Hierarchical avalanche prediction method and system based on support vector machine

Abstract

The invention discloses a hierarchical avalanche prediction method and system based on a support vector machine, and belongs to the technical field of avalanche disaster prediction, and the method comprises the following steps: (1) obtaining an avalanche hidden danger region through historical disaster records, literature literature and field investigation in combination with a multi-source remote sensing image; collecting and unifying terrain, weather, accumulated snow and water system data, and constructing a sample feature space; (2) taking the avalanche hidden danger point as a positive sample to generate a positive sample set; generating a negative sample set according to a negative sample construction strategy; (3) inputting the positive samples, the negative samples and the environmental factors into a support vector machine model for training to obtain an avalanche susceptibility prediction model; (4) evaluating model precision by adopting cross validation and an independent test set, and optimizing model parameters; and (5) grading prediction results, and outputting an avalanche susceptibility partition map. According to the method, the accuracy and robustness of sample sparse region avalanche prediction can be improved, and reliable support is provided for avalanche disaster early warning and risk management in a high mountain complex environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of avalanche disaster prediction, and particularly relates to a method and system for grading prediction of avalanches based on a support vector machine. BACKGROUND

[0002] The southeast region of Tibet is located in the southeast edge of the Qinghai-Tibet Plateau, with complex topography and landform, and overall presents the characteristics of high mountains and deep valleys, and significant terrain undulations. The region has valleys and gorges, with sharp terrain differences, forming a typical high mountain and gorge environment. At the same time, due to large altitude span and diverse climate types, meteorological elements such as precipitation, air temperature and wind field are highly uneven in spatial and temporal distribution. These natural conditions together lead to the situation of frequent and frequent avalanches in the region, becoming one of the regions with the most concentrated and active avalanche activities.

[0003] Avalanches not only directly threaten the safety of residents' lives and property, but also cause damage to transportation arteries, energy facilities, water conservancy projects and communication lines, and further have a serious impact on regional economic development and social stability. In recent years, with the continuous extension of major infrastructure construction such as transportation, energy and water conservancy to high mountain and gorge areas, the interaction between human activities and ice and snow environment intensifies, and the potential risk of avalanche disasters continues to rise.

[0004] The Qinghai-Tibet Plateau is one of the most sensitive regions to global climate change, and the type, frequency and scale of avalanche activities have shown new patterns of change: the frequency of dry snow avalanches gradually decreases, while the proportion and activity level of wet snow avalanches continuously rise, and the peak period of avalanches shows an advance trend. This complex change not only reshapes the spatiotemporal distribution of avalanches, but also significantly increases the uncertainty of risk prediction and management.

[0005] The disaster-pregnant mechanism of avalanche involves the coupling of multiple factors, including topographic factors (such as slope, aspect, and topographic curvature), meteorological factors (such as snowfall, air temperature, and wind speed), and snow cover factors (such as snow cover and snow depth). The relationship between these factors is complex and highly nonlinear. Traditional methods that rely on experience or statistics have obvious limitations in handling complex multi-source data, making it difficult to meet the needs of fine-grained risk assessment in complex mountainous areas. In recent years, with the development of machine learning, methods such as support vector machines (SVM) and random forests (RF) have been introduced into avalanche prediction research and have achieved good results in some areas. However, their application in southeastern Tibet is still relatively limited. The main reasons for the above limitations are: ① Avalanche samples are extremely sparse. Due to the complex environment of high mountains and valleys, it is difficult to observe and monitor avalanches, and the number of positive samples available for modeling is limited; ② The distribution of positive and negative samples is severely unbalanced. In the case of limited positive samples, the selection of negative samples plays a decisive role in the performance of the model; ③ The existing negative sample construction methods are single, mostly randomly or empirically selected, and lack systematic comparison and optimization, making it difficult to ensure the stability of the model and the reliability of the zoning results; ④ The terrain in southeastern Tibet is complex, and there are many environmental factors, and there is a lack of targeted and applicable negative sample construction methods.

[0006] Therefore, the existing technology has the following deficiencies in the field of avalanche prediction: it cannot solve the modeling needs under the condition of sample sparsity; it is difficult to balance the rationality of negative sample selection and the robustness of the model; and it lacks a universal method applicable to complex mountainous areas.

[0007] In summary, in typical high mountain areas where avalanche samples are sparse and the distribution of positive and negative samples is severely unbalanced, a systematic negative sample construction strategy is needed, combined with advanced machine learning methods, to improve the accuracy and stability of avalanche classification prediction, thereby providing effective support for disaster risk identification and prevention. SUMMARY

[0008] The purpose of the present application is to address the problem of sample sparsity and severe imbalance between positive and negative samples in high mountain and valley areas, and to propose a method and system for hierarchical prediction of avalanches based on support vector machines. By designing a reasonable negative sample construction strategy and combining multiple sources of topographic, meteorological, snow cover, and water system factors, a fine-grained prediction of avalanche susceptibility is achieved, improving the stability and robustness of the model, and providing technical support for avalanche disaster risk assessment and early warning in complex mountainous areas.

[0009] The technical solution adopted by the present application is as follows:

[0010] A method for hierarchical prediction of avalanches based on support vector machines, comprising the following steps:

[0011] (1) Obtain snow avalanche hazard area by combining historical disaster records, literature data, and on-site investigation with multi-source remote sensing images; collect and unify terrain, weather, snow cover, and water system data, extract environmental factors, and construct sample feature space;

[0012] (2) Obtain snow avalanche hazard points from historical points, on-site investigation points, and image interpretation results as positive samples to generate a positive sample set; generate a negative sample set according to a preset negative sample construction strategy;

[0013] (3) Input the positive samples, negative samples, and corresponding environmental factors into a support vector machine model for training to obtain a snow avalanche susceptibility prediction model;

[0014] (4) Use cross-validation and an independent test set to evaluate model accuracy, and optimize model parameters to improve robustness;

[0015] (5) Classify the prediction results and output a snow avalanche susceptibility zoning map for disaster warning and risk management.

[0016] Further, in step (2), the negative sample construction strategy includes at least one of the following:

[0017] a. Random negative sample selection: randomly sampling in areas where no snow avalanches have occurred;

[0018] b. Negative sample selection based on physical constraints: selecting negative samples in areas with a slope less than 35° and a snow cover degree less than 0.3;

[0019] c. Negative sample selection based on low susceptibility areas: selecting negative samples in areas with a very low probability of snow avalanche occurrence confirmed by historical data and on-site investigation.

[0020] Further, in step (2), a spatial offset strategy is introduced in the positive sample construction process to offset the original positive sample points by a preset distance to form a positive sample set, which simulates the uncertainty of snow avalanche identification and improves the robustness of the model.

[0021] Further, in step (3), the support vector machine model uses a radial basis function kernel function.

[0022] Further, in step (5), the prediction results are classified into five levels according to probability values: very low, low, medium, high, and very high.

[0023] Further, in step (1), the environmental factors extracted from terrain, weather, snow cover, and water system data are:

[0024] Terrain factors: elevation, slope, curvature, terrain relief, and elevation coefficient of variation, obtained from DEM data through GIS calculation;

[0025] Meteorological factors: wind speed, air temperature and precipitation data, obtained through multi-source meteorological data sets;

[0026] Snow cover factors: snow cover and snow depth data, combined with precipitation and temperature data to extract snowfall information;

[0027] Water system factors: three-level river data sets.

[0028] A hierarchical prediction system for snow avalanches, comprising the following modules:

[0029] Data processing module: for the collection, standardization and spatial interpolation of multi-source data;

[0030] Negative sample construction module: for performing random, physically constrained and low-occurrence-area negative sample generation strategies;

[0031] Model training module: for sample training and parameter optimization based on a support vector machine model;

[0032] Prediction and visualization module: for outputting snow avalanche occurrence probability classification results and generating regional risk maps.

[0033] In summary, due to the adoption of the above technical solutions, the present application has the following advantages:

[0034] 1. By introducing a negative sample construction strategy, the training ability of the model under limited data conditions is enhanced, effectively alleviating the modeling difficulties caused by insufficient snow avalanche observations and solving the sample sparsity problem.

[0035] 2. Based on reasonable negative sample construction (especially the low-occurrence-area selection strategy), the AUC, Recall and F1 scores of the support vector machine model are significantly improved, avoiding the noise interference caused by random negative samples, making the prediction results more stable and reliable, and improving the prediction accuracy and robustness.

[0036] 3. When there is a spatial bias in the positive samples, the negative sample construction strategy proposed by the present application can still alleviate the decline in model performance, showing strong robustness and enhancing the adaptability to uncertainty.

[0037] 4. Different negative sample strategies not only affect the model accuracy, but also change the model's sensitivity to environmental factors. The present application can highlight the role of key factors such as temperature and snow cover, improving the interpretability and scientificity of the model.

[0038] 5. The method of the present application is suitable for high mountain areas with sparse data and complex environment, and can provide snow avalanche risk classification prediction and disaster warning support for transportation, energy, water conservancy and other infrastructure construction areas, having wide application value. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be considered as limiting the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings, in which:

[0040] Figure 1 The avalanche sample distribution diagram of the present application research area is shown in the figure.

[0041] Figure 2 The flow framework diagram of the present application is shown in the figure.

[0042] Figure 3 The experimental result diagram of the avalanche prediction method in the specific embodiment of the present application is shown in the figure.

[0043] Figure 4 The support vector machine model factor importance comparison diagram under different negative sample construction strategies in the specific embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0045] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0046] It should be noted that: the numbers and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0047] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship commonly placed when the product of the application is used, which is only a simplified description for the convenience of describing the present application, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" and the like are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0048] In addition, the terms "horizontal", "vertical" and the like do not mean that the components must be absolutely horizontal or vertical, but can be slightly inclined. For example, "horizontal" only means that it is more horizontal relative to "vertical", and does not mean that the structure must be completely horizontal, but can be slightly inclined.

[0049] In the description of the present application, it should be noted that, unless otherwise explicitly specified and limited, the terms "arrangement", "installation", "connection", "connection" should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0050] The accompanying drawings are incorporated in the description of the present application Figures 1-4 ,

[0051] A method for grading prediction of snow avalanche based on support vector machine, comprising the following steps:

[0052] (1) Data collection and arrangement: obtain snow avalanche hidden danger area through historical disaster records, literature data and field investigation combined with multi-source remote sensing images; collect and unify multi-source data such as terrain, weather, snow and water system, and extract environmental factors; interpolate different data, unify to 30m spatial resolution, and construct sample feature space.

[0053] Among them, the historical snow avalanche records and related literature are used to sort out the distribution of snow avalanche activities in the study area; the key areas are investigated on site to obtain snow avalanche hidden danger point information; based on multi-source remote sensing images, the potential snow avalanche danger area is identified; so as to obtain the snow avalanche hidden danger point.

[0054] Among them, in order to establish the sample feature space, the following environmental factors are extracted by the present embodiment:

[0055] Terrain factors: elevation, slope, curvature, terrain relief, and coefficient of variation of elevation, which are calculated from DEM data through GIS;

[0056] Weather factors: wind speed, air temperature, and precipitation data, which are obtained from multi-source weather datasets;

[0057] Snow factors: snow cover and snow depth data, which are extracted by combining precipitation and temperature data;

[0058] Water system factors: national 1:250000 third-grade river dataset;

[0059] Specifically, all factors are unified to 30m resolution through bilinear interpolation for model input.

[0060] (2) Positive sample construction: Positive samples are derived from historical disaster records, literature data, and snow avalanche hazard points obtained from multi-source remote sensing images. The spatial bias strategy is introduced to move the original sample points 1000 meters to the surrounding area to form a new positive sample set, which simulates the influence of spatial uncertainty on the model.

[0061] Negative sample construction: Three types of negative sample selection strategies are designed to improve the training set:

[0062] Random selection method: Randomly select negative samples in non-avalanche areas as the baseline strategy;

[0063] Physical condition constraint method: According to the terrain and snow conditions, select negative samples from low-risk areas with slope less than 35° and snow cover less than 0.3;

[0064] Low-hazard area selection method: Based on historical data and field investigation results, extract negative samples from areas with extremely low snow avalanche probability.

[0065] (3) Model training:

[0066] Input positive samples, negative samples, and corresponding environmental factors into the support vector machine model for training to obtain the snow avalanche susceptibility prediction model;

[0067] Among them:

[0068] Support vector machine (SVM) classification model is used, and radial basis function (RBF) kernel function is used to enhance the fitting ability of non-linear features;

[0069] All features are standardized to ensure consistent numerical scale;

[0070] Confusion matrix, AUC, precision, recall, F1 score, and Kappa coefficient are used to evaluate the performance of the model;

[0071] Permutation Importance is introduced to reveal the contribution of environmental factors to the model prediction.

[0072] (4) Hierarchical prediction:

[0073] According to the avalanche occurrence probability output by the model, the study area is divided into different risk levels to realize hierarchical prediction.

[0074] Six groups of comparative experiments are set up:

[0075] S1: original positive samples + random negative samples

[0076] S2: original positive samples + physical condition negative samples

[0077] S3: original positive samples + low-prone area negative samples

[0078] C1: offset positive samples + random negative samples

[0079] C2: offset positive samples + physical condition negative samples

[0080] C3: offset positive samples + low-prone area negative samples

[0081] The system compares the performance of the model under different strategies to evaluate the applicability of the negative sample construction method in avalanche prediction.

[0082] Specifically, as shown in Figure 3 , the prediction performance comparison of each experimental group (S1-S3, C1-C3) under different negative sample selection strategies and positive sample space offset conditions is shown. Among them, S1 group uses random negative samples, S2 group uses negative samples based on slope and snow cover conditions, and S3 group uses low-prone area negative samples; C1-C3 correspond to three types of negative sample strategies under the condition of positive sample position offset. In the figure, the AUC index is compared, which directly reflects the influence of different sample construction methods on the prediction performance and robustness of the model.

[0083] Specifically, the results show that the S1 group adopts random negative samples, and the AUC is 0.951, and the overall performance is at the benchmark level. In contrast, the S2 group filters negative samples by slope and snow cover conditions, and the AUC is improved to 0.962. The S3 group selects negative samples in low-prone areas, and the AUC (0.955) is also higher than S1. In summary, compared with random negative samples, the use of certain negative sample selection strategies can significantly improve the performance of the model in precision and robustness, and the low-prone area selection method is more helpful to improve the sensitivity of the model to snow avalanche occurrence while maintaining high AUC. In the context of introducing spatial offset of positive samples (C1-C3), the performance of the model generally decreases to varying degrees, indicating that spatial uncertainty has a negative impact on the stability and precision of the model. However, by comparing the three groups of offset sample models (C1-C3) with the corresponding original sample models (S1-S3), it can be seen that the construction method of negative samples to some extent alleviates the performance degradation caused by the offset of positive samples, and even achieves performance rebound in some indicators. The AUC of C1 group (random negative samples) is 0.950, which is slightly lower than 0.951 of S1, indicating that the uncertainty of positive samples can directly weaken the recognition ability of the model. C2 group (slope < 35° and snow cover < 0.3) shows further decline, with AUC of 0.948, indicating that under the condition of positive sample offset, the simple filtering strategy of negative samples cannot offset the precision loss. However, the AUC of C3 group (low-prone area negative samples) reaches 0.962, not only effectively alleviating the performance degradation caused by the offset of positive samples, but also exceeding the overall level of S1 random negative samples. Therefore, the uncertainty of positive samples can lead to a decrease in model precision, but through reasonable negative sample strategies, the robustness of the model can still be improved.

[0084] Specifically, as shown in Figure 4 , (the meaning of the figure is explained: SF: snowfall; WS: wind speed; TMP: air temperature; SD: snow depth; SC: snow cover; UND: terrain relief; ELE: elevation; EVAR: elevation variation coefficient; SLP: slope; CUR: curvature; RIV: distance from river) which shows the sensitivity differences of the six groups of models to the main environmental factors under different experimental conditions, including:

[0085] S1, S2, and S3 correspond to the experimental groups of random negative samples, negative samples selected based on slope and snow cover conditions, and low-prone area negative samples under original positive sample conditions, respectively;

[0086] C1, C2, and C3 correspond to the three types of negative sample strategies under the condition of spatial offset of positive samples.

[0087] The figure takes air temperature, snow cover, and other environmental factors as examples to show the importance of the factors calculated by each group of models, and compares the changes in factor importance under different negative sample strategies. This figure can illustrate the influence of different negative sample selection methods on factor sensitivity in the model identification mechanism, and the role of negative sample construction in model stability and factor discrimination structure under the condition of spatial uncertainty of positive samples.

[0088] Specifically, the results show that different negative sample strategies can significantly change the model's sensitivity to key factors, thereby affecting overall accuracy. For example, air temperature is the most important in S2(0.233) and C2(0.207) selected by physical conditions, but it is significantly reduced in S1(0.122) and C1(0.077) with random negative samples. This indicates that when negative samples are too random, the differences between samples are diluted by noise, making it difficult for the model to highlight the discriminative power of air temperature. However, negative samples constructed by physical constraints can strengthen the dominant role of air temperature in avalanche occurrence. In contrast, snow cover reaches its highest value in S3(0.236), but almost loses its effect in C1(0.067) and C2(0.022), indicating that the original positive samples under zoning constraints can highlight the contribution of snow cover in spatial differentiation, while this feature is easily masked when positive samples have spatial uncertainty. That is, different negative sample selection strategies form differences in the model identification path, which explains why S2 has a clear advantage in Precision, while S3 performs better in Recall in the previous results. Under the condition of greater spatial uncertainty of positive samples (C group), reasonable negative sample construction not only alleviates the decline in accuracy, but also stabilizes the factor discrimination structure of the model. For example, after selecting negative samples in the low susceptibility zone, the role of snow cover (C1: 0.067; C2: 0.022; C3: 0.196) is restored, thereby improving the robustness of the overall model. Therefore, negative samples not only provide "control" information, but also determine the model's identification mechanism for environmental factors. Random negative samples may introduce noise, leading the model to learn ambiguous or even incorrect discrimination patterns; while negative samples based on physical constraints or susceptibility zoning can enhance the signal of key factors, reduce the interference of irrelevant or weakly related samples, and improve model accuracy and interpretability. This result provides methodological inspiration for snow avalanche disaster research in sparse sample areas: under the condition of limited or uncertain samples, negative sample construction not only affects model performance, but also determines its identification mechanism and dominant factor interpretation.

[0089] In summary, the negative sample strategy constructed by the method of the present application can effectively improve the prediction accuracy and robustness of the model in the sample sparse area. At the same time, the model can identify the environmental factors that have been overlooked in the past but actually play a key role in the mechanism of snow avalanche, providing reliable technical support for snow avalanche disaster warning. Through the positive sample offset experiment verification, the method still shows stability and generalizability under the condition of uncertainty in the positive sample space.

[0090] Specifically, a hierarchical prediction system for snow avalanches based on the method of the present application is provided, comprising the following modules:

[0091] Data processing module: for collecting, standardizing and spatially interpolating multi-source data;

[0092] Negative sample construction module: for executing random, physically constrained and low-occurrence-area negative sample generation strategies;

[0093] Model training module: for sample training and parameter optimization based on a support vector machine model;

[0094] Prediction and visualization module: for outputting the hierarchical results of snow avalanche occurrence probability and generating a regional risk map.

[0095] As described above is an embodiment of the present application. The foregoing is each preferred embodiment of the present application, and the preferred embodiments in each preferred embodiment can be arbitrarily stacked and combined for use if not obviously self-contradictory or with a certain preferred embodiment as a prerequisite. The embodiments and specific parameters in the embodiments are only for clearly describing the verification process of the application and are not intended to limit the patent protection scope of the present application, and the patent protection scope of the present application is still subject to its claims. Any equivalent structural changes made by using the content of the specification and drawings of the present application should also be included in the protection scope of the present application.

Claims

1. A hierarchical avalanche prediction method based on support vector machines, characterized in that, Includes the following steps: (1) Avalanche hazard areas were obtained by combining historical disaster records, literature and field surveys with multi-source remote sensing images; topographic, meteorological, snow cover and water system data were collected and unified, environmental factors were extracted and sample feature space was constructed; (2) Use avalanche history points, on-site investigation points and avalanche hazard points obtained from image interpretation results as positive samples to generate a positive sample set; generate a negative sample set according to the preset negative sample construction strategy; (3) Input positive samples, negative samples and corresponding environmental factors into the support vector machine model for training to obtain the avalanche susceptibility prediction model. (4) Use cross-validation and independent test sets to evaluate the model accuracy and optimize the model parameters to improve robustness; (5) The prediction results are classified and an avalanche susceptibility zoning map is output for disaster early warning and risk management.

2. The hierarchical avalanche prediction method based on support vector machines according to claim 1, characterized in that, In step (2), the negative sample construction strategy includes at least one of the following: a. Random negative sample selection: Random sampling is conducted within the study area where no avalanches have occurred; b. Negative sample selection based on physical constraints: Negative samples are selected in areas with a slope of less than 35° and a snow cover of less than 0.

3. c. Negative sample selection based on low-risk areas: Negative samples are selected from areas where the probability of avalanche occurrence is confirmed to be extremely low based on historical data and on-site investigations.

3. The hierarchical avalanche prediction method based on support vector machines according to claim 1, characterized in that, In step (2), a spatial offset strategy is introduced during the positive sample construction process to offset the original positive sample points by a preset distance to form a positive sample set, which is used to simulate the uncertainty of avalanche identification and improve the robustness of the model.

4. The hierarchical avalanche prediction method based on support vector machines according to claim 1, characterized in that, In step (3), the support vector machine model uses a radial basis function kernel function.

5. The hierarchical avalanche prediction method based on support vector machines according to claim 1, characterized in that, In step (5), the prediction results are divided into five levels according to the probability value: extremely low, low, medium, high, and extremely high.

6. The hierarchical avalanche prediction method based on support vector machines according to claim 1, characterized in that, In step (1), the environmental factors extracted from topographic, meteorological, snow cover, and water system data are specifically as follows: Topographic factors: elevation, slope, curvature, topographic relief, and elevation variation coefficient, which are calculated from DEM data using GIS. Meteorological factors: wind speed, temperature, and precipitation data, obtained through multi-source meteorological datasets; Snow cover factor: Snow cover and snow depth data, combined with precipitation and temperature data to extract snowfall information; Water system factors: Level 3 river dataset.

7. A system for hierarchical prediction of avalanches based on the method of any one of claims 1-6, characterized in that, Includes the following modules: Data processing module: used for the collection, standardization, and spatial interpolation of multi-source data; Negative sample construction module: used to execute negative sample generation strategies based on randomness, physical constraints, and low-vulnerability regions; Model training module: used for sample training and parameter optimization based on the support vector machine model; Prediction and Visualization Module: Used to output the probability classification results of avalanche occurrence and generate regional risk maps.

Images