-Tibet Plateau landslide based on artificial intelligence

Abstract

The invention relates to a Qinghai-Tibet Plateau landslide prediction method based on artificial intelligence, and the method comprises the following steps: S1, sorting Qinghai-Tibet Plateau landslide event data and influence factor data: carrying out the quantitative processing and normalization processing of 14 features, such as slope, orientation and the like, and dividing the features into three parts which are respectively used for constructing training, verification and test samples; s2, designing a model, and predicting the landslide in the permafrost region of the Qinghai-Tibet Plateau by using a machine learning technology: establishing a sample set containing 13 features; an interpretable artificial intelligence model is constructed; s3, selecting a black box model and performing classification by using a machine learning model; s4, interpreting the prediction of the random forest through an LIME model; s5, interpreting prediction of the random forest through an SHAP model; and S6, using the interpretable AI model to output and visualize the model. The method is high in precision, explainable and visualized in result, and can predict the susceptibility of the Qinghai-Tibet Plateau landslide.

Description

Technical Field

[0001] The present invention relates to the technical field of natural disaster prediction, and in particular to a method for predicting landslides on the Qinghai-Tibet Plateau based on artificial intelligence. Background Art

[0002] Landslides are one of the most common natural hazards and are becoming increasingly frequent and dangerous in both urban and rural areas, causing property damage, loss of life, and damage to infrastructure.

[0003] Landslides are triggered by a number of natural and man-made events, such as earthquakes, heavy rains, and extreme winds. Construction activities have further exacerbated the pressure on the environment. Due to the complexity of landslide hazards and the diversity of their causes, landslide hazard prediction has always been one of the difficulties. Many variables lead to the occurrence of landslides, and slope stability, topographic factors, weather and climate conditions, especially prolonged rainfall and snow melting, are often the causes of landslides. Heavy or continuous rainfall can saturate the soil, reduce soil viscosity, increase soil weight, and reduce slope stability. As snow melts rapidly and rainwater infiltrates the ground, it may further weaken the stability of the slope. Disturbances from urbanization and land use changes are also major factors in the occurrence of landslides. Understanding these variables is critical to mitigating the hazards associated with landslides and implementing effective disaster reduction strategies.

[0004] However, existing survey techniques and methods are time-consuming and labor-intensive in acquiring data, especially in the Qinghai-Tibet Plateau, where the topography is complex, climate change has strong spatial heterogeneity, the cryosphere is well developed, transportation and other infrastructure are widespread, and economic conditions are poor. Large-scale field surveys require considerable financial and human resources investment, and some field estimation methods may not be able to achieve high accuracy, especially in complex and changeable natural environments.

[0005] Landslide susceptibility is affected by many factors, including geological, meteorological, and topographic variables. Traditional landslide prediction models often rely on complex algorithms that are highly accurate but lack transparency and interpretability. Therefore, it is urgent to develop a new technical approach to overcome these drawbacks. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a method for predicting landslides on the Qinghai-Tibet Plateau based on artificial intelligence with high accuracy, explainability and visual results.

[0007] To solve the above problems, the method for predicting landslides on the Qinghai-Tibet Plateau based on artificial intelligence described in the present invention comprises the following steps:

[0008] S1 organizes data on landslide events and influencing factors on the Qinghai-Tibet Plateau:

[0009] Through various data sources, the observed and recorded landslide survey data in the permafrost area of ​​the Qinghai-Tibet Plateau were sorted out, and 14 features including slope aspect, curvature, earthquake, elevation, soil flow, lithology, normalized difference water index (NDWI), normalized difference vegetation index (NDVI), plan, precipitation intensity, profile, slope, human activities, and soil type were selected for quantification and normalization. The features were divided into three parts for constructing training, validation, and test samples respectively.

[0010] S2 designed a model to predict landslides in permafrost areas of the Qinghai-Tibet Plateau using machine learning techniques:

[0011] Using the existing data before and after geological disasters, a training dataset of landslide disasters in the permafrost area of ​​the Qinghai-Tibet Plateau was designed and established. After optimizing 14 features, a sample set containing 13 features was established, of which the input feature attributes were 1 geological feature processed by the embedding layer: landslide and 12 numerical features that were normalized before input: slope aspect, curvature, earthquake, elevation, soil flow, lithology, normalized difference water index (NDWI), normalized difference vegetation index (NDVI), plan, precipitation intensity, profile and slope. Then, an explainable artificial intelligence (XAI) model was constructed, whose architecture included explainability, machine learning model, knowledge graph construction, data preparation and subjective evaluation. Explanability was taken as the key factor, and the explainable AI model was used to explain the prediction of the random forest model.

[0012] S3 selects the black box model and uses the machine learning model for classification:

[0013] A black box model was selected and a machine learning model was used for classification. A neural network was added to the processing step. Then, based on the random forest algorithm, feature importance data was used to identify the key factors affecting the prediction of landslides on the Qinghai-Tibet Plateau. The landslide-prone training sample set on the Qinghai-Tibet Plateau was trained, the model was optimized, and the model was stored. The Random Forest (RF) algorithm was used to identify the features or variables that have the greatest impact on the occurrence of landslides on the Qinghai-Tibet Plateau, analyze the importance of the features, and evaluate the contribution of the features.

[0014] S4 explains the predictions of random forests through the LIME model:

[0015] A local interpretable model (LIME) for landslide susceptibility in the Qinghai-Tibet Plateau is constructed independently of the random forest model. The LIME model uses the complete training dataset to provide local explanations for the predicted reasoning (local explanations are model-agnostic) by explaining the positive and negative impacts of test vectors on landslide events.

[0016] S5 explains the predictions of random forests through the SHAP model:

[0017] The SHapley Additive exPlanations (SHAP) model of landslide susceptibility on the Qinghai-Tibet Plateau was constructed. The model quantifies the contribution of each feature in the model to the final prediction of the observed results, shows the relative contribution of each feature to the landslide occurrence outcome, and measures the individual contribution of each feature in the model prediction. The SHAP model is used to perform global and local interpretations and calculate the marginal contribution of features to the model output.

[0018] S6 uses explainable AI models to visualize model outputs.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] 1. This paper uses Explainable Artificial Intelligence (XAI) technology to create a very accurate landslide prediction model. Through SHAP (Additive Explanation), LIME (Local Interpretable Model Agnostic Explanation) and Random Forest (Feature Importance Analysis), it interprets the complex interrelationships between variables and provides an explanation of the factors that affect their prediction. By using XAI, we can fully understand the susceptibility of landslides on the Qinghai-Tibet Plateau and provide a deeper understanding of the mechanisms behind landslide events, which will help design a more effective landslide early warning system on the Qinghai-Tibet Plateau.

[0021] 2. The present invention is based on the characteristic attributes of landslides in permafrost areas of the Qinghai-Tibet Plateau. By constructing a prediction model in machine learning, the importance of parameters such as soil properties, slope, rainfall, and earthquake fitting is taken into account, thereby improving the accuracy of the prediction of the probability of landslide disasters on the Qinghai-Tibet Plateau.

[0022] 3. The present invention couples 14 characteristic attributes to construct training, validation and test samples. Based on these samples, the random forest method is used to predict the susceptibility of landslides in permafrost areas of the Qinghai-Tibet Plateau. The LIME model and the SHAP model are used to analyze the main factors of the landslide susceptibility of the Qinghai-Tibet Plateau. The characteristic attributes are linked to the landslide susceptibility of the Qinghai-Tibet Plateau to clarify different landslide susceptibility factors.

[0023] 4. The present invention uses an explainable artificial intelligence model to visualize the model output results, provide an understandable explanation for each model prediction, conduct verification and comparison, improve the overall reliability of the system, and improve the accuracy of model predictions.

[0024] 5. The method of the present invention can provide strong support for landslide susceptibility analysis and improve the accuracy of landslide susceptibility prediction on the Qinghai-Tibet Plateau, thereby providing a reference for landslide prediction, monitoring and early warning on the Qinghai-Tibet Plateau, in order to provide a reliable basis for landslide disaster prevention and mitigation work in the region. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The specific implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings.

[0026] Figure 1 It is the work flow chart of the present invention.

[0027] Figure 2 This is a flowchart of a method for predicting landslide prediction models in permafrost areas of the Qinghai-Tibet Plateau based on an explainable artificial intelligence model, as proposed in an embodiment of the present invention.

[0028] Figure 3 It is a schematic diagram of the landslide event prediction effect of the LIME model according to an embodiment of the present invention.

[0029] Figure 4 It is a schematic diagram of the prediction effect of the LIME model for non-landslide events according to an embodiment of the present invention.

[0030] Figure 5 It is a schematic diagram of the landslide event prediction effect of the SHAP model according to an embodiment of the present invention.

[0031] Figure 6 It is a schematic diagram of the prediction effect of non-landslide events of the SHAP model according to an embodiment of the present invention.

[0032] Figure 7 It is a schematic diagram of the dispersion effect of the SHAP value of each feature in an embodiment of the present invention.

[0033] Figure 8 It is a dependency graph of the entire data set according to an embodiment of the present invention.

[0034] Fig. 9It is a schematic diagram of the local interpretation effect of a landslide event according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] like Figure 1 As shown, a method for predicting landslides on the Qinghai-Tibet Plateau based on artificial intelligence includes the following steps:

[0036] S1 organizes data on landslide events and influencing factors on the Qinghai-Tibet Plateau:

[0037] Through various data sources, the observed and recorded landslide survey data in the permafrost area of ​​the Qinghai-Tibet Plateau were sorted out, and 14 features including slope aspect, curvature, earthquake, elevation, soil flow, lithology, normalized difference water index (NDWI), normalized difference vegetation index (NDVI), plan, precipitation intensity, profile, slope, human activities, and soil type were selected for quantification and normalization. The features were divided into three parts for constructing training, validation, and test samples respectively.

[0038] S2 designed a model to predict landslides in permafrost areas of the Qinghai-Tibet Plateau using machine learning techniques:

[0039] Using the existing data before and after geological disasters, a training dataset of landslide disasters in the permafrost area of ​​the Qinghai-Tibet Plateau was designed and established. After optimizing 14 features, a sample set containing 13 features was established, of which the input feature attributes were 1 geological feature processed by the embedding layer: landslide and 12 numerical features that were normalized before input: slope aspect, curvature, earthquake, elevation, soil flow, lithology, normalized difference water index (NDWI), normalized difference vegetation index (NDVI), plan, precipitation intensity, profile and slope. Then, an explainable artificial intelligence (XAI) model was constructed, whose architecture included explainability, machine learning model, knowledge graph construction, data preparation and subjective evaluation. Explanability was taken as the key factor, and the explainable AI model was used to explain the prediction of the random forest model.

[0040] S3 selects the black box model and uses the machine learning model for classification:

[0041] A black box model was selected and a machine learning model was used for classification. A neural network was added to the processing step. Then, based on the random forest algorithm, feature importance data was used to identify the key factors affecting the prediction of landslides on the Qinghai-Tibet Plateau. The landslide-prone training sample set on the Qinghai-Tibet Plateau was trained, the model was optimized, and the model was stored. The Random Forest (RF) algorithm was used to identify the features or variables that have the greatest impact on the occurrence of landslides on the Qinghai-Tibet Plateau, analyze the importance of the features, and evaluate the contribution of the features.

[0042] S4 explains the predictions of random forests through the LIME model:

[0043] A local interpretable model (LIME) for landslide susceptibility in the Qinghai-Tibet Plateau is constructed independently of the random forest model. The LIME model uses the complete training dataset to provide local explanations for the predicted reasoning (local explanations are model-agnostic) by explaining the positive and negative impacts of test vectors on landslide events.

[0044] S5 explains the predictions of random forests through the SHAP model:

[0045] The SHapley Additive exPlanations (SHAP) model of landslide susceptibility on the Qinghai-Tibet Plateau was constructed. The model quantifies the contribution of each feature in the model to the final prediction of the observed results, shows the relative contribution of each feature to the landslide occurrence outcome, and measures the individual contribution of each feature in the model prediction. The SHAP model is used to perform global and local interpretations and calculate the marginal contribution of features to the model output.

[0046] S6 uses explainable AI models to visualize model outputs.

[0047] Embodiment A method for predicting landslides on the Qinghai-Tibet Plateau based on artificial intelligence comprises the following steps:

[0048] S1 organizes data on landslide events and influencing factors on the Qinghai-Tibet Plateau:

[0049] Through various data sources, the observed and recorded landslide survey data in the permafrost area of ​​the Qinghai-Tibet Plateau were sorted out and preprocessed. Before the analysis, the influencing factors were selected according to the specific conditions of the Qinghai-Tibet Plateau, from which the geological environment factors of landslide occurrence, historical landslide data, physical laws of landslides and other factors were obtained. By extracting the characteristics of landslide-related data, the possible mutual combination relationship between various influencing factors that induce landslide disasters was comprehensively considered.

[0050] Fourteen features were selected, namely, aspect, curvature, earthquake, elevation, flow, lithology, normalized difference water index (NDWI), normalized difference vegetation index (NDVI), plan, precipitation intensity, profile, slope, human activities and soil type. These 14 features were effectively quantified and normalized, and divided into three parts for constructing training, validation and testing samples respectively. A landslide susceptibility model was constructed and analyzed for the permafrost area of ​​the Qinghai-Tibet Plateau to determine the probability of landslides in the Qinghai-Tibet Plateau in the future.

[0051] S2 designed a model to predict landslides in permafrost areas of the Qinghai-Tibet Plateau using machine learning techniques:

[0052] In order to make more effective predictions and forecasts of landslide occurrences, understand the relationships of prediction models and the importance of all features in model outputs, apply prepared samples to artificial intelligence technology, and train artificial intelligence models for landslide susceptibility prediction.

[0053] Since landslides are affected by many factors, based on field investigations and landslide disaster information cataloging, according to the multi-source data of geological disasters such as controlling factors, influencing factors, inducing factors, deformation signs, existing geological disasters, etc., and mainly using the comparative data before and after the existing geological disasters, a training dataset of landslide disasters in permafrost areas of the Qinghai-Tibet Plateau was designed and established. An optimization method for constructing a training sample set of landslide disasters in the Qinghai-Tibet Plateau was proposed. After optimizing 14 features, a sample set containing 13 features was established. The input feature attributes are 1 geological feature processed by the embedding layer: landslide, and 12 numerical features that have been normalized before input: slope aspect, curvature, earthquake, elevation, soil flow, lithology, normalized difference water index (NDWI), normalized difference vegetation index (NDVI), plan, precipitation intensity, profile, and slope. The architecture is mainly determined by the following factors: rainfall intensity, earthquakes, altitude, slope angle, soil flow and vegetation cover.

[0054] By using a normalized version of the Qinghai-Tibet Plateau landslide prediction dataset, a multi-step procedure was designed to predict the landslide susceptibility of the Qinghai-Tibet Plateau using machine learning techniques. To avoid overfitting and accurately predict landslides on the Qinghai-Tibet Plateau, an explainable artificial intelligence (XAI) was constructed, which uses a simple deep learning model to make diagnosis easier.

[0055] The designed architecture includes interpretability, machine learning models, knowledge graph construction, data preparation, and subjective evaluation. Finally, interpretability is taken as a key factor, and an explanatory AI model is used to explain the predictions of the random forest model, thereby achieving the prediction and prevention of future landslides.

[0056] S3 selects the black box model and uses the machine learning model for classification:

[0057] Select the Black Box model and use the machine learning model for classification. The Black Box model uses feature importance data to identify key factors that affect the prediction. A neural network is added to the processing step to improve the model's ability to recognize complex patterns in the input. With the addition of a neural network, the Black Box model is improved in recognizing complex patterns in the input.

[0058] Based on the random forest algorithm, feature importance data is used to identify the key factors affecting the prediction of landslides on the Qinghai-Tibet Plateau. Learning training, model optimization and model storage are carried out on the landslide-prone training sample set on the Qinghai-Tibet Plateau. The Random Forest (RF) algorithm is used to identify the features or variables that have the greatest impact on the occurrence of landslides on the Qinghai-Tibet Plateau, analyze the importance of features, and evaluate feature contributions.

[0059] By analyzing the importance of 14 features, the possible factors leading to landslides in the Qinghai-Tibet Plateau are estimated. These features include slope aspect, curvature, earthquake, elevation, soil flow, lithology, normalized difference water index (NDWI), normalized difference vegetation index (NDVI), plan, precipitation intensity, profile and slope, human activities, and soil type.

[0060] From the results of feature importance, it can be seen that rainfall intensity, slope and lithology are all very important factors for the landslide risk in the Qinghai-Tibet Plateau. According to the accuracy of the confusion matrix analysis and the importance of factor features output by the model, when the feature importance is the highest, that is, the feature contribution is the largest, the intervention can be given priority in the landslide-prone features to mitigate the occurrence of landslides by targeting the most influential factors.

[0061] S4 explains the predictions of random forests through the LIME model:

[0062] In order to explain the predictions of random forests, a local interpretable model (Local Interpretable Model-Agnostic Explanations, LIME) of landslide susceptibility in the Qinghai-Tibet Plateau was constructed. The LIME model is independent of the random forest model. The LIME model uses the complete training data set to provide local explanations for the reasoning of the predictions (local explanation model independence). Using the trained LIME model, only each sample is explained, generating a new data set by explaining the positive and negative impacts of the test vector on the landslide event. While keeping the complexity of the interpretable model low, the minimization loss L (loss) is minimized as much as possible (the closeness of the interpretable model to the original model, that is, the fit to the dataset generated after the disturbance).

[0063] The LIME algorithm helps us understand the decision-making process behind a specific model result. The algorithm steps of the LIME model are as follows:

[0064] ⑴Start.

[0065] ⑵Input data set.

[0066] ⑶ Select specific instances from the dataset for which to generate explanations.

[0067] (4) Fit a local interpretable model to the samples and corresponding predictions.

[0068] ⑸ Calculate feature importance / / Extract feature weights or coefficients from the fitted interpretable model to determine the importance of each feature to the prediction of the selected instance. Positive coefficients indicate features that contribute positively to the prediction, while negative coefficients indicate features that contribute negatively.

[0069] ⑹Visualize feature importance scores or coefficients to understand which features have the greatest impact on the prediction of the selected instance.

[0070] ⑺Determine which features play an important role in the prediction and how their values ​​affect the results.

[0071] ⑻Stop.

[0072] The LIME model is used to explain the landslide events on the Qinghai-Tibet Plateau. The reasoning provided by LIME for each result x is obtained using the following formula. The LIME algorithm is shown in formula (1):

[0073]

[0074] Where: Represents the defined objective function; G is the set of interpretable models in the mathematical equation, G is processed using the random forest algorithm; g∈G is a measure of the proximity between instances z and x; Ω(g) is a measure of the complexity of the explanation provided by g∈G, and when Ω(g) (the complexity of the explanation model) is low enough to be understood by humans, the optimal solution of the objective function is obtained by minimizing the L function. L is a metric that measures the degree of inconsistency between the approximation provided by g and the actual function f within the range defined by π(ⅹ).

[0075] By interpreting the predictions given by the LIME prediction model for specific test features, LIME can make more certain judgments about landslide and non-landslide events. Using the LIME model to interpret landslide events on the Qinghai-Tibet Plateau, such as Figure 3 As shown in the figure, the confidence level of the predicted landslide is 94%. Since the altitude is equal to 1, the profile is greater than 4, the rainfall intensity is greater than 4, the slope aspect ratio is greater than 2, the curvature is less than 2, the NDWI is greater than 2, the earthquake is greater than 2, and the NDVI is greater than 2, the model predicts that this event is a landslide event. Use the LIME model to predict a non-landslide event on the Qinghai-Tibet Plateau, such as Figure 4 As shown: The confidence level is 83%. Considering factors such as orientation and curvature, since the rainfall intensity is less than 3, the NDVI is greater than 4, the altitude is greater than 2, the profile is less than 2, and the earthquake is less than 2, the model predicts that this event is a non-landslide event.

[0076] S5 explains the predictions of random forests through the SHAP model:

[0077] In order to explain the predictions of random forests, the SHapley Additive exPlanations (SHAP) model of landslide susceptibility in the Qinghai-Tibet Plateau was constructed. SHAP is an XAI method that uses cooperative game theory to evaluate the importance of variables with multicollinearity. The model quantifies the contribution of each feature in the model to the final prediction of the observed results, which can show the relative contribution of each feature to the landslide occurrence outcome and measure the individual contribution of each feature in the model prediction.

[0078] The SHAP value is calculated as a weighted average of the impact of adding a feature on the prediction of all possible feature combinations. This method takes into account the different effects of a feature's interaction with other features in the model. It provides a diagram to explain the impact of the predictor variable (feature) on the dependent variable, and thoroughly explains the components and framework of the model, more objectively and comprehensively evaluating the importance of feature factors, thereby improving the performance and accuracy of the model.

[0079] The SHAP model is used for global and local interpretation, and the marginal contribution of features to the model output is calculated. The results generated by SHAP are easier to interpret and more accurate. SHAP values ​​are similar to regression coefficients, which can be positive or negative, and large or small. A SHAP value greater than 0 indicates a positive impact, and a value less than 0 indicates a negative impact.

[0080] The algorithm steps of the SHAP model are as follows:

[0081] ⑴Start.

[0082] ⑵Input data set.

[0083] ⑶ Use a trained machine learning model (the proposed method uses random forest).

[0084] ⑷ Initialize the SHAP interpreter object using the trained machine learning model and dataset.

[0085] ⑸ Calculate the SHAP value of all instances in the dataset.

[0086] ⑹ Generate summary plots or statistics to visualize the overall impact of features on model predictions.

[0087] ⑺ Explain a single prediction or a specific instance, analyzing a single SHAP value to understand why the model made a specific prediction.

[0088] ⑻ Use force plots or other visualization techniques to illustrate the impact of each feature on the model’s predictions for that instance.

[0089] ⑼ Feature Importance Ranking - Sort features according to their average absolute SHAP value across all instances.

[0090] ⑽Model Validation and Improvement — Validate that SHAP explains the performance of the model using appropriate evaluation metrics and techniques.

[0091] ⑾ Improve the model based on insights gained from interpreting SHAP values ​​(such as feature selection, hyperparameter tuning, or data preprocessing).

[0092] ⑿Stop.

[0093] The SHAP model is used to explain landslide events on the Qinghai-Tibet Plateau. The algorithm of SHAP value (Shapley value) is as shown in formulas (2) and (3):

[0094]

[0095] Where: is the Shapley value of feature i, indicating the impact of feature i on the overall prediction; N represents the set of all features; n represents the number of features in N; S represents the subset containing feature i, and N\{i} represents the set after feature i is deleted from set N; = means that set S is contained in set N; |S| is the number of features in subset S, n! represents the factorial of n; v(S) represents a real function v on N as the feature function; S∪{i} represents the set with feature i added from set S; g(z′) represents the model output for a given feature vector x, g is the explanation model; is the Shapley value without considering any features; Z i ′ is the simplified feature vector after removing feature i; M is the number of features.

[0096] The predictions given by the SHAP model for specific test features were interpreted, and a SHAP analysis of the factors affecting landslide events on the Qinghai-Tibet Plateau was performed (see Figure 5 ), the length of the bar graph shows the relative importance of the features. The longer the bar, the higher the priority of the corresponding feature in predicting landslides. Among all the features, altitude has the highest priority in predicting landslides on the Qinghai-Tibet Plateau. Figure 6 ), showing the average impact of each feature on the model output, with the “flow” feature having the largest positive impact on the model output. In contrast, the “lithology” feature has the largest negative impact, meaning that, on average, higher flow values ​​tend to increase the model’s predictions of landslides on the Qinghai-Tibet Plateau. In contrast, higher lithology values ​​tend to reduce the probability of landslide events on the Qinghai-Tibet Plateau. Features are further analyzed using SHAP summary plots, which combine feature importance with feature effects. Each point on the SHAP summary plot is a Shapley value for a feature and an instance. Based on the dispersion of the SHAP values ​​for each feature (see Figure 7 ), showing that rainfall intensity, slope and profile characteristics strongly affect the probability of landslides on the Qinghai-Tibet Plateau. SHAP dependency diagram (see Figure 8) illustrates how a given feature affects the entire data set. By plotting the value of the feature slope orientation on the x-axis and then plotting the SHAP value of the feature slope orientation on the y-axis, it can be found that the interaction effect drives the vertical dispersion of the SHAP value at a single feature value, and there is a potential interaction between feature values. In addition, a local explanation of a landslide event on the Qinghai-Tibet Plateau is given (see Fig. 9 ), and found that among all the characteristics, lithological characteristics have the greatest positive impact on landslide outcomes, which is 0.36, with a probability value of 0.91 and an expected value (EV) of 0.484 for landslide occurrence in machine learning simulations.

[0097] S6 uses explainable AI models to visualize model outputs.

[0098] Through comprehensive analysis and visualization of model output, Figure 2 A flow chart of a preferred landslide prediction method provided in an embodiment of the present invention examines the results of the LIME model and the SHAP model and the interpretation effect of the random forest model, and generates an accurate landslide prediction based on the obtained results.

[0099] Both SHAP and LIME models use the complete training dataset to provide global and local explanations. The SHAP model allows global and local explanations, while the LIME model only allows local explanations for each instance. Through its model results, the results generated by SHAP are easier to interpret and more accurate. Compared with LIME, SHAP has a higher computational complexity. The proposed model better mines the relationship between landslide occurrence and characteristic factors, provides a clear explanation for the characteristics that lead to landslide susceptibility, and effectively improves the model prediction accuracy.

[0100] Using an explainable AI model, the output of the model is visualized at the final stage. With this visualization, questioning the random forest model becomes easier and the prediction accuracy of the random forest model is improved, which makes it possible to conduct a thorough analysis of the results.

[0101] The model also completed a subjective evaluation using real-time data to ensure that the results met human standards and expectations. The human-centric viewpoint compares the model's predictions with domain expertise and common sense to build a real-time monitoring and early warning system, thereby improving the overall reliability of the system and achieving rapid response to landslide hazards.

Claims

1. A method for predicting landslides on the Qinghai-Tibet Plateau based on artificial intelligence, comprising the following steps: S1 organizes data on landslide events and influencing factors on the Qinghai-Tibet Plateau: Through various data sources, the observed and recorded landslide survey data in the permafrost area of ​​the Qinghai-Tibet Plateau were sorted out, and 14 features including slope orientation, curvature, earthquake, altitude, soil flow, lithology, normalized difference water index, normalized difference vegetation index, plane, rainfall intensity, profile and slope, human activities, and soil type were selected for quantification and normalization. They were divided into three parts, which were used to construct training, validation, and test samples respectively; S2 designed a model to predict landslides in permafrost areas of the Qinghai-Tibet Plateau using machine learning techniques: Using the existing comparative data before and after geological disasters, a training dataset for landslide disasters in the permafrost area of ​​the Qinghai-Tibet Plateau was designed and established. After optimizing 14 features, a sample set containing 13 features was established, of which the input feature attributes were 1 geological feature processed by the embedding layer: landslide and 12 numerical features that had been normalized before input: slope orientation, curvature, earthquake, altitude, soil flow, lithology, normalized difference water index, normalized difference vegetation index, plane, rainfall intensity, profile and slope; and then an explainable artificial intelligence model was constructed, whose architecture included explainability, machine learning model, knowledge graph construction, data preparation and subjective evaluation, and interpretability was taken as a key factor, using an explainable AI model to explain the predictions of the random forest model; S3 selects the black box model and uses the machine learning model for classification: Select a black box model and use a machine learning model for classification, and add a neural network in the processing step; then based on the random forest algorithm, use feature importance data to identify the key factors affecting the prediction of landslides on the Qinghai-Tibet Plateau, conduct learning training, model optimization and model storage on the landslide-prone training sample set on the Qinghai-Tibet Plateau, identify the features or variables that have the greatest impact on the occurrence of landslides on the Qinghai-Tibet Plateau through the random forest algorithm, analyze the importance of features, and evaluate feature contributions; S4 explains the predictions of random forests through the LIME model: A local explanation LIME model for landslide susceptibility in the Qinghai-Tibet Plateau is constructed independently of the random forest model. The LIME model uses the complete training dataset to provide local explanations for the predicted reasoning by explaining the positive and negative impacts of the test vectors on landslide events. S5 explains the predictions of random forests through the SHAP model: The additive explanatory SHAP model of landslide susceptibility in the Qinghai-Tibet Plateau was constructed. This model quantifies the contribution of each feature in the model to the final prediction of the observed results, and shows the relative contribution of each feature to the landslide occurrence outcome, measuring the individual contribution of each feature in the model prediction; The SHAP model is used to perform global and local explanations and calculate the marginal contribution of features to the model output. S6 uses explainable AI models to visualize model outputs.

Images