Non-landslide sample selection optimization method and system based on information amount

Abstract

The invention relates to the technical field of geological disaster assessment, in particular to a non-landslide sample selection optimization method and system based on information amount, and the method comprises the steps: collecting historical landslide data and environmental factor data in a to-be-predicted region, carrying out the importance evaluation of environmental factors, and screening out landslide characteristic factors which have large impact on the landslide susceptibility; carrying out information quantity analysis on the screened landslide characteristic factors, carrying out weight assignment on the factor information quantity according to the importance of the landslide characteristic factors to obtain comprehensive information quantity, and selecting reliable non-landslide samples in an area with relatively low comprehensive information quantity; inputting the data into a prediction model to obtain a prediction result; and carrying out landslide susceptibility evaluation on the to-be-predicted region, generating a landslide susceptibility partition map, and formulating targeted disaster prevention and reduction measures. According to the method, a machine learning algorithm and an information amount method are combined, the objectivity of weight assignment of the influence factors is improved, so that more representative non-landslide samples are selected, and the precision of a landslide susceptibility prediction model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster assessment, and in particular to an information-based non-landslide sample selection optimization method and system. Background Art

[0002] Landslides, as a common geological disaster, are extremely harmful to human society. They are sudden and destructive, often causing casualties, property losses and environmental damage.

[0003] The selection of non-landslide samples is one of the key steps in building a landslide susceptibility prediction model. The selection of non-landslide samples directly affects the effect of model training and prediction accuracy. However, the selection of non-landslide samples faces many challenges. The existing non-landslide sample selection methods mainly include the whole-area random selection method, the landslide buffer method, the similarity-based method, and the sampling based on specific conditions. The whole-area random selection method randomly selects a certain number of sample points in the area to be predicted as non-landslide samples. This method is simple and easy to implement, but the sample distribution is uneven and the representativeness is insufficient, making it difficult to ensure the effect of model training. The landslide buffer method sets a buffer zone around the landslide. The area within the buffer zone is considered to be the area affected by the landslide, and the area outside the buffer zone is considered to be the non-landslide area. This method can exclude some areas affected by the landslide, but the determination of the buffer zone radius lacks a unified standard, which may lead to uneven sample distribution. The similarity-based method selects samples that are similar to the landslide samples in environmental factors as non-landslide samples. This method can ensure the representativeness of the samples, but requires a large amount of environmental factor data to support it, and it is difficult to ensure the number of samples. Sampling based on specific conditions selects non-landslide samples according to specific conditions. For example, samples from areas with gentle slopes, high vegetation coverage, and less rainfall are selected as non-landslide samples. This method is highly targeted, but may result in insufficient sample quantities. Summary of the invention

[0004] In order to overcome the limitations of the existing non-landslide sample selection method, the present invention provides the following technical solutions:

[0005] According to a first aspect of the present invention, the present invention proposes an information-based non-landslide sample selection optimization method, comprising:

[0006] Collect historical landslide data and environmental factor data in the area to be predicted, and use the recursive feature elimination method to screen out environmental factors that have a great impact on landslide susceptibility as landslide characteristic factors;

[0007] Randomly select initial non-landslide samples in the area to be predicted, use the random forest model to evaluate the importance of the landslide characteristic factors, assign weights to the information model according to the importance evaluation scores, divide extremely low and low information areas based on the calculation results of the information value, and select reliable non-landslide samples from the extremely low and low information areas in the enhanced information model;

[0008] Using a random forest model to construct a landslide susceptibility prediction model, constructing a model training set and a model test set based on the historical landslide data and reliable non-landslide samples, and inputting the landslide susceptibility prediction model to obtain a first prediction result;

[0009] The prediction accuracy and uncertainty of the first prediction result of the landslide susceptibility prediction model are evaluated by using ROC curve, confusion matrix and relative density method;

[0010] The landslide susceptibility prediction model is used to evaluate the landslide susceptibility of the predicted area, generate a landslide susceptibility zoning map, and formulate targeted disaster prevention and mitigation measures.

[0011] Furthermore, the collecting of historical landslide data and environmental factor data in the area to be predicted, and using a recursive feature elimination method to screen out landslide characteristic factors that have a great impact on landslide susceptibility, also includes:

[0012] Collect data on topography, geology, climate, soil, and vegetation environmental factors in the area to be predicted;

[0013] A machine learning algorithm is used to analyze the impact of the environmental factor data on the occurrence of landslides, and factors that have a great impact on the occurrence of landslides are screened out as landslide characteristic factors, which serve as the basis for subsequent information calculations.

[0014] Furthermore, randomly selecting initial non-landslide samples in the area to be predicted, using a random forest model to evaluate the importance of the landslide characteristic factors, assigning weights to the information content model according to the importance evaluation scores, dividing extremely low and low information content areas based on the calculation results of the information content values, and selecting reliable non-landslide samples from the extremely low and low information content areas in the enhanced information content model, further comprising:

[0015] The information content of the screened landslide characteristic factors is calculated to obtain the information content value of each landslide characteristic factor at different levels;

[0016] The importance of the landslide characteristic factors is evaluated using a random forest model to obtain the importance score of each landslide characteristic factor;

[0017] According to the importance scores of different landslide characteristic factors, the information value is weighted to obtain the enhanced information value;

[0018] Based on the calculation result of the enhanced information value, extremely low and low information content areas are divided, and reliable non-landslide samples are selected from the extremely low and low information content areas in the enhanced information content model.

[0019] Furthermore, the method of using the random forest model to construct a landslide susceptibility prediction model, constructing a model training set and a model test set based on the historical landslide data and reliable non-landslide samples, and inputting the landslide susceptibility prediction model to obtain a first prediction result, further includes:

[0020] Conduct a quality assessment of the representativeness, quantity, distribution and independence of the selected non-landslide samples to verify the credibility of the non-landslide samples;

[0021] The random forest model was selected to use the optimized non-landslide samples together with the landslide samples for model training, and the parameters of the landslide susceptibility prediction model were optimized using grid search and cross-validation methods;

[0022] The parameters of the landslide susceptibility prediction model include at least the number of decision trees, the maximum depth, and the minimum number of samples;

[0023] After training, the landslide susceptibility prediction model parameters are used to predict the landslide susceptibility in unknown areas.

[0024] Furthermore, the method of using the ROC curve, confusion matrix and relative density method to evaluate the prediction accuracy and uncertainty of the first prediction result of the landslide susceptibility prediction model also includes:

[0025] The prediction accuracy and consistency of the landslide susceptibility prediction model were evaluated using ROC curve, confusion matrix and Kappa coefficient indicators;

[0026] The ROC curve shows the performance of the model, and the area under the curve (AUC) analyzes the accuracy of the prediction results. The higher the AUC value, the better the model performance;

[0027] Confusion matrix analysis of the classification results of the landslide susceptibility prediction model, including accuracy, precision, and recall;

[0028] The Kappa coefficient evaluates the consistency of the landslide susceptibility prediction model, and the higher the value, the better the model consistency;

[0029] The relative density method is used to analyze the susceptibility distribution characteristics predicted by the landslide susceptibility prediction model and compare them with the actual landslide distribution.

[0030] According to a second aspect of the present invention, the present invention proposes a non-landslide sample selection optimization system based on information quantity, comprising:

[0031] The acquisition module collects historical landslide data and environmental factor data in the area to be predicted, and uses the recursive feature elimination method to screen out environmental factors that have a great impact on the susceptibility of landslides as landslide characteristic factors;

[0032] The sample screening module randomly selects initial non-landslide samples in the area to be predicted, uses the random forest model to evaluate the importance of the landslide characteristic factors, assigns weights to the information model according to the importance evaluation scores, divides extremely low and low information areas based on the calculation results of the information value, and selects reliable non-landslide samples from the extremely low and low information areas in the enhanced information model;

[0033] A model testing module, using a random forest model to build a landslide susceptibility prediction model, building a model training set and a model testing set based on the historical landslide data and reliable non-landslide samples, and inputting the landslide susceptibility prediction model to obtain a first prediction result;

[0034] A model evaluation module, which uses ROC curve, confusion matrix and relative density method to evaluate the prediction accuracy and uncertainty of the first prediction result of the landslide susceptibility prediction model;

[0035] The landslide assessment module uses the landslide susceptibility prediction model to conduct landslide susceptibility assessment on the predicted area, generates a landslide susceptibility zoning map, and formulates targeted disaster prevention and mitigation measures.

[0036] Furthermore, the acquisition module further includes:

[0037] Collect data on topography, geology, climate, soil, and vegetation environmental factors in the area to be predicted;

[0038] A machine learning algorithm is used to analyze the impact of the environmental factor data on the occurrence of landslides, and environmental factors that have a great impact on the occurrence of landslides are screened out as landslide characteristic factors, which serve as the basis for subsequent information calculation.

[0039] Furthermore, the sample screening module further includes:

[0040] The information content of the screened landslide characteristic factors is calculated to obtain the information content value of each landslide characteristic factor at different levels;

[0041] The importance of the landslide characteristic factors is evaluated using a random forest model to obtain the importance score of each landslide characteristic factor;

[0042] According to the importance scores of different landslide characteristic factors, the information value is weighted to obtain the enhanced information value;

[0043] Based on the calculation result of the enhanced information value, extremely low and low information areas are divided, and reliable non-landslide samples are selected from the extremely low and low information areas in the enhanced information model.

[0044] Furthermore, the model testing module further includes:

[0045] Conduct a quality assessment of the representativeness, quantity, distribution and independence of the selected non-landslide samples to verify the credibility of the non-landslide samples;

[0046] The random forest model was selected to use the optimized non-landslide samples together with the landslide samples for model training, and the parameters of the landslide susceptibility prediction model were optimized using grid search and cross-validation methods;

[0047] The parameters of the landslide susceptibility prediction model include at least the number of decision trees, the maximum depth, and the minimum number of samples;

[0048] After training, the landslide susceptibility prediction model parameters are used to predict the landslide susceptibility in unknown areas.

[0049] Furthermore, the model evaluation module further includes:

[0050] The prediction accuracy and consistency of the landslide susceptibility prediction model were evaluated using ROC curve, confusion matrix and Kappa coefficient indicators;

[0051] The ROC curve shows the performance of the model, and the area under the curve (AUC) analyzes the accuracy of the prediction results. The higher the AUC value, the better the model performance;

[0052] Confusion matrix analysis of the classification results of the landslide susceptibility prediction model, including accuracy, precision, and recall;

[0053] The Kappa coefficient evaluates the consistency of the landslide susceptibility prediction model, and the higher the value, the better the model consistency;

[0054] The relative density method is used to analyze the susceptibility distribution characteristics predicted by the landslide susceptibility prediction model and compare them with the actual landslide distribution.

[0055] The present invention relates to the field of geological disaster assessment technology, and in particular to a non-landslide sample selection optimization method and system based on information volume, collecting historical landslide data and environmental factor data in the area to be predicted, screening out factors with great influence on landslide susceptibility as landslide characteristic factors; selecting initial non-landslide samples, evaluating the importance of landslide characteristic factors and selecting reliable non-landslide samples; constructing a landslide susceptibility prediction model and a training set and a test set, inputting the prediction model to obtain a first prediction result; evaluating the prediction accuracy and uncertainty of the landslide susceptibility prediction model, evaluating the landslide susceptibility of the area to be predicted, generating a landslide susceptibility zoning map, and formulating targeted disaster prevention and mitigation measures. The present invention combines a machine learning algorithm with an information volume method, improves the objectivity of the weight assignment of influencing factors, thereby selecting more representative non-landslide samples, improving the accuracy of the landslide susceptibility prediction model and reducing uncertainty. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A workflow diagram of a non-landslide sample selection optimization method based on information quantity claimed in an embodiment of the present invention;

[0057] Figure 2 A schematic diagram of the structure of a random forest model used in an information-based non-landslide sample selection optimization method claimed in an embodiment of the present invention;

[0058] Figure 3 A landslide impact factor importance ranking diagram for a non-landslide sample selection optimization method based on information quantity claimed in an embodiment of the present invention;

[0059] Figure 4 An information quantity diagram of landslide influencing factors for a non-landslide sample selection optimization method based on information quantity claimed in an embodiment of the present invention;

[0060] Figure 5 An enhanced information map obtained by weighting according to importance in a non-landslide sample selection optimization method based on information quantity claimed in an embodiment of the present invention;

[0061] Figure 6 A distribution map of the locations of landslide samples and non-landslide samples according to a non-landslide sample selection optimization method based on information quantity as claimed in an embodiment of the present invention;

[0062] Figure 7 A ROC curve comparison diagram of different non-landslide sample selection mode models of a non-landslide sample selection optimization method based on information quantity claimed in an embodiment of the present invention;

[0063] Figure 8A landslide susceptibility zoning result map based on EIV-RF of a non-landslide sample selection optimization method based on information quantity claimed in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] The terms "first", "second" and "third" in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined as "first", "second" and "third" can explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "multiple" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined. All directional indications (such as up, down, left, right, front, back...) in the embodiments of the present invention are only used to explain the relative position relationship, movement, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication also changes accordingly. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices.

[0066] Reference to an "embodiment" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiment may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0067] The landslide susceptibility prediction model is an important tool for assessing the probability of landslides in a specific area. Its core idea is to analyze the correlation between historical landslides and environmental factors, and use this information to predict the possibility of future landslides. Common landslide susceptibility prediction models mainly include models based on expert experience, models based on physical mechanisms, and data-driven models. The model based on expert experience mainly relies on the experience and knowledge of experts to evaluate the possibility of landslides through qualitative analysis. Its advantages are simplicity and ease of use, but its disadvantages are strong subjectivity and lack of objectivity. The model based on physical mechanisms is based on the physical mechanism of landslides, and uses mechanical principles and numerical simulation methods to predict the possibility of landslides. Its advantages are solid theoretical foundations, but its disadvantages are complex models, difficulty in obtaining parameters, and difficulty in applying to large-scale predictions. The data-driven model mainly uses machine learning algorithms to establish a prediction model by analyzing historical landslide data and environmental factor data. Its advantages are strong objectivity, high prediction accuracy, and easy promotion and application.

[0068] The selection of non-landslide samples is one of the key steps in building a landslide susceptibility prediction model. The selection of non-landslide samples directly affects the effect of model training and prediction accuracy. However, the selection of non-landslide samples faces the following challenges:

[0069] (I) Uneven sample distribution: Landslides occur in a spatial clustering pattern, which results in uneven distribution of landslide samples in space. Traditional non-landslide sample selection methods often fail to take this into account, resulting in uneven sample distribution and affecting the effect of model training.

[0070] (II) Insufficient sample representativeness: Traditional non-landslide sample selection methods often lack specificity and cannot effectively distinguish the impact of different types of terrain, geology, hydrology and other environmental factors on landslide occurrence, resulting in insufficient sample representativeness and affecting the model prediction accuracy.

[0071] (III) Insufficient number of samples: Landslides are low-probability events, resulting in a relatively small number of landslide samples. Traditional non-landslide sample selection methods often require a large number of non-landslide samples for training, resulting in insufficient sample numbers and affecting the stability of model training.

[0072] According to the first embodiment of the present invention, the present invention claims a non-landslide sample selection optimization method based on information volume, referring to Figure 1 ,include:

[0073] Collect historical landslide data and environmental factor data in the area to be predicted, and use the recursive feature elimination method to screen out environmental factors that have a great impact on landslide susceptibility as landslide characteristic factors;

[0074] Randomly select initial non-landslide samples in the area to be predicted, use the random forest model to evaluate the importance of the landslide characteristic factors, assign weights to the information model according to the importance evaluation scores, divide extremely low and low information areas based on the calculation results of the information value, and select reliable non-landslide samples from the extremely low and low information areas in the enhanced information model;

[0075] Using a random forest model to construct a landslide susceptibility prediction model, constructing a model training set and a model test set based on the historical landslide data and reliable non-landslide samples, and inputting the landslide susceptibility prediction model to obtain a first prediction result;

[0076] The prediction accuracy and uncertainty of the first prediction result of the landslide susceptibility prediction model are evaluated by using ROC curve, confusion matrix and relative density method;

[0077] The landslide susceptibility prediction model is used to evaluate the landslide susceptibility of the predicted area, generate a landslide susceptibility zoning map, and formulate targeted disaster prevention and mitigation measures.

[0078] Furthermore, the method of collecting historical landslide data and environmental factor data in the area to be predicted, and using a recursive feature elimination method to screen out environmental factors that have a great impact on the susceptibility of landslides as landslide characteristic factors, also includes:

[0079] Collect data on topography, geology, climate, soil, and vegetation environmental factors in the area to be predicted;

[0080] A machine learning algorithm is used to analyze the impact of the environmental factor data on the occurrence of landslides, and factors that have a great impact on the occurrence of landslides are screened out as landslide characteristic factors, which serve as the basis for subsequent information calculations.

[0081] Among them, in this embodiment, the collected characteristic factor data include: various types of data including topography, geological elements, vegetation impact, meteorology, hydrology and human activities;

[0082] The feature index selection includes elevation, slope, aspect, plan curvature, profile curvature, soil dataset, precipitation, humidity, population density, and vegetation index (NDVI).

[0083] The feature selection process includes: selecting the random forest algorithm as the basic learner, which has good generalization ability and maturity;

[0084] Divide the data set into training set and test set, using 70% of the data as the training set and 30% of the data as the test set;

[0085] Use the random forest algorithm to train the model on the training set and calculate the performance indicators of the model;

[0086] Using the characteristics of the random forest algorithm, the contribution of each feature factor to the model prediction performance is calculated. Feature Importance is an important output in the random forest algorithm, which is used to measure the contribution of each feature to the model prediction performance. First, the model performance (such as accuracy or mean square error) on the original data set is calculated. For each feature, the value of the feature is randomly shuffled, and then the model performance is calculated again. Feature importance is measured by the difference between the original model performance and the shuffled model performance. The larger the difference, the more important the feature.

[0087] Sort the feature factors according to the feature importance score, with the features with higher importance at the front;

[0088] Select the least important characteristic factor to eliminate;

[0089] Remove the selected feature factors from the feature set and update the feature set;

[0090] According to the feature importance, the feature factors that have the greatest impact on the model are selected as the final feature set.

[0091] During the screening process, factors such as the importance, relevance, and data quality of the factors can be comprehensively considered to ensure that the final screened factor set can effectively reflect the driving factors of landslides.

[0092] By analyzing the relationship between each factor and the frequency of landslide occurrence, as well as the correlation between factors, we can determine which factors have a greater impact on landslide occurrence and retain these factors as model inputs.

[0093] Clean the filtered data, including processing missing values, outlier detection and processing, data format conversion, and removal of duplicate data.

[0094] Missing values ​​can be processed through interpolation, deletion, etc., outliers can be processed through statistical analysis, data smoothing, etc., data format conversion can convert data from different sources into a unified format, and removing duplicate data can avoid data redundancy.

[0095] Furthermore, randomly selecting initial non-landslide samples in the area to be predicted, using a random forest model to evaluate the importance of the landslide characteristic factors, assigning weights to the information content model according to the importance evaluation scores, dividing extremely low and low information content areas based on the calculation results of the information content values, and selecting reliable non-landslide samples from the extremely low and low information content areas in the enhanced information content model, further comprising:

[0096] The information content of the screened landslide characteristic factors is calculated to obtain the information content value of each landslide characteristic factor at different levels;

[0097] The importance of the landslide characteristic factors is evaluated using a random forest model to obtain the importance score of each landslide characteristic factor;

[0098] According to the importance scores of different landslide characteristic factors, the information value is weighted to obtain the enhanced information value;

[0099] Based on the calculation result of the enhanced information value, extremely low and low information areas are divided, and reliable non-landslide samples are selected from the extremely low and low information areas in the enhanced information model.

[0100] In this embodiment, the calculation of the enhanced information volume includes:

[0101] Obtain updated landslide catalog data through field investigation, landslide history catalog and remote sensing images; extract landslide susceptibility evaluation index data using digital elevation models, geological maps and other data;

[0102] Load the landslide characteristic factor layer and reclassify the factors;

[0103] Add landslide point data, extract multi-values ​​of landslide characteristic factor layer to landslide points, and calculate the number of landslide points within the specified range;

[0104] Perform statistical calculations on the amount of information, export the attribute table, reconnect the calculated data to the vector layer, and then convert it to a raster using the surface-to-raster tool to obtain the characteristic factor information layer.

[0105] Based on the feature screening results, the importance scores of different feature factors are considered, and the information values ​​of the feature factors are weighted to obtain the enhanced information value.

[0106] According to the enhanced information value, the area to be predicted is divided into areas of different levels.

[0107] Non-landslide samples were selected in low-information areas (e.g., very low-information areas and low-information areas) to avoid misselection of samples in high-susceptibility areas.

[0108] Furthermore, a suitable machine learning algorithm is selected to evaluate the importance of each feature factor. Commonly used algorithms include random forest, gradient boosting tree, support vector machine, etc. The random forest algorithm is widely used in feature importance assessment due to its strong maturity and high prediction accuracy.

[0109] The selected machine learning algorithm is used to train the landslide samples and calculate the importance score of each feature in the model. The feature importance index of the random forest model (such as Gini importance) is used to evaluate the influence of each feature on the occurrence of landslides.

[0110] Based on the importance score of each feature, its weight is calculated. For example, the inverse of the feature importance score can be used as the weight, or a linear transformation of the feature importance score can be used as the weight.

[0111] The enhanced information content model is used to partition the proneness of the prediction area.

[0112] The information volume model is a statistical model based on the principles of information theory. Its core idea is to calculate the probability of landslide occurrence under different environmental factors and divide the area into susceptibility groups according to the probability.

[0113] The enhanced information model takes into account the weights of different characteristic factors on the basis of the information model, making the susceptibility division results more reasonable and accurate.

[0114] According to the information value of each factor at different susceptibility levels, the area to be predicted is divided into extremely low, low, medium, high and extremely high susceptibility areas.

[0115] Non-landslide samples were randomly selected in the very low and low susceptibility areas. The number of samples was selected based on the number of landslide samples and the balance of the data set was considered.

[0116] In order to ensure the generalization ability of the model, the number of non-landslide samples should be equivalent to the number of landslide samples, and the sample distribution should have a certain degree of randomness.

[0117] Using the random sampling method, a number of samples equal to the number of landslide sites were randomly selected from the low and very low susceptibility classes as non-landslide samples.

[0118] Furthermore, the use of the random forest model to construct a landslide susceptibility prediction model, constructing a model training set and a model test set based on the reliable non-landslide samples, and inputting the landslide susceptibility prediction model to obtain a first prediction result, further includes:

[0119] Conduct a quality assessment of the representativeness, quantity, distribution and independence of the selected non-landslide samples to verify the credibility of the non-landslide samples;

[0120] The random forest model was selected to use the optimized non-landslide samples together with the landslide samples for model training, and the parameters of the landslide susceptibility prediction model were optimized using grid search and cross-validation methods;

[0121] The parameters of the landslide susceptibility prediction model include at least the number of decision trees, the maximum depth, and the minimum number of samples;

[0122] After training, the landslide susceptibility prediction model parameters are used to predict the landslide susceptibility in unknown areas.

[0123] In this embodiment, the quality assessment of the selected non-landslide samples includes:

[0124] Representativeness: Non-landslide points should be representative of areas where landslides have not occurred in the predicted area, and avoid selecting potential high-risk areas.

[0125] Quantity: The number of non-landslide points should be close to 1:1 with the number of landslide samples to ensure the adequacy of model training.

[0126] Distribution: Non-landslide points should be evenly distributed throughout the area to be predicted to avoid concentration in a specific area.

[0127] Independence: Non-landslide points should be independent of landslide points, and avoid selecting areas that are spatially adjacent to landslide points.

[0128] Specifically, the random forest model in this embodiment is an integrated learning algorithm, and its parameter selection is crucial to the prediction accuracy and generalization ability of the model. In this embodiment of the present invention, the parameter selection of the random forest model includes:

[0129] Number of trees (n_estimators): The more decision trees there are, the more complex the model is and the higher the prediction accuracy is, but too high a number of trees will lead to overfitting. In the present invention, 300 decision trees are selected as the optimal number of trees based on the cross-validation results.

[0130] Splitting threshold (max_depth): The greater the depth of the decision tree, the more complex patterns the model can learn, but it is also prone to overfitting. In the present invention, according to the cross-validation results, 15 layers are selected as the optimal splitting threshold.

[0131] At each split, the random forest model selects a certain number of features from all features for consideration. In the present invention, according to the cross-validation results, sqrt(m) features are selected as the maximum number of features, where m is the total number of features.

[0132] Different from the traditional random forest, the random forest model of the present invention does not use replacement sampling (bootstrap=False), so that each tree is constructed based on the complete data set, which can improve the stability of the model.

[0133] The landslide samples and non-landslide samples selected by the enhanced information method are used as training data sets, and the sampling data with attributes are divided into test data sets. The training data set is used to train the landslide model, and the test data set is used to verify the performance of the landslide model. 70% of the data is used as the training data set and the remaining 30% of the data is used as the test data set.

[0134] The trained random forest model is used to predict the area to be predicted and the landslide susceptibility index of each grid cell is obtained.

[0135] According to the landslide susceptibility index, the area to be predicted is divided into five levels: very low, low, medium, high and very high.

[0136] The susceptibility levels are plotted into a map to visually display the distribution of landslide susceptibility in the area to be predicted.

[0137] Furthermore, the method of using the ROC curve, confusion matrix and relative density method to evaluate the prediction accuracy and uncertainty of the first prediction result of the landslide susceptibility prediction model also includes:

[0138] The prediction accuracy and consistency of the landslide susceptibility prediction model were evaluated using ROC curve, confusion matrix and Kappa coefficient indicators;

[0139] The ROC curve shows the performance of the model, and the area under the curve (AUC) analyzes the accuracy of the prediction results. The higher the AUC value, the better the model performance;

[0140] Confusion matrix analysis of the classification results of the landslide susceptibility prediction model, including accuracy, precision, and recall;

[0141] The Kappa coefficient evaluates the consistency of the landslide susceptibility prediction model, and the higher the value, the better the model consistency;

[0142] The relative density method is used to analyze the susceptibility distribution characteristics predicted by the landslide susceptibility prediction model and compare them with the actual landslide distribution.

[0143] In this embodiment, indicators such as ROC (receiver operating characteristic curve), confusion matrix, Kappa coefficient, etc. are used to evaluate the prediction accuracy and uncertainty of the model.

[0144] Precision analysis based on ROC curve is a key means to evaluate the performance of binary classification models. It intuitively displays the performance of the model by plotting the relationship curve between the true positive rate (TPR) and the false positive rate (FPR) under different thresholds.

[0145] The drawing of ROC curve includes data preparation, threshold determination, calculation of TPR and FPR, and connecting the points with FPR as the horizontal axis and TPR as the vertical axis.

[0146] The performance of the model is quantified by the area under the ROC curve (AUC value). The larger the AUC value, the better the model classification effect. The ideal model should have an AUC value close to 1.

[0147] By comparing the ROC curves of different models, we can intuitively judge the performance and select the appropriate threshold according to the ROC curve in practical applications to balance the true positive rate and false positive rate, so as to make the best choice in terms of precision, recall and other indicators.

[0148] The confusion matrix is ​​another important tool for evaluating the performance of classification models. Accuracy and precision derived from the confusion matrix are two basic indicators in the performance evaluation of classification models. The Kappa coefficient is an indicator for measuring model consistency. The higher the value, the better the consistency between the model prediction results and the actual results. The meaning of each variable is:

[0149] True Positive (TP): The true class of the sample is the positive class, and the result of model recognition is also the positive class.

[0150] False Negative (FN): False negative class. The true category of the sample is the positive class, but the model identifies it as the negative class.

[0151] False Positive (FP): False positive. The true category of the sample is the negative class, but the model identifies it as the positive class.

[0152] True Negative (TN): True negative class. The true class of the sample is negative, and the model identifies it as negative. The evaluation index is calculated by formulas (1)-(3):

[0153]

[0154] In formula (3), the higher the Kappa coefficient, the better the consistency of the model. Generally, a Kappa coefficient above 0.8 indicates very good consistency, 0.61-0.8 indicates good consistency, 0.41-0.6 indicates moderate consistency, 0.21-0.4 indicates general consistency, and 0.0-0.2 indicates almost no consistency. Among them, P is the sum of the number of correctly classified samples in each category divided by the total number of samples, that is, the overall classification accuracy. PO is the expected consistency rate, that is, the consistency rate of the two test results due to chance:

[0155]

[0156] According to a second embodiment of the present invention, the present invention claims a non-landslide sample selection optimization system based on information quantity, comprising:

[0157] The acquisition module collects historical landslide data and environmental factor data in the area to be predicted, and uses the recursive feature elimination method to screen out environmental factors that have a great impact on the susceptibility of landslides as landslide characteristic factors;

[0158] The sample screening module randomly selects initial non-landslide samples in the area to be predicted, uses the random forest model to evaluate the importance of the landslide characteristic factors, assigns weights to the information model according to the importance evaluation scores, divides extremely low and low information areas based on the calculation results of the information value, and selects reliable non-landslide samples from the extremely low and low information areas in the enhanced information model;

[0159] A model testing module, using a random forest model to build a landslide susceptibility prediction model, building a model training set and a model testing set based on the historical landslide data and reliable non-landslide samples, and inputting the landslide susceptibility prediction model to obtain a first prediction result;

[0160] A model evaluation module, which uses ROC curve, confusion matrix and relative density method to evaluate the prediction accuracy and uncertainty of the first prediction result of the landslide susceptibility prediction model;

[0161] The landslide assessment module uses the landslide susceptibility prediction model to conduct landslide susceptibility assessment on the predicted area, generates a landslide susceptibility zoning map, and formulates targeted disaster prevention and mitigation measures.

[0162] Furthermore, the acquisition module further includes:

[0163] Collect data on topography, geology, climate, soil, and vegetation environmental factors in the area to be predicted;

[0164] A machine learning algorithm is used to analyze the impact of the environmental factor data on the occurrence of landslides, and environmental factors that have a great impact on the occurrence of landslides are screened out as landslide characteristic factors, which serve as the basis for subsequent information calculation.

[0165] Furthermore, the sample screening module further includes:

[0166] The information content of the screened landslide characteristic factors is calculated to obtain the information content value of each landslide characteristic factor at different levels;

[0167] The importance of the landslide characteristic factors is evaluated using a random forest model to obtain the importance score of each landslide characteristic factor;

[0168] According to the importance scores of different landslide characteristic factors, the information value is weighted to obtain the enhanced information value;

[0169] Based on the calculation result of the enhanced information value, extremely low and low information areas are divided, and reliable non-landslide samples are selected from the extremely low and low information areas in the enhanced information model.

[0170] Furthermore, the model testing module further includes:

[0171] Conduct a quality assessment of the representativeness, quantity, distribution and independence of the selected non-landslide samples to verify the credibility of the non-landslide samples;

[0172] The random forest model was selected to use the optimized non-landslide samples together with the landslide samples for model training, and the parameters of the landslide susceptibility prediction model were optimized using grid search and cross-validation methods;

[0173] The parameters of the landslide susceptibility prediction model include at least the number of decision trees, the maximum depth, and the minimum number of samples;

[0174] After training, the landslide susceptibility prediction model parameters are used to predict the landslide susceptibility in unknown areas.

[0175] Furthermore, the model evaluation module further includes:

[0176] The prediction accuracy and consistency of the landslide susceptibility prediction model were evaluated using ROC curve, confusion matrix and Kappa coefficient indicators;

[0177] The ROC curve shows the performance of the model, and the area under the curve (AUC) analyzes the accuracy of the prediction results. The higher the AUC value, the better the model performance;

[0178] Confusion matrix analysis of the classification results of the landslide susceptibility prediction model, including accuracy, precision, and recall;

[0179] The Kappa coefficient evaluates the consistency of the landslide susceptibility prediction model, and the higher the value, the better the model consistency;

[0180] The relative density method is used to analyze the susceptibility distribution characteristics predicted by the landslide susceptibility prediction model and compare them with the actual landslide distribution.

[0181] The following is a specific experimental example:

[0182] The distribution data of 1021 landslide points in Henan Province were obtained from the Geographic Remote Sensing Ecological Network, and combined with historical map image data and landslide lists to ensure the accuracy of landslide location data;

[0183] 20 basic environmental factor data were collected, including elevation, slope, aspect, curvature, rainfall, humidity, population density, vegetation index and other factor data. These data cover information on topography, geological structure, soil, vegetation, precipitation and other aspects, and can fully reflect the geological environment and potential risks of the area to be predicted;

[0184] Calculate the information amount of each factor and make a characteristic factor information amount diagram;

[0185] The factors with the highest impact on landslides are screened out through recursive feature elimination algorithm, and the information weight is allocated according to the importance score of the characteristic factors to obtain the enhanced information layer.

[0186] Three different methods, namely, the whole area random selection method, the landslide buffer method and the enhanced information method, were used to select non-landslide samples, and the landslide samples and non-landslide samples were used as the modeling data sets.

[0187] The sampling data with attributes are divided into a training data set and a test data set. The training data set is used to train the landslide model, and the test data set is used to verify the performance of the landslide model. 70% of the data is used as the training data set, and the remaining 30% of the data is used as the test data set.

[0188] Use the training data set to train the selected model and adjust the model parameters to improve the prediction accuracy of the model. Use methods such as grid search and cross-validation to evaluate the model performance under different model parameters and select the optimal model parameters;

[0189] The non-landslide samples selected by three different methods were coupled with the random forest model to construct three models: random-RF, buffer-RF, and EIV-RF for landslide susceptibility prediction. The landslide susceptibility model obtained through training can generate a landslide susceptibility zoning map.

[0190] The prediction accuracy and uncertainty of the model were evaluated by ROC curve, confusion matrix and relative density analysis.

[0191] The characteristic data in the data collection includes:

[0192] Various types of data including topographic conditions, geological elements, soil conditions, meteorology and hydrology, human activities, etc.

[0193] Feature indicators include elevation, slope, aspect, plan curvature, profile curvature, soil data set, precipitation, humidity, population density, NDVI, etc.

[0194] like Figure 2 As shown, this embodiment provides a schematic diagram of the structure of the random forest learner used:

[0195] The Random Forest model (RF) is an integrated learning algorithm that improves the accuracy and stability of predictions by building multiple decision trees. Each decision tree is trained based on a subset of samples and features, and then the results of all decision trees are combined for prediction. The following is a detailed explanation of each step in the Random Forest model structure diagram:

[0196] Training samples: These samples contain all feature variables, such as elevation, slope, rainfall, etc., as well as the corresponding label variables (if the model is a classification task) or target variables (if the model is a regression task);

[0197] Random selection of features and samples: When building each decision tree, the random forest model randomly selects a portion of features (random selection of features) and a portion of samples (random selection of samples) to grow the tree. This randomness makes each decision tree different, which helps improve the generalization ability of the model;

[0198] Training set: Each training set is composed of randomly selected features and samples and is used to train a decision tree. Each training set corresponds to a decision tree.

[0199] Decision tree: A decision tree is grown for each training set. A decision tree is a tree structure used to classify or regress samples. Each node contains a feature and a threshold value used to divide the sample into two subsets. The path on the tree represents the classification or regression result of the sample;

[0200] Prediction results: Each decision tree predicts the training sample and obtains different prediction results Y1, Y2, Y3...Y k For classification tasks, the prediction result is the probability that each sample belongs to a certain category; for regression tasks, the prediction result is the predicted value of each sample;

[0201] Final prediction: Finally, the prediction results of all decision trees are averaged or voted to get the final prediction result Y. For classification tasks, Y is the average of all decision tree prediction results; for regression tasks, Y is the voting result of all decision tree prediction results.

[0202] like Figure 3 As shown, the importance ranking of landslide characteristic factors based on the recursive feature elimination algorithm used in this embodiment is:

[0203] In this implementation, 20 environmental factors such as elevation, slope, aspect, and curvature were selected as initial landslide characteristic factors. Recursive feature elimination (RFE) was used for analysis. Landslide samples and non-landslide samples were used as modeling data sets. An initial feature set was first selected from the 20 feature factors. Then, the feature set was trained using a random forest model to evaluate the importance of each feature. The algorithm eliminated the least important features and updated the feature set.

[0204] This process is repeated until a preset stopping condition is reached, such as the number of features reaching a specified value or the model performance no longer improving.

[0205] By repeatedly training the model and recursively performing feature set regression on the 20 initial feature factors, we can identify important factors that play a key role in the model performance, eliminate factors with low impact or irrelevant factors, and screen out the best feature subset in the classification.

[0206] In this example, 12 landslide characteristic factors with higher importance scores (elevation, slope, aspect, plane curvature, profile curvature, precipitation, humidity, population density, vegetation index, t_oc (0-30cm soil organic carbon content), t_silt (0-30cm silt content), s_clay (30-100cm clay content)) were finally selected for landslide susceptibility modeling.

[0207] These characteristic factors cover information on topography, geological structure, soil, vegetation, precipitation, etc., and can comprehensively reflect the geological environment and potential risks of the area to be predicted.

[0208] Through importance analysis, the relative importance of each characteristic factor is obtained. The higher the value, the stronger the impact on the landslide model. Among them, the importance of elevation, slope and rainfall are 0.227, 0.120 and 0.119 respectively, which are the three factors with the highest scores among the landslide characteristic factors.

[0209] like Figure 4 In this embodiment, the information amount of each characteristic factor is calculated by respectively calculating the information amount of the selected characteristic factors to obtain the information amount layer of each characteristic factor:

[0210] To calculate the information content of each landslide characteristic factor, it is necessary to first collect data related to the landslide, perform data preprocessing, and then use the information content model to calculate the information content of each landslide characteristic factor. The main steps include:

[0211] Obtain updated landslide catalog data through field investigation, landslide history catalog and remote sensing images; extract landslide susceptibility evaluation index data using digital elevation models, geological maps and other data;

[0212] Load the landslide characteristic factor layer in ArcGIS and reclassify the factors;

[0213] Add landslide point data, extract multi-values ​​of landslide characteristic factor layer to landslide points, and calculate the number of landslide points within the specified range;

[0214] Perform statistical calculations on the amount of information, export the attribute table, reconnect the calculated data to the vector layer, and then convert it to raster using the polygon to raster tool;

[0215] like Figure 5 In this embodiment, information weights are allocated according to the feature factor importance scores to obtain an enhanced information map, which is described by the following steps:

[0216] Step 1: Based on the information volume model, this study combines machine learning with the information volume method and screens out factors with a higher impact on landslides through a recursive feature elimination algorithm.

[0217] Step 2: The information of characteristic factors such as elevation, slope, rainfall, curvature, and population density is weighted according to importance to obtain an enhanced information map.

[0218] Step 3: Enhanced information volume The model takes into account the different contributions of various characteristic factors to landslide occurrence, so it can more accurately assess the relative importance of various factors to landslide risk, thereby improving the accuracy of model prediction.

[0219] Step 4: Use the natural breakpoint method to initially partition the enhanced information content model and select non-landslide samples in the extremely low and low information content areas.

[0220] like Figure 6 This embodiment also provides a schematic diagram of the distribution positions of landslide samples and non-landslide samples in the enhanced information map:

[0221] Based on the enhanced information volume model, non-landslide samples were selected and the comprehensive information volume map was deeply analyzed to understand the distribution of information volume values ​​in different regions and their relationship with geological disaster risks.

[0222] Through this analysis, areas that are prone to geological disasters and those that are not prone to geological disaster risks are identified.

[0223] The natural break point method is used to initially partition the enhanced information content model into extremely low, low, medium, high, and extremely high information content areas, and non-landslide samples are selected in the extremely low and low information content areas.

[0224] A random sampling method was used to extract non-landslide samples. At the same time, the demand for data volume for machine learning was taken into account. The number of non-landslide points and landslide points was selected in a 1:1 ratio.

[0225] In this example, there are 1021 landslide points in total. 1021 non-landslide points are evenly selected in the extremely low and low information areas where the enhanced information value is lower than 0 and geological disasters are not likely to occur. The selected non-landslide points are evenly distributed in the entire study area.

[0226] Finally, through verification, it can be found that the landslide samples are basically distributed in the area with higher information value, that is, the extremely high and high information value areas initially divided, while the non-landslide samples are also distributed in the low information value area, which is consistent with the facts.

[0227] like Figure 7 ,This embodiment also provides ROC curve diagrams of different non-landslide sample selection ,models for comparing the prediction accuracy and effect of the ,models.

[0228] The ROC curve of the random-RF model is less convex, with an AUC value of 0.83. As can be seen from the figure, the performance of the random-RF model is relatively low at all false positive rates. This indicates that the random-RF model has a weak ability to distinguish between positive and negative samples. The AUC value of the buffer-RF model is 0.90. The buffer-RF model performs well at low false positive rates, but its performance decreases as the false positive rate increases. It can be seen that selecting non-landslide samples outside the buffer area of ​​the landslide point improves the accuracy of the random forest model; the ROC curve of the EIV-RF model is higher than that of the above two methods. The AUC value of EIV-RF is 0.96. As the false positive rate increases, the true positive rate also increases accordingly, indicating that the EIV-RF model has a strong ability to distinguish between positive and negative samples. This shows that the non-landslide samples selected in the extremely low and low information areas of the initial division based on the EIV model can significantly improve the model accuracy.

[0229] The ROC curves of the above three models show that the EIV-RF model has the best prediction effect and accuracy.

[0230] like Figure 8 , this embodiment also provides a landslide susceptibility zoning result map based on the EIV-RF model.

[0231] According to the landslide susceptibility prediction results, it can be learned that the extremely high and high susceptibility areas are mainly distributed in the Taihang Mountain area, the Sanmenxia-Luoyang Loess Hilly Mountain area, the Xiaoshan-Xiong'ershan-Funiushan Mountain area, and the Tongbai-Dabie Mountain area. The complex terrain and geological structure environment in these areas, coupled with frequent human engineering activities, have led to an increase in landslide risks. The undulating terrain and slope changes make the soil more prone to instability, increasing the risk of landslides. The medium-prone areas are mainly distributed in the southern foot of the Funiushan Mountains, the eastern foot of the Waifang Mountains, the eastern foot of the Songshan Mountains, and the northern foot of the Dabie Mountains. These areas have relatively steep terrain and large terrain changes, which is also one of the important factors for the susceptibility of landslides. The low and extremely low susceptibility areas are mainly distributed in the Yellow River alluvial plain area, the Huaihe River alluvial plain area, and the Nanyang Basin area. These areas have flat terrain and lack the basic environmental conditions for landslides to occur, so landslide disasters are relatively rare.

[0232] This embodiment also provides the results of different coupling model accuracy evaluation indicators:

[0233] The order of magnitude of accuracy, precision, and Kappa coefficient is Random - RF < Buffer - RF < EIV - RF. The results show that compared with Random - RF, the Buffer - RF and EIV - RF coupling models have higher prediction accuracy. Among them, the prediction accuracy of the Buffer - RF model is 83.65%, and the prediction accuracy of the EIV - RF model is as high as 92.95%, further proving that the EIV - RF model has higher reliability. At the same time, the Kappa coefficient of EIV - RF is 0.8591, indicating that the prediction results of the EIV - RF model have very good consistency. The Kappa coefficient of the Buffer - RF model is 0.6707, indicating that the prediction results of Buffer - RF have good consistency, while the Kappa coefficient of the Random - RF model is 0.4934, indicating that the prediction results of the Random - RF model have medium consistency. Through the above model evaluation indicators, it can be concluded that the EIV - RF model has the best prediction effect among the three models for landslide susceptibility prediction.

[0234] This embodiment also provides the statistical results of the susceptibility evaluation levels based on 3 non - landslide sample selection methods:

[0235] In landslide susceptibility prediction, the distribution of landslide points in different ranges of each characteristic factor is different. The relative density can describe the influence degree of each factor on landslide disasters in different range intervals. The relative density refers to the ratio of the proportion of landslides occurring at different grades in a certain area to the total area occupied by landslides of that grade. This concept is often used in geological disaster research to evaluate the occurrence frequency and influence degree of landslides at different grades. Formula 4 is as follows:

[0236]

[0237] In the formula, D is the relative density of landslide points, Ni is the number of landslide points in the i - th level area of the factor, N is the total number of landslide points in the study area, Si is the area of the i - th level area of the factor, and S is the total area of the study area.

[0238] As the landslide susceptibility level increases, the corresponding relative density D value also gradually increases. Most of the landslide grid units are included in the extremely high and high - susceptibility areas predicted by each method. Thus, it can be seen that the susceptibility maps predicted by the 3 selection methods are generally similar, but differences still exist when comparing the details of the prediction results. When using an equal - proportion sample set (landslide: non - landslide = 1:1), the proportions of historical landslides in the extremely high and high - susceptibility areas of the Random - RF, Buffer - RF, and EIV - RF models are 72.18%, 74.83%, and 87.37% respectively. Among them, the EIV - RF model predicts the largest number of historical landslides in the extremely high and high - susceptibility areas, indicating that the susceptibility results of the EIV - RF model are more consistent with the distribution characteristics of known landslides and have better modeling performance.

[0239] The beneficial effects of the present invention are as follows:

[0240] The non-landslide sample optimization strategy based on enhanced information method proposed in this paper can effectively improve the accuracy of landslide susceptibility prediction model and reduce uncertainty, which is mainly reflected in the following aspects:

[0241] Traditional non-landslide sample selection methods (such as random selection and buffer zone selection) may mistakenly select potential high-risk areas as non-landslide samples, resulting in erroneous information in the model training data and reducing the model prediction accuracy. The EVI method can identify "safe" areas with lower landslide risks by analyzing the weights of each characteristic factor and combining the information volume model, thereby avoiding selecting potential high-risk areas as non-landslide samples.

[0242] The EVI method not only considers areas with a low probability of landslides, but also considers the importance of each characteristic factor to ensure that the selected non-landslide samples can better represent the overall characteristics of the area to be predicted. For example, the EVI method will give priority to samples in areas with gentle slopes, high vegetation coverage, and low rainfall. These areas account for a larger proportion of the area to be predicted and can better reflect the overall characteristics of the region.

[0243] The non-landslide samples selected by the EVI method are more representative, and the constructed model can more accurately identify potential high-risk areas. The model can more effectively learn the laws of landslide occurrence and thus more accurately predict the possibility of future landslides.

[0244] The EVI method can reduce the model's sensitivity to outliers and noise, making the model's prediction results more stable and reliable. Even in the absence of data or in the presence of noise, the model can maintain a high level of prediction accuracy, providing a more reliable basis for decision-making in disaster prevention and mitigation.

[0245] The EVI method can select more representative non-landslide samples, so that the model can better learn the laws of landslide occurrence and improve the generalization ability of the model. The model can better adapt to the characteristics of different regions and maintain a high prediction accuracy even when making predictions in new regions.

[0246] The non-landslide sample optimization strategy based on the enhanced information method can effectively improve the accuracy of the landslide susceptibility prediction model and reduce uncertainty, providing more reliable technical support for disaster prevention and mitigation. This strategy can select more representative non-landslide samples and build a more accurate prediction model, thereby more effectively identifying potential high-prone areas and reducing the risk of landslide disasters.

[0247] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, systems and methods can be implemented in other ways. For example, the system embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of systems or units, which can be electrical, mechanical or other forms.

[0248] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above integrated unit may be implemented in the form of hardware or in the form of software functional units. The above is only an implementation mode of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly used in other related technical fields, is also included in the patent protection scope of the present invention.

[0249] The specific implementation methods of the invention are described in detail above, but they are only examples, and the invention is not limited to the specific implementation methods described above. For those skilled in the art, any equivalent modification or substitution of the invention is also within the scope of the invention, therefore, the equalization, modification, improvement, etc. made without departing from the spirit and principle of the invention should be included in the scope of the invention.

Claims

1. A non-landslide sample selection optimization method based on information quantity, characterized in that: include: Collect historical landslide data and environmental factor data in the area to be predicted, and use the recursive feature elimination method to screen out environmental factors that have a great impact on landslide susceptibility as landslide characteristic factors; Initial non-landslide samples are randomly selected in the area to be predicted, and the importance of all landslide characteristic factors is evaluated using the random forest model. The weight of the information model is assigned according to the importance evaluation score of the landslide characteristic factors, and reliable non-landslide samples are selected in the area with low information in the obtained information model. Using a random forest model to construct a landslide susceptibility prediction model, constructing a model training set and a model test set based on the historical landslide data and reliable non-landslide samples, and inputting the landslide susceptibility prediction model to obtain a first prediction result; The prediction accuracy and uncertainty of the first prediction result of the landslide susceptibility prediction model are evaluated by using ROC curve, confusion matrix and relative density method; The landslide susceptibility prediction model is used to evaluate the landslide susceptibility of the predicted area, generate a landslide susceptibility zoning map, and formulate targeted disaster prevention and mitigation measures.

2. The information-based non-landslide sample selection optimization method according to claim 1 is characterized in that: The collecting of historical landslide data and environmental factor data in the area to be predicted, and using a recursive feature elimination method to screen out environmental factors that have a great impact on the susceptibility of landslides as landslide characteristic factors, also includes: Collect data on topography, geology, climate, soil, and vegetation environmental factors in the area to be predicted; A machine learning algorithm is used to analyze the impact of the environmental factor data on the occurrence of landslides, and factors that have a great impact on the occurrence of landslides are screened out as landslide characteristic factors, which serve as the basis for subsequent information calculations.

3. The method for selecting non-landslide samples based on information quantity according to claim 1 is characterized in that: Randomly select initial non-landslide samples in the area to be predicted, use the random forest model to evaluate the importance of the landslide characteristic factors, assign weights to the information model according to the importance evaluation scores of the landslide characteristic factors, and select reliable non-landslide samples from the area with low information content in the obtained information model, and also include: The information content of the screened landslide characteristic factors is calculated to obtain the information content value of each landslide characteristic factor at different levels; The importance of the landslide characteristic factors is evaluated using a random forest model to obtain the importance score of each landslide characteristic factor; According to the importance scores of different landslide characteristic factors, the information value is weighted to obtain the enhanced information value; Based on the calculation result of the enhanced information value, extremely low and low information areas are divided, and reliable non-landslide samples are selected from the extremely low and low information areas in the enhanced information model.

4. The method for selecting non-landslide samples based on information quantity according to claim 1 is characterized in that: The method of using the random forest model to construct a landslide susceptibility prediction model, constructing a model training set and a model test set based on the historical landslide data and reliable non-landslide samples, and inputting the landslide susceptibility prediction model to obtain a first prediction result, further includes: Conduct a quality assessment of the representativeness, quantity, distribution and independence of the selected non-landslide samples to verify the credibility of the non-landslide samples; The random forest model was selected to use the optimized non-landslide samples together with the landslide samples for model training, and the parameters of the landslide susceptibility prediction model were optimized using grid search and cross-validation methods; The parameters of the landslide susceptibility prediction model include at least the number of decision trees, the maximum depth, and the minimum number of samples; After training, the landslide susceptibility prediction model parameters are used to predict the landslide susceptibility of unknown areas.

5. The method for selecting non-landslide samples based on information quantity according to claim 1 is characterized in that: The method of using ROC curve, confusion matrix and relative density method to evaluate the prediction accuracy and uncertainty of the first prediction result of the landslide susceptibility prediction model also includes: The prediction accuracy and consistency of the landslide susceptibility prediction model were evaluated using ROC curve, confusion matrix and Kappa coefficient indicators; The ROC curve shows the performance of the model, and the area under the curve AUC analyzes the accuracy of the prediction results. The higher the AUC value, the better the model performance; Confusion matrix analysis of the classification results of the landslide susceptibility prediction model, including accuracy, precision, and recall; The Kappa coefficient evaluates the consistency of the landslide susceptibility prediction model. The higher the Kappa value, the better the model consistency. The relative density method is used to analyze the susceptibility distribution characteristics predicted by the landslide susceptibility prediction model and compare them with the actual landslide distribution.

6. A non-landslide sample selection optimization system based on information quantity, characterized in that: include: The acquisition module collects historical landslide data and environmental factor data in the area to be predicted, and uses the recursive feature elimination method to screen out environmental factors that have a great impact on the susceptibility of landslides as landslide characteristic factors; The sample screening module randomly selects initial non-landslide samples in the area to be predicted, uses the random forest model to evaluate the importance of the landslide characteristic factors, assigns weights to the information model according to the importance evaluation scores, and selects reliable non-landslide samples from the area with low information content in the obtained information model; A model testing module, using a random forest model to construct a landslide susceptibility prediction model, constructing a model training set and a model testing set based on the historical landslide data and reliable non-landslide samples, and inputting the landslide susceptibility prediction model to obtain a first prediction result; A model evaluation module, which uses ROC curve, confusion matrix and relative density method to evaluate the prediction accuracy and uncertainty of the first prediction result of the landslide susceptibility prediction model; The landslide assessment module uses the landslide susceptibility prediction model to conduct landslide susceptibility assessment on the predicted area, generates a landslide susceptibility zoning map, and formulates targeted disaster prevention and mitigation measures.

7. The information-based non-landslide sample selection optimization system according to claim 6 is characterized in that: The acquisition module further includes: Collect data on topography, geology, climate, soil, and vegetation environmental factors in the area to be predicted; A machine learning algorithm is used to analyze the impact of the environmental factor data on the occurrence of landslides, and environmental factors that have a great impact on the occurrence of landslides are screened out as landslide characteristic factors, which serve as the basis for subsequent information calculation.

8. The information-based non-landslide sample selection optimization system according to claim 6 is characterized in that: The sample screening module further includes: The information content of the screened landslide characteristic factors is calculated to obtain the information content value of each landslide characteristic factor at different levels; The importance of the landslide characteristic factors is evaluated using a random forest model to obtain the importance score of each landslide characteristic factor; According to the importance scores of different landslide characteristic factors, the information value is weighted to obtain the enhanced information value; Based on the calculation result of the enhanced information value, extremely low and low information areas are divided, and reliable non-landslide samples are selected from the extremely low and low information areas in the enhanced information model.

9. The information-based non-landslide sample selection optimization system according to claim 6, characterized in that: The model testing module further includes: Conduct a quality assessment of the representativeness, quantity, distribution and independence of the selected non-landslide samples to verify the credibility of the non-landslide samples; The random forest model was selected to use the optimized non-landslide samples together with the landslide samples for model training, and the parameters of the landslide susceptibility prediction model were optimized using grid search and cross-validation methods; The parameters of the landslide susceptibility prediction model include at least the number of decision trees, the maximum depth, and the minimum number of samples; After training, the landslide susceptibility prediction model parameters are used to predict the landslide susceptibility of unknown areas.

10. The information-based non-landslide sample selection optimization system according to claim 6, characterized in that: The model evaluation module further includes: The prediction accuracy and consistency of the landslide susceptibility prediction model were evaluated using ROC curve, confusion matrix and Kappa coefficient indicators; The ROC curve shows the performance of the model, and the area under the curve (AUC) analyzes the accuracy of the prediction results. The higher the AUC value, the better the model performance; Confusion matrix analysis of the classification results of the landslide susceptibility prediction model, including accuracy, precision, and recall; The Kappa coefficient evaluates the consistency of the landslide susceptibility prediction model, and the higher the value, the better the model consistency; The relative density method is used to analyze the susceptibility distribution characteristics predicted by the landslide susceptibility prediction model and compare them with the actual landslide distribution.

Images