High mountain avalanche susceptibility step-coupling element feature space prediction method

Abstract

The invention discloses a mountain avalanche susceptibility step-coupling element feature space prediction method, and relates to the technical field of mountain disaster space intelligent evaluation and risk mapping. The method comprises the following steps: firstly, constructing a four-dimensional high-resolution factor grid of terrain-weather-vegetation-accumulated snow; balancing avalanche positive and negative samples by step sampling; then, through an extreme random tree calibrated by Platt, an AdaBoost and naive Bayes ternary heterogeneous learning device, outputting a first-level posterior probability in parallel, and splicing the first-level posterior probability with the original features to form high-dimensional element features; the meta-features are input into a'sparse attention-lightweight Transform 'dual-channel dynamic gating network, and secondary deep coupling and noise suppression are realized; and finally, enhancing and outputting an avalanche susceptibility probability graph with the resolution of 12.5 m during multi-scale testing. According to the avalanche risk management method, GIS interpretability is reserved, the avalanche susceptibility evaluation precision, robustness and migration capability are remarkably improved, and the avalanche risk management method can be widely applied to avalanche risk management of western traffic corridors, ski fields and national defense channels.

Description

Technical Field

[0001] This invention relates to the field of spatial intelligent assessment and risk mapping technology for mountain disasters, specifically a spatial prediction method for the tiered-coupled meta-features of high-altitude avalanche susceptibility. Background Technology

[0002] Avalanches are among the most destructive natural geological hazards in high-altitude regions. Their suddenness and high kinetic energy often lead to road disruptions, casualties, and infrastructure damage. With the rapid expansion of high-altitude transportation corridors in western China, border defense lines, and the snow and ice tourism industry, traditional risk assessment methods relying on manual inspections and expert experience are no longer sufficient to meet the needs of large-scale, high-resolution, and rapidly updated safety management.

[0003] In existing technologies, avalanche susceptibility assessment mainly adopts the following two approaches:

[0004] Deterministic methods based on physical mechanics models require a large number of field mechanical parameters as input, are computationally complex, and are difficult to generalize to regional scales.

[0005] While empirical models based on statistics or machine learning are computationally efficient, they generally suffer from problems such as imbalanced samples, reliance on human experience for feature engineering, and overfitting by a single algorithm. Furthermore, some studies have attempted to introduce deep learning into disaster susceptibility assessment; however, when faced with avalanche data characterized by small samples, high dimensionality, and strong spatial autocorrelation, they are prone to exhibiting drawbacks such as being "black box" and uninterpretable, unstable training, and poor generalization performance.

[0006] In recent years, the concept of meta-learning has been introduced into the field of geological hazards, improving prediction accuracy by "learning the basic model first, then learning the high-level representation." However, existing meta-learning frameworks often directly stack the probabilities of basic classifiers without considering the complementarity between original features and probabilistic features, and they do not design network structures specifically for the dual characteristics of avalanche factors—"local terrain control + long-range meteorological driving"—resulting in limited accuracy improvement in high-resolution scenarios of 10m–30m. Furthermore, the spatial resolution of risk maps output by existing studies is generally coarse (≥30m), making it difficult to meet the needs of refined risk management at the road and cableway levels.

[0007] Therefore, there is an urgent need for a new avalanche susceptibility prediction method that takes into account "GIS interpretability, small sample robustness, and high-resolution output". While maintaining clear physical meaning, it should fully explore the local-long-range coupling relationship between multi-source heterogeneous factors and achieve 12.5m resolution risk mapping to provide technical support for major projects and public safety in high-altitude areas. Summary of the Invention

[0008] The purpose of this invention is to provide a step-coupled meta-feature space prediction method for high-altitude avalanche susceptibility, in order to solve the technical problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the step-coupled meta-feature space of high-altitude avalanche susceptibility, comprising at least the following steps:

[0010] S1: Acquire a 12.5m resolution digital elevation model, meteorological reanalysis data, vegetation index time series data and snow cover parameter products for the study area, construct a four-dimensional gridded factor set of topographic factor set, meteorological factor set, vegetation factor set and snow cover factor set, and perform normalization processing.

[0011] S2: Based on historical avalanche cataloging, a 1:2 step sampling strategy is used to balance positive and negative samples to obtain a training set and an independent test set;

[0012] S3: Based on the training set, parallel training is performed on extreme random trees, AdaBoost and Naive Bayes triple heterogeneous learners calibrated by Platt with 5-fold cross-validation, and the output is a first-level posterior probability vector.

[0013] S4: The original normalized features are horizontally concatenated with the first-level posterior probability vector to form a high-dimensional meta-feature, which is then input into the "Sparse Attention-Lightweight Transformer" dual-channel dynamic gating network for secondary deep coupling to obtain the avalanche susceptibility probability.

[0014] S5: Multi-scale test-time enhancement and 12.5m risk map output. Multi-scale test-time enhancement is performed on the test set, and the spatial resolution of the output avalanche susceptibility probability map is 12.5m×12.5m, which is consistent with the ASTERGDEM digital elevation model. It can be directly used for subsequent path planning and disaster reduction decision-making in the GIS platform, and is divided into five risk zones according to the natural breakpoint method.

[0015] Furthermore, the set of topographic factors includes slope, aspect, curvature, topographic humidity index (TWI), and surface cutting depth;

[0016] The meteorological factor set includes annual average maximum wind speed, daily precipitation extremes, and daily temperature range;

[0017] The vegetation factor set includes the Normalized Difference Vegetation Index (NDVI) and forest canopy height;

[0018] The set of snow accumulation factors includes the historical maximum snow depth, snow density, and spring 0°C layer height.

[0019] Furthermore, S2 includes at least the following steps:

[0020] S201: Avalanche cataloging points are stratified by altitude, that is, divided into three layers according to altitude zones ≤2500m, 2500–3500m and >3500m;

[0021] S202: Randomly select positive and negative samples within each layer, with ≥30 positive samples in each layer. If insufficient, supplement with adjacent elevation points; ultimately, obtain 200 positive samples and 400 negative samples, forming a 1:2 step ratio.

[0022] S203: The training set and the independent test set are randomly divided in a 7:3 ratio, and a spatial interval of ≥500m is maintained to avoid neighboring leakage.

[0023] Furthermore, the Platt calibration in S3 uses the sigmoid calibration function and retains the calibrated probability values ​​instead of hard classification labels.

[0024] Furthermore, the "sparse attention-lightweight Transformer" dual-channel dynamic gating network in S4 includes:

[0025] TabNet-Like sparse attention towers are used to extract local feature masks. These towers achieve feature sparsity through learnable masks, with the mask sparsity constrained to between 0.6 and 0.8.

[0026] A lightweight Transformer tower is used to capture long-range factor associations. The lightweight Transformer tower retains only 2 encoder layers, has 4 heads, and the feedforward dimension is twice the embedding dimension to reduce the risk of overfitting in sparse samples in high-altitude areas.

[0027] After concatenating the two channel outputs, dynamic fusion is achieved through learnable gating weights to obtain secondary coupled features. The dynamic gating network weights the two channel features through softmax gating coefficients, the sum of which is 1, and the weights are adaptively adjusted as the samples change.

[0028] Furthermore, in the multi-scale test of S5, the enhanced noise ratio is 0.03, the enhancement number is 5, and the noise follows a uniform distribution of [-1, 1].

[0029] Furthermore, S1 includes at least the following steps:

[0030] S101: Multi-source data acquisition is carried out, including 12.5m resolution DEM data, Landsat-8 multispectral imagery, MODIS snow cover products and meteorological station observation data of the study area, and then a four-dimensional gridded factor set of topographic factor set, meteorological factor set, vegetation factor set and snow cover factor set is constructed to extract 10 evaluation factors.

[0031] S102: Perform spatial registration and resampling, unify all data to the CGCS2000 / UTMZone46N coordinate system, and resample to 12.5m resolution using bilinear interpolation to ensure raster alignment;

[0032] S103: Finally, normalization is performed, mapping all continuous factors to the [0, 1] interval using a normalization method:

[0033]

[0034] Where x is the original data value, xmin and xmax are the minimum and maximum values ​​of the data, respectively, and x′ is the normalized data value.

[0035] Furthermore, S3 includes at least the following steps:

[0036] S301: Learner parameter settings:

[0037] Extremely Random Tree (ET):

[0038] Number of trees = 300; maximum tree depth = 11; minimum number of samples for node splits = 3; prediction formula is:

[0039]

[0040] in, Let i be the prediction result for the i-th tree. The total number of trees;

[0041] AdaBoost:

[0042] Number of weak classifiers = 300; learning rate = 0.1; prediction formula is:

[0043]

[0044] in, Let t be the t-th weak classifier; Weights for weak classifiers;

[0045] Naive Bayes (GNB):

[0046] Gaussian prior assumptions and variance smoothing parameters are used. To avoid the zero probability problem, its prediction formula is:

[0047]

[0048] in, Let be the prior probability of category k. Features Conditional probability under category k;

[0049] S302: 5-fold cross-validation and Platt calibration:

[0050] Perform 5-fold cross-validation on the three heterogeneous learners mentioned above, that is, divide the training set into 5 subsets, use 4 subsets for training and 1 subset for validation each time, and repeat 5 times.

[0051] Simultaneously, the Plattsigmoid function is used to probabilistically calibrate the learner's original decision values. The calibration formula is as follows:

[0052]

[0053] in, These are the learner's initial decision values; The probability values ​​are obtained by fitting using the maximum likelihood method, retaining the calibrated probability values ​​instead of hard classification labels, thus improving the reliability of the probability output.

[0054] S303: Output the first-order posterior probability:

[0055] After training, output the first-level posterior probability vectors of the training set and independent test set on the three learners respectively. , , Each element in the vector takes values ​​in the range [0, 1], representing the probability of an avalanche occurring for the corresponding sample.

[0056] Furthermore, S4 includes at least the following steps:

[0057] S401: High-dimensional feature construction, which horizontally concatenates the 10 evaluation factors with the 3-dimensional first-order posterior probability vector output by S303 to form a 13-dimensional high-dimensional feature. This achieves the initial fusion of physical characteristics and model probabilistic characteristics;

[0058] S402: TabNet-Like Sparse Attention Tower Operation:

[0059] Embedding Mapping: The 13-dimensional features are mapped to a 64-dimensional embedding vector through a linear transformation and the ReLU activation function, i.e.:

[0060]

[0061] in, This is the weight matrix; For bias terms;

[0062] Mask generation: Feature masks are generated through two linear transformations and the softmax function. And introduce sparsity loss: );

[0063] The sparsity of the constraint mask is 0.6-0.8, that is:

[0064]

[0065] in, and It is the weight matrix of the linear transformation; Indicates the activation function; and It is a bias term; it achieves redundant feature filtering and highlights key local features.

[0066] Feature weighting and output:

[0067] The weighted features are obtained by multiplying the meta-features element-by-element by the mask. After linear transformation, ReLU activation, and layer normalization, the local features are output:

[0068]

[0069] in, Representation layer normalization; It is a weight matrix; Paranoia;

[0070] S403: Lightweight Transformer Tower Operations

[0071] Feature-wise dimensionality enhancement: Meta-features are reshaped into tensors Z of dimension (B, 12, 64) using an embedding function.

[0072]

[0073] Where B is the batch size; This is a shape transformation operation; For embedding vectors;

[0074] Long-range correlation extraction: A 2-layer Transformer encoder (dmodel=64, nhead=4, feedforward dimension=128, dropout=0.1) is used to capture long-range correlations between factors (such as the influence of distant meteorological factors on snow cover stability).

[0075] Global pooling: Perform one-dimensional average pooling on the encoder output to obtain global long-range features. .

[0076] Furthermore, the enhancement and 12.5m risk map output during the multi-scale testing includes at least the following steps:

[0077] S501: Add uniform noise δ~U[-0.03,0.03] to the test set each time;

[0078] S502: Repeat the reasoning 5 times, and take the average probability pTTA = 1 / 5·Σpk;

[0079] S503: Paste the probability values ​​back to the corresponding grid to obtain a 12.5m×12.5m avalanche susceptibility probability map;

[0080] S504: The risk zone is divided into five levels: extremely low, low, medium, high, and extremely high, using the natural breakpoint method, and a map is completed.

[0081] Compared with the prior art, the beneficial effects of the present invention are:

[0082] 1. The sample application of the present invention is more efficient. The tiered sampling still maintains the representativeness of altitude stratification at a 1:2 positive-negative ratio, and can be trained in small sample scenarios.

[0083] 2. The meta-features of this invention are interpretable; the original factors and calibration probabilities are concatenated, which preserves the physical meaning and introduces model confidence.

[0084] 3. The dual channels of this invention can complement each other, TabNet-Like local masking filters redundancy, Transformer captures long-range meteorological-snow accumulation correlation, and gating weights adaptively balance the contributions of both.

[0085] 4. This invention can output high resolution, perform 12.5m grid calculations throughout, and the final risk map can be directly connected to the fine route selection of linear projects such as roads and cableways;

[0086] 5. The present invention is robust. By using TTA to reduce the noise of a single sample, the AUC is improved by ≥3.1% on the independent test set, and it has good transferability to mountainous areas with scarce data. Attached Figure Description

[0087] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0088] Figure 1 This is the overall flowchart of the present invention;

[0089] Figure 2 This is a schematic diagram of the ROC curve of the present invention;

[0090] Figure 3 This is a schematic diagram of the PR curve of the present invention. Detailed Implementation

[0091] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0092] This invention aims to address the problems of insufficient accuracy in avalanche hazard identification, weak model generalization ability, and unstable prediction results in high-altitude areas. The method achieves high-precision output of spatial probabilities of avalanche susceptibility and risk level classification by constructing an integrated "data-model-augmentation-mapping" workflow.

[0093] Specifically as follows:

[0094] Please see Figures 1-3 A step-coupled meta-feature space prediction method for high-altitude avalanche susceptibility includes at least the following steps:

[0095] S1: Acquire a 12.5m resolution digital elevation model, meteorological reanalysis data, vegetation index time series data and snow cover parameter products for the study area, construct a four-dimensional gridded factor set of topographic factor set, meteorological factor set, vegetation factor set and snow cover factor set, and perform normalization processing.

[0096] S2: Based on historical avalanche cataloging, a 1:2 step sampling strategy is used to balance positive and negative samples to obtain a training set and an independent test set;

[0097] S3: Based on the training set, parallel training is performed on extreme random trees, AdaBoost and Naive Bayes triple heterogeneous learners calibrated by Platt with 5-fold cross-validation, and the output is a first-level posterior probability vector.

[0098] S4: The original normalized features are horizontally concatenated with the first-level posterior probability vector to form a high-dimensional meta-feature, which is then input into the "Sparse Attention-Lightweight Transformer" dual-channel dynamic gating network for secondary deep coupling to obtain the avalanche susceptibility probability.

[0099] S5: Multi-scale test-time enhancement and 12.5m risk map output. Multi-scale test-time enhancement is performed on the test set, and the spatial resolution of the output avalanche susceptibility probability map is 12.5m×12.5m, which is consistent with the ASTERGDEM digital elevation model. It can be directly used for subsequent path planning and disaster reduction decision-making in the GIS platform, and is divided into five risk zones according to the natural breakpoint method.

[0100] The set of topographic factors includes slope, aspect, curvature, topographic moisture index (TWI), and surface incision depth.

[0101] The meteorological factor set includes annual average maximum wind speed, daily precipitation extremes, and diurnal temperature range;

[0102] The vegetation factor set includes the Normalized Difference Vegetation Index (NDVI) and forest canopy height;

[0103] The snow accumulation factor set includes the historical maximum snow depth, snow density, and spring 0°C layer height.

[0104] S2 includes at least the following steps:

[0105] S201: Avalanche cataloging points are stratified by altitude, that is, divided into three layers according to altitude zones ≤2500m, 2500–3500m and >3500m;

[0106] S202: Randomly select positive and negative samples within each layer, with ≥30 positive samples in each layer. If insufficient, supplement with adjacent elevation points; ultimately, obtain 200 positive samples and 400 negative samples, forming a 1:2 step ratio.

[0107] S203: The training set and the independent test set are randomly divided in a 7:3 ratio, and a spatial interval of ≥500m is maintained to avoid neighboring leakage.

[0108] The Platt calibration in S3 uses the sigmoid calibration function and retains the calibrated probability values ​​instead of hard classification labels.

[0109] The "Sparse Attention-Lightweight Transformer" dual-channel dynamic gating network in S4 includes:

[0110] TabNet-Like sparse attention towers are used to extract local feature masks. The TabNet-Like sparse attention towers achieve feature sparsity through learnable masks, with mask sparsity constrained to between 0.6 and 0.8.

[0111] Lightweight Transformer Tower is used to capture long-range factor associations. The lightweight Transformer Tower retains only 2 encoder layers with 4 heads and the feedforward dimension is twice the embedding dimension to reduce the risk of overfitting for sparse samples in high-altitude areas.

[0112] After concatenating the outputs of the two channels, dynamic fusion is achieved through learnable gating weights to obtain secondary coupled features. The dynamic gating network weights the features of the two channels through softmax gating coefficients, the sum of which is 1, and the weights are adaptively adjusted as the samples change.

[0113] In the multi-scale test in S5, the enhanced noise ratio is 0.03, the enhancement number is 5, and the noise follows a uniform distribution of [-1, 1].

[0114] S1 includes at least the following steps:

[0115] S101: Multi-source data acquisition is carried out, including 12.5m resolution DEM data, Landsat-8 multispectral imagery, MODIS snow cover products and meteorological station observation data of the study area, and then a four-dimensional gridded factor set of topographic factor set, meteorological factor set, vegetation factor set and snow cover factor set is constructed to extract 10 evaluation factors.

[0116] S102: Perform spatial registration and resampling, unify all data to the CGCS2000 / UTMZone46N coordinate system, and resample to 12.5m resolution using bilinear interpolation to ensure raster alignment;

[0117] S103: Finally, normalization is performed, mapping all continuous factors to the [0, 1] interval using a normalization method:

[0118]

[0119] Where x is the original data value, xmin and xmax are the minimum and maximum values ​​of the data, respectively, and x′ is the normalized data value.

[0120] S3 includes at least the following steps:

[0121] S301: Learner parameter settings:

[0122] Extremely Random Tree (ET):

[0123] Number of trees = 300; maximum tree depth = 11; minimum number of samples for node splits = 3; prediction formula is:

[0124]

[0125] in, Let i be the prediction result for the i-th tree. The total number of trees;

[0126] AdaBoost:

[0127] Number of weak classifiers = 300; learning rate = 0.1; prediction formula is:

[0128]

[0129] in, Let t be the t-th weak classifier; Weights for weak classifiers;

[0130] Naive Bayes (GNB):

[0131] Gaussian prior assumptions and variance smoothing parameters are used. To avoid the zero probability problem, its prediction formula is:

[0132]

[0133] in, Let be the prior probability of category k. Features Conditional probability under category k;

[0134] S302: 5-fold cross-validation and Platt calibration:

[0135] Perform 5-fold cross-validation on the three heterogeneous learners mentioned above, that is, divide the training set into 5 subsets, use 4 subsets for training and 1 subset for validation each time, and repeat 5 times.

[0136] Simultaneously, the Plattsigmoid function is used to probabilistically calibrate the learner's original decision values. The calibration formula is as follows:

[0137]

[0138] in, These are the learner's initial decision values; The probability values ​​are obtained by fitting using the maximum likelihood method, retaining the calibrated probability values ​​instead of hard classification labels, thus improving the reliability of the probability output.

[0139] S303: Output the first-order posterior probability:

[0140] After training, output the first-level posterior probability vectors of the training set and independent test set on the three learners respectively. , , Each element in the vector takes values ​​in the range [0, 1], representing the probability of an avalanche occurring for the corresponding sample.

[0141] S4 includes at least the following steps:

[0142] S401: High-dimensional feature construction, which horizontally concatenates the 10 evaluation factors with the 3-dimensional first-order posterior probability vector output by S303 to form a 13-dimensional high-dimensional feature. This achieves the initial fusion of physical characteristics and model probabilistic characteristics;

[0143] S402: TabNet-Like Sparse Attention Tower Operation:

[0144] Embedding Mapping: The 13-dimensional features are mapped to a 64-dimensional embedding vector through a linear transformation and the ReLU activation function, i.e.:

[0145]

[0146] in, This is the weight matrix; For bias terms;

[0147] Mask generation: Feature masks are generated through two linear transformations and the softmax function. And introduce sparsity loss: );

[0148] The sparsity of the constraint mask is 0.6-0.8, that is:

[0149]

[0150] in, and It is the weight matrix of the linear transformation; Indicates the activation function; and It is a bias term; it achieves redundant feature filtering and highlights key local features.

[0151] Feature weighting and output:

[0152] The weighted features are obtained by multiplying the meta-features element-by-element by the mask. After linear transformation, ReLU activation, and layer normalization, the local features are output:

[0153]

[0154] in, Representation layer normalization; It is a weight matrix; Paranoia;

[0155] S403: Lightweight Transformer Tower Operations

[0156] Feature-wise dimensionality enhancement: Meta-features are reshaped into tensors Z of dimension (B, 12, 64) using an embedding function.

[0157]

[0158] Where B is the batch size; This is a shape transformation operation; For embedding vectors;

[0159] Long-range correlation extraction: A 2-layer Transformer encoder (dmodel=64, nhead=4, feedforward dimension=128, dropout=0.1) is used to capture long-range correlations between factors (such as the influence of distant meteorological factors on snow cover stability).

[0160] Global pooling: Perform one-dimensional average pooling on the encoder output to obtain global long-range features. .

[0161] The enhancement and 12.5m risk map output during multi-scale testing should include at least the following steps:

[0162] S501: Add uniform noise δ~U[-0.03,0.03] to the test set each time;

[0163] S502: Repeat the reasoning 5 times, and take the average probability pTTA = 1 / 5·Σpk;

[0164] S503: Paste the probability values ​​back to the corresponding grid to obtain a 12.5m×12.5m avalanche susceptibility probability map;

[0165] S504: The risk zone is divided into five levels: extremely low, low, medium, high, and extremely high, using the natural breakpoint method, and a map is completed.

[0166] In summary:

[0167] This invention first constructs a four-dimensional high-resolution factor raster of "topography-meteorology-vegetation-snow cover"; then, it balances avalanche positive and negative samples using stepped sampling; subsequently, it outputs the first-level posterior probability in parallel using a Platt-calibrated extreme random tree, AdaBoost, and Naive Bayes ternary heterogeneous learner, and concatenates it with the original features to form high-dimensional meta-features; then, it inputs the meta-features into a dual-channel dynamic gating network of "sparse attention-lightweight Transformer" to achieve secondary deep coupling and noise suppression; finally, it enhances the output of an avalanche susceptibility probability map with a resolution of 12.5m after multi-scale testing; this invention significantly improves the accuracy, robustness, and transferability of avalanche susceptibility assessment while retaining GIS interpretability, and can be widely applied to avalanche risk management in western transportation corridors, ski resorts, and national defense corridors.

[0168] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A step-coupling element feature space prediction method for high mountain avalanche liability, characterized by: At least comprising the following steps: S1: Obtain the 12.5m resolution digital elevation model, meteorological reanalysis data, vegetation index time series data and snow parameter products of the study area, construct the four-dimensional grid factor set of terrain factor set, meteorological factor set, vegetation factor set and snow factor set, and carry out normalization processing; S2: According to the historical avalanche catalog, a 1:2 step sampling strategy is adopted to balance the positive and negative samples, and the training set and independent test set are obtained; S3: Based on the training set, the extreme random tree, AdaBoost and naive Bayes three heterogeneous learners calibrated by 5-fold cross-validation Platt are trained in parallel, and the first-level posterior probability vector is output; S4: The original normalized features and the first-level posterior probability vector are transversely spliced to form high-dimensional meta features, which are input into the "sparse attention-lightweight Transformer" dual-channel dynamic gating network for secondary deep coupling to obtain the snow avalanche susceptibility probability; S5: Multi-scale test time augmentation and 12.5m risk map output, multi-scale test time augmentation is performed on the test set, and the spatial resolution of the output snow avalanche susceptibility probability map is 12.5m×12.5m, which is consistent with the ASTER GDEM digital elevation model. It can be directly used for subsequent path planning and disaster reduction decision-making on GIS platform, and is divided into five risk zones according to the natural breakpoint method.

2. The step-coupling element feature space prediction method for high mountain snow avalanche susceptibility according to claim 1, characterized in that: The terrain factor set includes slope, aspect, curvature, terrain wetness index TWI, and surface cutting depth; The meteorological factor set includes annual average maximum wind speed, daily precipitation extreme value, and temperature range; The vegetation factor set includes normalized difference vegetation index NDVI and forest canopy height; The snow factor set includes historical maximum snow depth, snow density and spring 0℃ layer height.

3. The step-coupling element feature space prediction method for high mountain snow avalanche susceptibility according to claim 1, characterized in that: S2 at least comprises the following steps: S201: Stratify the avalanche catalog points by altitude, i.e. stratify them into three layers according to the altitude bands ≤2500m, 2500-3500m and >3500m; S202: Randomly extract positive and negative samples within each layer, with ≥30 positive samples in each layer, and supplement adjacent altitude points if necessary; finally obtain 200 positive samples and 400 negative samples, forming a 1:2 step ratio; S203: Stratify and randomly divide the training set and independent test set according to the 7:3 ratio, and keep the spatial interval ≥500m to avoid adjacent leakage.

4. The step-coupling element feature space prediction method for high mountain snow avalanche susceptibility according to claim 1, characterized in that: The Platt calibration in S3 uses a sigmoid calibration function, and the calibrated probability value is retained instead of a hard classification label.

5. The step-coupling element feature space prediction method for high mountain snow avalanche susceptibility according to claim 1, characterized in that: The "sparse attention-lightweight Transformer" dual-channel dynamic gating network in S4 includes: TabNet-like sparse attention tower for extracting local feature masks, which realizes feature sparsification through a learnable mask with a mask sparsity constraint of 0.6-0.8 Lightweight Transformer tower for capturing long-range factor associations, which only retains 2 layers of encoder with a head number of 4 and a feedforward dimension of 2 times the embedding dimension to reduce the overfitting risk of sparse samples in high mountain areas; The two-channel outputs are concatenated, and a learnable gating weight is used to dynamically fuse the two channels to obtain secondary coupled features.

6. The step-coupling element feature space prediction method for high mountain snow avalanche susceptibility according to claim 1, characterized in that: The enhanced noise ratio in the multi-scale test in S5 is 0.03, the enhancement times are 5, and the noise is uniformly distributed in [-1, 1].

7. The step-coupling element feature space prediction method for high mountain snow avalanche susceptibility according to claim 1, characterized in that: The S1 at least includes the following steps: S101: Multi-source data acquisition is performed to acquire 12.5m resolution DEM data, Landsat-8 multispectral images, MODIS snow products and meteorological station observation data in the study area, and then four-dimensional grid factor sets of terrain factor set, meteorological factor set, vegetation factor set and snow factor set are constructed to extract 10 types of evaluation factors; S102: Spatial registration and resampling are performed to unify all data to the CGCS2000 / UTMZone46N coordinate system, and the bilinear interpolation method is used to resample to 12.5m resolution to ensure grid alignment; S103: Finally, normalization processing is performed, and all continuous factors are mapped to the [0, 1] interval using a normalization method: Where x is the original data value, xmin and xmax are the minimum and maximum values of the data respectively, and x' is the normalized data value.

8. The step-coupling element feature space prediction method for high mountain snow avalanche susceptibility according to claim 7, characterized in that: The S3 at least includes the following steps: S301: Learner parameter setting: Extreme Random Tree: Number of trees = 300; Maximum depth of tree = 11; Minimum number of samples for node splitting = 3, and its prediction formula is: wherein, is the prediction result for the i-th tree, is the total number of trees; AdaBoost: Number of weak classifiers = 300; Learning rate = 0.1; and its prediction formula is: wherein, is the tth weak classifier; is the weak classifier weight; Naive Bayes: With Gaussian prior assumption, variance smoothing parameter , avoiding zero probability problem, its prediction formula is: wherein, prior probability of class k, feature conditional probability under class k; S302: 5-fold cross-validation and Platt calibration: 5-fold cross-validation is performed on the above three heterogeneous learners, that is, the training set is divided into 5 subsets, 4 subsets are used for training and 1 subset is used for validation each time, and the cycle is repeated 5 times; At the same time, the Platt sigmoid function is used to calibrate the probability of the original decision value of the learner, and the calibration formula is: wherein, is the original decision value of the learner; The maximum likelihood method is used for fitting, and the calibrated probability value is reserved instead of a hard classification label, so that the reliability of the probability output is improved. S303: Output of first-level posterior probability: After training, the first-level posterior probability vectors of the training set and the independent test set on the three learners are output respectively , , Each element in the vector ranges from 0 to 1, representing the probability of avalanche occurrence of the corresponding sample.

9. The step-coupling element feature space prediction method for high mountain snow avalanche susceptibility according to claim 8, characterized in that: The S4 at least includes the following steps: S401: high-dimensional element feature construction, 10 evaluation factors are transversely spliced with the 3-dimensional first-order posterior probability vector output by S303 to form a 13-dimensional high-dimensional feature , realizing preliminary fusion of physical features and model probability features; S402: TabNet-Like sparse attention tower operation: Embedding mapping: 13-dimensional meta-features are mapped to 64-dimensional embedding vectors through linear transformation and ReLU activation function, that is: wherein, is a weight matrix; is a bias term; Mask generation: generate feature mask by two linear transformations and softmax function and introduce sparsity loss: ); The constraint mask sparsity is 0.6-0.8, that is: wherein, and is a weight matrix of the linear transformation; denotes an activation function; and is a bias term; implement redundant feature filtering, highlight key local features; Feature weighting and output: element-wise multiplication of the meta-feature and the mask to obtain weighted features and output local features after linear transformation, ReLU activation and layer normalization wherein, represents a layer normalization; is a weight matrix; is a bias term; S403: Lightweight Transformer tower operation: Feature-wise dimensionality increase: The meta-features are reshaped into a tensor Z with dimensions (B, 12, 64) through an embedding function, that is where B is a batch size; is a shape transform operation; is an embedding vector; Long-range association extraction: A 2-layer Transformer encoder is used to capture the long-range association between factors; Global pooling: One-dimensional average pooling is performed on the encoder output to obtain global long-range features: .

10. The step-coupling element feature space prediction method for high mountain snow avalanche susceptibility according to claim 9, characterized in that: The multi-scale test enhancement and 12.5m risk map output at least include the following steps: S501: Add uniform noise δ ~ U[-0.03, 0.03] to the test set each time; S502: Repeat the inference 5 times, and take the average probability pTTA = 1 / 5·Σpk; S503: Paste the probability value to the corresponding grid to get the 12.5m×12.5m avalanche susceptibility probability map; S504: Divide into five levels of risk areas of very low, low, medium, high, and very high by using the natural breakpoint method, and complete the mapping.

Images