A debris flow prediction method based on material initiation-propagation process

Abstract

The present invention relates to the field of debris flow prediction technology, and in particular to a debris flow prediction method based on a material triggering-propagation process, comprising: collecting debris flow events and rainfall events; obtaining geographical background conditions and rainfall triggering conditions based on the debris flow events and rainfall events; constructing an index system for predicting debris flows based on the geographical background conditions and rainfall triggering conditions; inputting the indicators in the index system into a debris flow prediction model, and outputting the spatiotemporal probability value of the occurrence of a debris flow, wherein the debris flow prediction model is constructed based on a whale optimization algorithm and a machine learning model. The present invention integrates triggering conditions and geographical background features through a hybrid machine learning model, and evaluates the spatiotemporal possibility of debris flows based on hourly watersheds from the perspective of material triggering-propagation of debris flows. The method is implemented in a typical terrain abrupt change zone, provides technical guidance for early warning and prediction, and serves the prevention and mitigation of debris flows in mountainous areas.

Description

Technical Field

[0001] The present invention relates to the technical field of debris flow prediction, and in particular to a debris flow prediction method based on a material initiation-propagation process. Background Art

[0002] Numerous studies have investigated the spatial probability of debris flow occurrence, focusing on debris flow susceptibility. These methods include heuristic approaches that rely heavily on expert experience; statistical methods such as frequency ratios, principal component analysis, and logistic regression; and physical methods that require high data processing and parameter calibration, such as trigger-propagation modeling supported by the TRIGRS and FLO-2D models. Although these methods can identify the distribution of hazard-prone locations, they remain insufficient for disaster prevention and mitigation decision-making. The temporal probability of future debris flow occurrences still needs to be foreseen.

[0003] Many researchers and institutions have been working to predict the spatiotemporal likelihood of debris flows using data-driven or statistical models. However, these approaches still have significant limitations due to their reliance on historical disaster records. Specifically, these methods are unsuitable for regions with limited historical disaster information, as they rely on analyzing the relationship between historical disaster occurrences and rainfall characteristics to predict debris flow likelihood. Overall, generalizing and updating ID thresholds is challenging.

[0004] Basins with a continuous geographic distribution often experience clustered debris flows, which occur simultaneously during the same rainfall event. Similar rainfall thresholds within clustered basins are thought to be due to similar geographic contextual conditions that contribute to the hazards, such as topography, tectonic conditions, hydrology, and source material conditions. These observations have led researchers to understand that rainfall thresholds for triggering debris flows need to be updated to account for changes in the geographic context. Consequently, researchers have proposed that the key factors driving debris flow formation can be divided into two categories: geographic contextual conditions that control a location's susceptibility to debris flows, and triggering conditions, such as extreme rainfall and localized increases in soil moisture, that control when debris flows occur. Subsequently, several studies have enhanced rainfall threshold methods by incorporating geographic contextual variables. This allows for the prediction of debris flow occurrence in locations where hazard data is scarce. For example, a global prediction model, called LHASA version 2, has been described that uses a global landslide susceptibility map to represent geographic contextual conditions, employing a data-driven approach. Although a grid map of potential hazardous areas over a 60° north-south range can be generated every 3 hours by combining satellite precipitation products with global sensitivity maps, the ground application of this model is still limited due to its coarse kilometer-level spatial resolution and 3-hour temporal resolution. Bordoni, M., V. Vivaldi, L. Lucchelli,L. Ciabatta, L. Brocca, JP Galve, and C. Meisina. 2021. Development of adata-driven model for spatial and temporal shallow landslide probability ofoccurrence at catchment scale. Landslides 18(4): 1209-1229. developed a data-driven approach to predict the spatiotemporal probability of rainfall-triggered shallow landslides, supporting the sensitivity results calculated from geological, geomorphological and hydrological predictors, with temporal triggering conditions involving short-term accumulated rainfall, antecedent rainfall and soil hydrological conditions. While these methods allow researchers to predict the spatiotemporal probability of landslides or debris flows in areas lacking a historical record of disasters, they are based on grid-based assessment objects and are only suitable for coarse regional predictions. This limits their application to predicting debris flows developing within a watershed. This is due to a lack of consideration of the physical mechanisms by which unstable material feeds debris flows and is transported to gully mouths, causing damage.

[0005] Therefore, the present invention proposes a debris flow prediction method based on the material initiation-propagation process. Summary of the Invention

[0006] The purpose of this invention is to provide a debris flow prediction method based on the material initiation-propagation process. By integrating triggering conditions and geographical background characteristics through a hybrid machine learning model, the spatiotemporal possibility of debris flow occurrence based on hourly watersheds is evaluated from the perspective of material triggering-propagation of debris flows. This method is implemented in a typical terrain transition zone, providing technical guidance for early warning and prediction, and serving the prevention and mitigation of debris flows in mountainous areas.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A debris flow prediction method based on the material initiation-propagation process includes:

[0009] Collect debris flow events and rainfall events;

[0010] According to the debris flow event and the rainfall event, obtaining geographical background conditions and rainfall excitation conditions;

[0011] Constructing an index system for predicting debris flow based on the geographical background conditions and the rainfall excitation conditions;

[0012] The indicators in the indicator system are input into the debris flow prediction model, and the spatiotemporal probability value of the occurrence of debris flow is output, wherein the debris flow prediction model is constructed based on the whale optimization algorithm and the machine learning model.

[0013] Optionally, obtaining the geographical background conditions includes: selecting key geographical background indicators based on the conditions for the formation of debris flows to obtain the geographical background conditions, wherein the key geographical background indicators include geological and structural indicators, topographic and geomorphological indicators, hydrological and meteorological indicators, and ecological and human indicators.

[0014] Optionally, obtaining the rainfall triggering condition includes:

[0015] Acquiring a rainfall sequence according to the debris flow event and the rainfall event;

[0016] Extracting features from the rainfall sequence to obtain rainfall features;

[0017] The TRIGRS model is used to identify the distribution of unstable substances and characterize the safety factor of substance stability, and to calculate the substance trigger-propagation index;

[0018] The rainfall excitation condition is obtained according to the rainfall characteristics and the material trigger-propagation index.

[0019] Optionally, performing feature extraction on the rainfall sequence to obtain rainfall features includes:

[0020] According to the traditional ID threshold curve theory, traditional rainfall statistical features are extracted from the rainfall sequence, wherein the traditional rainfall statistical features include average rainfall intensity and rainfall event duration;

[0021] Extract features of the rainfall sequence using the Tsfresh tool to obtain rainfall driving features, wherein the rainfall driving features include absolute energy, sum value, approximate entropy, continuous wavelet transform coefficient, autocorrelation, upper quantile, and lower quantile;

[0022] The rainfall characteristics are obtained by combining the traditional rainfall statistical characteristics and the rainfall driving characteristics.

[0023] Optionally, obtaining the average rainfall intensity includes: calculating the average rainfall intensity by dividing the accumulated rainfall by the duration of the rainfall event.

[0024] Optionally, calculating the substance trigger-propagation index includes:

[0025] ;

[0026] ;

[0027] Among them, MTP g Indicates the material trigger-propagation index of stage g, A ust,g represents the unstable area in the basin at stage g, V ust,g represents the volume of unstable material in the basin determined by the TRIGRS model, IC std,g represents the normalized connectivity index of stage g, A wetershed represents the total area of ​​the basin, A unit Denotes the area of ​​the computational unit, D sld,g It represents the failure surface depth of the calculation unit with the minimum safety factor output by the TRIGRS model.

[0028] Optionally, the indicator system for predicting debris flows includes: MTP indicator, ID threshold indicator, Tsfresh indicator, and ENV indicator, wherein the MTP indicator is the material trigger-propagation index, the ID threshold indicator includes the average rainfall intensity and the duration of the rainfall event, the Tsfresh indicator is the rainfall driving feature extracted from the rainfall sequence by the Tsfresh tool, and the ENV indicator includes the geographical background key indicators and the effective rainfall before the debris flow breaks out.

[0029] Optionally, constructing the debris flow prediction model based on the whale optimization algorithm and the machine learning model includes: training the optimal hyperparameters of the machine learning model through the whale optimization algorithm to construct the debris flow prediction model, wherein the machine learning model is support vector classification, random forest and extreme gradient boosting.

[0030] The beneficial effects of the present invention are as follows: the present invention first constructs a dynamic hazard assessment index system for debris flow from the perspective of the debris flow formation mechanism, and the system consists of two parts. One part is the geographical background conditions, which are used to characterize the susceptibility of debris flow occurrence, that is, the spatial possibility of debris flow in a certain area. The other part is the triggering conditions, which are used to characterize the triggering characteristics related to rainfall. Subsequently, by integrating three advanced machine learning models and a hyperparameter optimization algorithm, a hybrid machine learning model was established, and the model convergence speed and generalization ability were optimized to simulate the impact of susceptibility differences controlled by geographical background conditions on changes in rainfall thresholds. In this way, the hybrid machine learning model can be used to fit the complex nonlinear relationship between debris flow occurrence and disaster-causing conditions, thereby predicting the spatiotemporal possibility of debris flow occurrence and providing technical support for debris flow forecasting and early warning in complex mountainous areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0032] Figure 1 is a schematic diagram of an infinite slope model according to an embodiment of the present invention;

[0033] FIG2 is a comparison of ROC curves of different hybrid machine learning models under different parameter design schemes according to an embodiment of the present invention, wherein (a) is the ROC verification curve of the WOA-SVC model under different parameter schemes, (b) is the ROC verification curve of the WOA-RF model under different parameter schemes, and (c) is the ROC verification curve of the WOA-XGB model under different parameter schemes;

[0034] FIG3 is a spatiotemporal probability prediction of a debris flow in a target area according to an embodiment of the present invention, wherein (a) is the probability level, and (b) is the hourly rainfall conditions and material triggering-propagation conditions from the start of the rainfall event to the onset of the debris flow;

[0035] Figure 4 shows the spatiotemporal probability prediction of the second debris flow in the target area according to an embodiment of the present invention, where (a) is the probability level, and (b) is the hourly rainfall conditions and material trigger-propagation conditions from the start of the rainfall event to the onset of the debris flow;

[0036] Figure 5 This is a flow chart of a debris flow prediction method based on the material initiation-propagation process according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0038] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 5 This embodiment provides a debris flow prediction method based on the material initiation-propagation process, including:

[0040] Collect debris flow events and rainfall events;

[0041] According to debris flow events and rainfall events, obtain geographical background conditions and rainfall triggering conditions;

[0042] Construct an index system for predicting debris flows based on geographical background conditions and rainfall triggering conditions;

[0043] The indicators in the indicator system are input into the debris flow prediction model, and the spatiotemporal probability value of the occurrence of debris flow is output. The debris flow prediction model is constructed based on the whale optimization algorithm and the machine learning model.

[0044] Specifically, the collection of debris flow events includes: The main purpose of collecting debris flow records is to train the hybrid machine learning model and evaluate its prediction accuracy and practical applicability. The collected debris flow events include a comprehensive record of rainfall events, covering rainfall sequences, locations and times of occurrence. This extensive dataset helps to explore the factors that trigger debris flows and conduct in-depth research on the physical mechanisms of their formation. In addition, supplementary information on debris flow events was retrieved from literature databases. Using 15 years of literature records collected by the Institute of Plateau Meteorology in Chengdu, China, the corresponding hourly rainfall sequences that triggered debris flows were compiled.

[0045] Rainfall event collection involves defining a rainfall event that triggers a debris flow: a rainfall event is considered to begin when the hourly rainfall intensity reaches or exceeds 4 mm / hour; a rainfall event is considered to end when the rainfall intensity remains below 4 mm / hour for six consecutive hours. Thus, rainfall sequences closely associated with debris flow formation were identified to capture the temporal variation in rainfall statistical characteristics. Subsequently, a rolling method was used to partition the rainfall sequences into subsequences at one-hour intervals. Ultimately, 209 rainfall subsequences were generated, and the SMOTE tool was used to balance the samples, resulting in 360 samples.

[0046] Furthermore, obtaining geographical background conditions includes: selecting key geographical background indicators according to the conditions for debris flow formation to obtain geographical background conditions, wherein the key geographical background indicators include geological and structural indicators, topographic and geomorphological indicators, hydrological and meteorological indicators, and ecological and human indicators.

[0047] Specifically, "debris flow susceptibility" refers to the relative likelihood of a debris flow occurring or being triggered within a specific area. This comprehensive concept can be applied in a variety of practical scenarios, such as risk assessment and disaster prevention planning. To describe and assess debris flow susceptibility and characterize the critical conditions for debris flow occurrence within a watershed, four key indicators were selected based on the three primary conditions for debris flow formation: steep terrain, abundant source material, and sufficient water resources. These indicators include: (i) geological and tectonic indicators, such as lithology (LIT), soil erosion index (SE), soil depth (SD), distance to fault (DFF), distance to the epicenter of historical earthquakes (DFE) (magnitude greater than 4 since 2000), and peak seismic acceleration (PSA); (ii) topographic and geomorphological indicators, such as channel gradient (CG) and relief of terrain (REL); (iii) hydrometeorological indicators, such as drought index (AI), average annual precipitation (AP), and average annual temperature (MAT); and (iv) ecological and anthropogenic indicators, such as road density (RD), population density (PD), land use type (LD), and normalized difference vegetation index (NDVI).

[0048] Furthermore, obtaining rainfall triggering conditions includes:

[0049] Obtain rainfall sequences based on debris flow events and rainfall events;

[0050] Extract features from rainfall sequences to obtain rainfall characteristics;

[0051] The TRIGRS model is used to identify the distribution of unstable substances and characterize the safety factor of substance stability, and to calculate the substance trigger-propagation index;

[0052] According to rainfall characteristics and material trigger-propagation index, rainfall excitation conditions are obtained.

[0053] Specifically, debris flows are soil-water mixtures formed by rainfall runoff. Therefore, this embodiment combines dynamic rainfall statistical features and rainfall-driven material replenishment characteristics to characterize the triggering conditions of debris flows. Rainfall-driven triggering features are therefore divided into two categories: (i) statistical features extracted from rainfall sequences, and (ii) rainfall-driven material triggering and propagation features.

[0054] Furthermore, feature extraction is performed on the rainfall sequence to obtain rainfall features including:

[0055] According to the traditional ID threshold curve theory, traditional rainfall statistical features are extracted from the rainfall sequence, where the traditional rainfall statistical features include average rainfall intensity and rainfall event duration;

[0056] The Tsfresh tool is used to extract the features of the rainfall sequence to obtain rainfall driving characteristics, including absolute energy, sum value, approximate entropy, continuous wavelet transform coefficient, autocorrelation, upper quantile and lower quantile.

[0057] The rainfall characteristics are obtained by combining traditional rainfall statistical characteristics and rainfall driving characteristics.

[0058] Specifically, obtaining rainfall characteristics includes:

[0059] (1) Statistical feature extraction of rainfall series:

[0060] Rainfall statistical characteristics are quantitative statistical information extracted from rainfall events. These characteristics include the spatial and temporal distribution of rainfall events, as well as rainfall intensity, rainfall amount, duration, and frequency. Analyzing these characteristics helps determine the scale, intensity, and duration of rainfall events that trigger debris flows.

[0061] Traditional rainfall statistical characteristics:

[0062] Based on the traditional ID threshold curve theory, traditional rainfall statistical features, including average rainfall intensity (ARI) and rainfall event duration (RED), are extracted from rainfall sequences. The average rainfall intensity (ARI) of each rainfall subsequence is calculated by dividing the cumulative rainfall by the rainfall event duration (RED).

[0063] In addition, the effective rainfall before the debris flow occurs can be quantified by the formula.

[0064] (1);

[0065] Among them, ERA k represents the effective rainfall in the k days before the debris flow occurs; K represents the reduction coefficient, which is 0.84; k represents the number of days, R i represents the rainfall on the previous i-th day.

[0066] More statistical features extracted by the Tsfresh tool:

[0067] Extracting additional rainfall statistical features through the Tsfresh tool is crucial for building comprehensive debris flow prediction models, complementing traditional rainfall statistical features. Relying solely on traditional indicators may not fully capture the inherent complexity and variability of rainfall events. The Tsfresh tool automatically extracts features from time series data to capture a more comprehensive and richer set of rainfall statistical characteristics, thereby enhancing the model's expressive power. Furthermore, debris flow processes are complex, and there is a potential nonlinear relationship between rainfall and debris flows. By introducing the additional rainfall statistical features extracted by the Tsfresh tool, the complex correlations between rainfall and debris flows can be better captured, thereby improving the prediction model's ability to identify potential nonlinear relationships. Furthermore, the Tsfresh tool can extract various rainfall statistical features, such as temporal, spectral, and information entropy. This enables the prediction model to comprehensively analyze rainfall from multiple perspectives and dimensions, reflecting key characteristics such as the spatiotemporal distribution and intensity variations of rainfall events, thereby improving the model's robustness and adaptability.

[0068] The Python package Tsfresh was used to extract several features that significantly contribute to debris flow formation from rainfall time series. 143 features were extracted from the rainfall time series. Machine learning techniques were used to analyze the relationship between these features and debris flow occurrence within the basin, ultimately identifying seven features that significantly contribute to debris flow formation. These seven rainfall-driven features include absolute energy (AE), sum value (SV), approximate entropy (AEMR), continuous wavelet transform coefficient (CWT), autocorrelation coefficient (ATC), upper quantile (CQ1), and lower quantile (CQ2). A detailed description of these metrics is provided in Table 1. The variable x represents the rainfall time series.

[0069] Table 1

[0070]

[0071] (2) Calculation of rainfall-driven material propagation characteristics:

[0072] Dynamic characterization of material replenishment during rainfall is crucial for debris flow prediction, as it is an integral component of debris flow composition. The research method utilizes the TRIGRS model to capture the distribution of unstable materials within a watershed by inputting sediment characteristics, spatial sediment distribution (i.e., soil depth), rainfall data, and topographic conditions, and uses this as the primary source of debris flow material. Subsequently, a propagation model (FLO-2D) is used to simulate the triggering and propagation processes of debris flows. However, due to the limitations of the detailed numerical simulation of debris flow propagation in this research method in terms of computational resources, time cost, and application restrictions, this method has certain shortcomings in multi-basin early warning tasks.

[0073] Geomorphic connectivity refers to the probability of material transport and convergence, which can be quantified using the Geomorphic Connectivity Index (IC). This index describes the likelihood of an object being transported from point A to point B within a watershed, providing a simplified approach for modeling the triggering and propagation of debris flows. This example uses the TRIGRS model to quantify the distribution of unstable materials within a watershed. The connectivity index then represents the degree of connection between the location of unstable materials and the convergence points of debris flow basins, describing the ability of unstable materials to feed debris flows. The specific implementation process is as follows:

[0074] i. Using the TRIGRS model to identify unstable species distribution:

[0075] To assess the distribution of unstable materials within a given watershed during a specific rainfall event, Fortran-compiled software uses the Transfer of Soil and Rock Stability by Rainfall and Infiltration (TRIGRS) model. By combining TRIGRS output with geomorphic connectivity, the transport potential of identified unstable materials is quantified. This information is then incorporated as a hazard factor into a hybrid machine learning model to predict the likelihood of debris flow occurrence. Figure 1 A schematic diagram showing the model parameter partitioning setup.

[0076] ψ(Z,t) represents the pressure head at vertical depth Z and time t. ψ(Z,t) is calculated using Equation (2).

[0077] (2);

[0078] Where d is the initial water level depth, the square of the cosine of the slope β=cos²δ, δ is the terrain slope, I nZ is the surface flux of a given intensity in the nth time interval, H(·) is the Heaviside step function, t n is the time of the nth time interval in the rainfall sequence, D1=D0 / cos²δ, D0 is the saturated hydraulic diffusivity (equal to K s 200 times of d LZ is the thickness of the permeable soil layer, , erfc(·) is the complementary error function. K s is the saturated hydraulic conductivity, N is the total number of time intervals, and m is the number of terms in the infinite series. Since the infinite series converges quickly in this formula, only the first few terms are needed in the calculation process.

[0079] The model is based on the following assumption: the solution given by equation (2) is only applicable when the initial condition is tension saturation. Therefore, it can be inferred that the linearized form of Richards' equation is valid in this case, and the hydraulic conductivity can be approximated by its saturation value. It should be noted that within the framework of this embodiment, rainfall will not produce runoff that exceeds the infiltration capacity. After determining the pore water pressure head through equation (2), the safety factor (F) representing the stability of the material can be calculated based on the infinite slope model, as shown in equation (3). s ).

[0080] (3);

[0081] Among them, F s (Z,t) represents the safety factor. c′ and represent soil cohesion and soil internal friction angle respectively; γ w and γ s Denote the bulk density of water and soil respectively. According to existing research, the damage occurs at the base boundary Z=d LZ The pore water pressure reaches its maximum depth at this point. Basic soil mechanics parameters (such as cohesion, internal friction angle, and bulk density) for different soil types are assigned values ​​based on the established mapping relationship between soil types and basic soil parameters.

[0082] ii: Material Trigger-Propagation (MTP) Index Calculation:

[0083] The Material Triggering-Propagation (MTP) index is determined by calculating the average potential for unstable materials to be eroded and enter the river channel using equations (4) and (5). Equation (4) quantifies the erosive intensity of potentially unstable materials, while equation (5) assesses the probability of their migration along the terrain into the river channel. This index comprehensively reflects the coupled effects of topographic characteristics, soil erodibility, and hydrodynamic conditions on material transport processes.

[0084] (4);

[0085] (5);

[0086] Among them, MTP g Indicates the material trigger-propagation index of stage g, A ust,g represents the unstable area in the basin at stage g, V ust,grepresents the volume of unstable material in the basin determined by the TRIGRS model, IC std,g represents the normalized connectivity index of stage g, A wetershed represents the total area of ​​the basin, A unit Denotes the area of ​​the computational unit, D sld,g It represents the failure surface depth of the calculation unit with the minimum safety factor output by the TRIGRS model.

[0087] Furthermore, the indicator system used to predict debris flows includes: MTP indicator, ID threshold indicator, Tsfresh indicator, and ENV indicator. Among them, the MTP indicator is the material trigger-propagation index, the ID threshold indicator includes the average rainfall intensity and rainfall event duration, the Tsfresh indicator is the rainfall driving characteristics extracted from the rainfall sequence through the Tsfresh tool, and the ENV indicator includes key indicators of the geographical background and the effective rainfall before the debris flow outbreak.

[0088] Specifically, to explore the optimal combination of indicators, all indicators were divided into different groups using three schemes, as shown in Table 2. The MTP index refers to the material trigger-propagation (MTP) index; the ID threshold index includes the average rainfall intensity (ARI) and the rainfall event duration (RED); the Tsfresh index refers to rainfall statistical characteristics extracted using the Tsfresh tool, including absolute energy (AE), sum value (SV), approximate entropy (AEMR), continuous wavelet transform coefficient (CWT), autocorrelation coefficient (ATC), upper quantile (CQ1), and lower quantile (CQ2); and the ENV index refers to the geographic background environmental conditions used to assess debris flow susceptibility, including four key indicators of geographic background conditions and the effective rainfall amount (ERA) before the debris flow outbreak. Furthermore, the advantages of incorporating the MTP index were verified by comparing the predictive performance of the hybrid machine learning model in Plan A and Plan B. Similarly, the advantages of using the Tsfresh index over the traditional ID index were revealed by comparing the predictive performance of the hybrid machine learning model in Plan B and Plan C.

[0089] Table 2

[0090]

[0091] Furthermore, building a debris flow prediction model based on the whale optimization algorithm and the machine learning model includes: training the optimal hyperparameters of the machine learning model through the whale optimization algorithm to build a debris flow prediction model, wherein the machine learning model is support vector classification, random forest and extreme gradient boosting.

[0092] Specifically, machine learning (ML) is a complex data-driven method that can predict the spatiotemporal probability of debris flow occurrence by modeling the complex relationships between predictor variables and response variables. Before training a machine learning model, it is usually necessary to manually set some hyperparameters, which significantly limits the model's prediction accuracy and output objectivity. Therefore, this embodiment combines the Whale Optimization Algorithm (WOA) with advanced machine learning algorithms to establish several hybrid machine learning models for predicting debris flows. These algorithms include support vector classification (SVC), random forest (RF), and extreme gradient boosting (XGBoost). The WOA algorithm is a heuristic hyperparameter optimization algorithm that can automatically search for the optimal hyperparameters for training machine learning models and reduce the influence of subjective factors. The above three machine learning algorithms were selected because they have performed well in previous debris flow prediction tasks. In addition, similar results obtained by multiple different machine learning models help verify the reliability of the model results.

[0093] In summary, this example proposes a debris flow prediction method based on the material initiation and propagation process. First, the study establishes an indicator system that combines rainfall triggering conditions with geographic background conditions that characterize debris flow susceptibility. Based on this indicator system, a comprehensive parameter quantification scheme is provided for each indicator. Then, a machine learning model with automatic hyperparameter optimization is used to integrate these indicators to estimate the probability of debris flow occurrence.

[0094] Among the model's input parameters, the rainfall sequence serves as a specific dynamic predictor of all dynamic driving features. This enables automated execution of subsequent debris flow warning tasks, as the model simply extracts the rainfall sequence from radar rainfall products, real-time station data, and predicted precipitation to achieve dynamic debris flow warnings. Furthermore, although different dynamic features (such as rainfall statistics and material triggering and propagation characteristics) have different units, they are all dimensionlessly converted before input into the machine learning model to accelerate convergence and improve robustness.

[0095] The model output is a probability value ranging from 0 (no debris flow) to 1 (debris flow). The index is divided into five levels: very low, low, medium, high, and very high, using cutoff points of 0.2, 0.4, 0.6, and 0.8. These outputs quantify the potential risk of debris flow across five different levels.

[0096] Before model training, there was a significant imbalance between the number of positive examples (debris flow occurrence) and negative examples (no debris flow occurrence) in the samples, indicating a data imbalance problem. To address this issue, oversampling was used to increase the number of positive examples, generating a more balanced dataset and thus improving the robustness of the model. Subsequently, the dataset was randomly split into a training set (70%) for model training and a test set (30%) for model validation. To prevent overfitting and better control model complexity, cross-validation was used. During the sample partitioning and cross-validation process, samples from the same watershed with similar geographical conditions were grouped into Q+1 groups based on the number of debris flow events. Thus, the region was divided into Q sub-regions, the model was trained on Q-1 sub-regions, and validated on the Qth sub-region. This approach is designed to assess the robustness of the model for geospatial problems.

[0097] It's worth noting that while current state-of-the-art machine learning models (such as SVC, RF, and XGBoost models) can be used directly as predictors, this example opts for building hybrid machine learning models because they offer greater adaptability in hyperparameter optimization. Compared to manual parameter tuning during training, machine learning models enhanced with the WOA hyperparameter optimization algorithm have stronger parameter search capabilities.

[0098] The reliability and practicality of the debris flow prediction model were evaluated using receiver operating characteristic (ROC) curves and confusion matrices (CM). Specifically, model performance was estimated using the formula ACC = (TP + TN) / (TP + TN + FP + FN), where ACC represents the prediction accuracy, TP and TN represent the number of test points predicted correctly, and FP and FN represent the number of test points predicted incorrectly. Model quality was assessed by calculating the area under the ROC curve (AUC). A higher AUC value indicates better model prediction accuracy.

[0099] The method of this embodiment is analyzed below based on the target area:

[0100] The target geographical area of ​​this embodiment is located in a fault zone with frequent tectonic activity and dense secondary faults. From a lithological point of view, the main exposed strata include the Proterozoic, Sinian, Cambrian, Silurian, Devonian, Triassic and Quaternary. The study area has a wide range of altitudes, ranging from 804 meters to 5933 meters, and about 66% of the area has a slope of more than 30°. The study area is divided into two natural climate zones: the rainy zone in the south and the semi-arid river valley zone in the north. The average annual rainfall in the central area of ​​the rainy zone is 1235 mm, and the maximum daily rainfall is 270 mm.

[0101] The semi-arid valley region has an average annual rainfall of 526 mm and 80 mm, respectively. The region's rich research achievements and comprehensive disaster records make it an ideal location for debris flow prediction modeling.

[0102] This example uses three parameter design schemes, named Plan A, Plan B, and Plan C, to train three hybrid machine learning (HML) models: WOA-SVC, WOA-RF, and WOA-XGBoost. Model performance is evaluated and quantified using confusion matrices and receiver operating characteristic (ROC) curves. The results demonstrate that the proposed method achieves high prediction accuracy, with ACC scores ranging from 0.815 to 0.917 and AUC values ​​ranging from 0.863 to 0.967. Figures 2(a)-2(c) compare the ROC curves of different hybrid machine learning models under different parameter design schemes.

[0103] This example outlines model improvements from three perspectives. First, the integration of the Whale Optimization Algorithm (WOA) facilitates the automated and precise identification of machine learning model hyperparameters, significantly improving the model's convergence speed and prediction accuracy. Model performance evaluation demonstrates that WOA-optimized machine learning (ML) models outperform standalone ML models. Based on ACC scores, the WOA optimization algorithm improves the performance of SVC, RF, and XGBoost by 5.64%, 3.31%, and 5.58%, respectively.

[0104] Secondly, a hybrid machine learning (HML) model trained with additional rainfall statistical features extracted by the Tsfresh model outperformed a model trained solely on traditional ID threshold features in terms of predictive performance. The Tsfresh tool is used to extract a wider range of statistical features from rainfall sequences to characterize the impact of rainfall triggering mechanisms on debris flow formation. Integrating these rainfall statistical features into the prediction model significantly improved the performance of the HML model. Specifically, under the parameter configuration of Plan C, the performance of the HML model compared to Plan B was improved by 4.25% for SVC, 2.21% for RF, and 2.12% for XGBoost (based on ACC scores). The results show that the rainfall statistical features extracted by the Tsfresh tool have stronger predictive capabilities for debris flow occurrence than traditional ID thresholds.

[0105] The integration of metrics characterizing the material triggering-propagation (MPT) process further enhanced model performance. Incorporating the MPT index into the proposed method significantly improved the performance of the hybrid machine learning (HML) model. Specifically, under the parameter design of Plan B, the HML model improved support vector classification (SVC) by 1.05%, random forest (RF) by 2.26%, and extreme gradient boosting (XGBoost) by 1.01% compared to Plan A (based on ACC scores). These findings highlight the effectiveness of prediction methods that combine rainfall statistics with material triggering-propagation processes, thereby improving debris flow prediction performance by comprehensively considering the essential components of debris flows (i.e., water and material). Notably, the MPT index, because its derivation considers the physical mechanisms of debris flow formation (including the triggering and propagation of unstable materials), has become a key indicator in debris flow monitoring and early warning systems.

[0106] In order to further study the joint decision-making mechanism of rainfall statistical characteristics and material triggering-propagation characteristics on debris flow prediction results, the spatiotemporal probability prediction process of two typical debris flow events was analyzed.

[0107] As shown in Figures 3(a) and 3(b), the material trigger-propagation (MTP) index in target area 1 gradually increased as the rainfall event progressed. Initially, the growth rate was rapid, then slowed, and stabilized approximately two hours before the debris flow onset. The absolute energy (AE) of the rainfall increased exponentially throughout the event, with the rate of increase gradually accelerating, especially during the last two hours, when it experienced an explosive growth. Correspondingly, the dynamically calculated probability index initially remained low (0-0.2) before gradually climbing to high (0.6-0.8) and very high (0.8-1.0) levels approximately two hours before the debris flow onset.

[0108] As shown in Figures 4(a) and 4(b), for target area 2, the AE index, which describes the triggering and propagation of material, exhibited a step-like increase as the rainfall event progressed. Initially, the increase was rapid, then slowed, and then rapidly increased again. It stabilized approximately two hours before the debris flow erupted, then rapidly increased. The MTP index and the AE index exhibited similar growth trends, but did not exhibit a step-like pattern. Ultimately, the MTP index and the AE of the rainfall increased simultaneously, causing the dynamically calculated probability index to climb from a low level (0-0.2) to a high level (0.6-0.8).

[0109] Overall, material trigger-propagation (MTP) and absolute energy (AE) appear to predict debris flow outbreaks in a complementary manner. When MTP conditions reach a bottleneck, higher rainfall absolute energy may also trigger debris flows. Generally speaking, AE, SV, and MTP are all worthy of attention when developing debris flow monitoring and early warning systems to support disaster prevention and control efforts in mountainous areas.

[0110] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A debris flow prediction method based on the material initiation-propagation process, characterized in that: include: Collect debris flow events and rainfall events; According to the debris flow event and the rainfall event, obtaining geographical background conditions and rainfall excitation conditions; Obtaining the rainfall triggering condition includes: Acquiring a rainfall sequence according to the debris flow event and the rainfall event; Extracting features from the rainfall sequence to obtain rainfall features; The TRIGRS model is used to identify the distribution of unstable substances and characterize the safety factor of substance stability, and to calculate the substance trigger-propagation index; Acquiring the rainfall excitation condition according to the rainfall characteristics and the material trigger-propagation index; Constructing an index system for predicting debris flow based on the geographical background conditions and the rainfall excitation conditions; The indicator system used to predict debris flows includes: MTP index, ID threshold index, Tsfresh index, and ENV index. Among them, the MTP index is the material trigger-propagation index, the ID threshold index includes the average rainfall intensity and rainfall event duration, the Tsfresh index is the rainfall driving characteristics extracted from the rainfall sequence using the Tsfresh tool, and the ENV index includes key indicators of the geographical background and the effective rainfall before the debris flow outbreak. The indicators in the indicator system are input into the debris flow prediction model, and the spatiotemporal probability value of the occurrence of debris flow is output, wherein the debris flow prediction model is constructed based on the whale optimization algorithm and the machine learning model.

2. The debris flow prediction method based on material initiation-propagation process according to claim 1 is characterized in that: Obtaining the geographical background conditions includes: selecting key geographical background indicators according to the conditions for the formation of debris flows to obtain the geographical background conditions, wherein the key geographical background indicators include geological and structural indicators, topographic and geomorphological indicators, hydrological and meteorological indicators, and ecological and human indicators.

3. The debris flow prediction method based on material initiation-propagation process according to claim 1 is characterized in that: Extracting features from the rainfall sequence to obtain rainfall features includes: According to the traditional ID threshold curve theory, traditional rainfall statistical features are extracted from the rainfall sequence, wherein the traditional rainfall statistical features include average rainfall intensity and rainfall event duration; Extracting features of the rainfall sequence using the Tsfresh tool to obtain rainfall driving features, wherein the rainfall driving features include absolute energy, sum value, approximate entropy, continuous wavelet transform coefficient, autocorrelation, upper quantile, and lower quantile; The rainfall characteristics are obtained by combining the traditional rainfall statistical characteristics and the rainfall driving characteristics.

4. The debris flow prediction method based on material initiation-propagation process according to claim 3 is characterized in that: Obtaining the average rainfall intensity includes: calculating the average rainfall intensity by dividing the accumulated rainfall by the duration of the rainfall event.

5. The debris flow prediction method based on material initiation-propagation process according to claim 1 is characterized in that: Calculating the substance trigger-propagation index includes: ； ； Among them, MTP g Indicates the material trigger-propagation index of stage g, A ust,g represents the unstable area in the basin at stage g, V ust,g represents the volume of unstable material in the basin determined by the TRIGRS model, IC std,g represents the normalized connectivity index of stage g, A wetershed represents the total area of ​​the basin, A unit Denotes the area of ​​the computational unit, D sld,g It represents the failure surface depth of the calculation unit with the minimum safety factor output by the TRIGRS model.

6. The debris flow prediction method based on material initiation-propagation process according to claim 1, characterized in that: Constructing the debris flow prediction model based on the whale optimization algorithm and the machine learning model includes: training the optimal hyperparameters of the machine learning model through the whale optimization algorithm to construct the debris flow prediction model, wherein the machine learning model is support vector classification, random forest and extreme gradient boosting.

Images