A method for probabilistic assessment of the susceptibility of mountain floods and debris flows based on a cascade learning model

Abstract

The present invention discloses a method for evaluating the susceptibility probability of mountain torrents and debris flows based on a cascaded learning model, which relates to the technical field of the distribution evaluation of mountain torrents and debris flows, and includes: calculating index factors according to preprocessed data, where the index factors include hydrological index factors, material index factors, and topographic energy index factors, standardizing and basinizing the index factors to obtain processed index factors; eliminating highly correlated factors or autocorrelated factors in the processed index factors based on the Pearson correlation index to obtain screened data; model training, where model training includes training a preset model according to the screened data, retaining the contributing index factors in this training process based on a recursive feature elimination-cross-validation model to obtain a cascaded dominant learning model, and inputting second information into the cascaded dominant learning model to obtain the evaluation result of the susceptibility probability of mountain torrents and debris flows. The present invention realizes the long-term scale susceptibility evaluation of mountain torrents and debris flows at the basin scale.

Description

Technical Field

[0001] The present invention relates to the technical field of mountain flood and debris flow distribution assessment, and particularly to a method for assessing the susceptibility probability of mountain floods and debris flows based on a cascading learning model. Background Art

[0002] At present, three traditional methods are mainly used to evaluate the distribution of mountain floods and debris flows. The hydrological model method simulates the runoff formation mechanism based on the basin hydrological process. Although it can reflect the water collection situation, it is difficult to obtain various hydrological parameters and is not universal when predicting source-dominated disasters. The ground Internet of Things monitoring system collects and monitors data in real time by deploying a large number of sensors (such as rain gauges, temperature and humidity meters, seismographs, ultrasonic recorders, etc.) to accurately warn of disasters in potential danger areas, but it has high costs, is vulnerable to damage, difficult to maintain, and has limited application scope. The statistical method based on rainfall intensity and daily rainfall establishes a rainfall-disaster correlation model through historical data, which is difficult to reflect complex geological environmental factors and lacks consideration of source conditions, so the evaluation accuracy is limited. Therefore, due to the need for quantitative assessment of the susceptibility of mountain floods and debris flows in large areas of mountainous areas, the remote sensing monitoring method has become quite popular in recent years. This method uses satellite images to dynamically monitor and obtain surface information in real time over a large area. However, in mountainous areas with variable weather, affected by seasons of cloud, fog, rain, snow, and the spatial resolution of images, the image quality is difficult to guarantee, and it can only capture disasters with obvious image features, with a certain degree of prediction lag, and also lacks an assessment of the causes of disasters. Therefore, a method for assessing the susceptibility probability of mountain floods and debris flows based on a cascading learning model is developed to solve the above problems. Summary of the Invention

[0003] The present invention proposes a method for assessing the susceptibility probability of mountain floods and debris flows based on a cascading learning model to solve the problems of parameter heterogeneity characterization, monitoring blind spots, insufficient mechanism-data fusion, and unstable algorithms existing in the assessment of the susceptibility probability of mountain floods and debris flows in the prior art.

[0004] The present invention achieves the above object through the following technical solutions:

[0005] A method for assessing the susceptibility probability of mountain floods and debris flows based on a cascading learning model of the present invention includes:

[0006] Obtaining first information and second information, where the first information includes historical hydrological element data, historical material element data, historical topographic energy element data, and disaster binary classification data, and the second information includes a dataset to be evaluated;

[0007] Preprocess the first information to obtain preprocessed data, calculate index factors based on the preprocessed data, where the index factors include hydrological index factors, material index factors, and topographic energy index factors, standardize and basinize the index factors to obtain processed index factors;

[0008] Based on the Pearson correlation index and preset conditions, eliminate the strongly correlated factors in the processed index factors to obtain screening data;

[0009] Model training, which includes training a preset random forest model according to the screening data, retaining the contributing index factors in this training process based on the recursive feature elimination-cross-validation model to obtain a trained random forest classifier, embedding the random forest classifier into the RFECV framework to obtain the RFECV-RF model, and also includes training a preset XGBoost model according to the prediction results of the RFECV-RF model to obtain a trained XGBoost model;

[0010] Fuse the RFECV-RF model and the XGBoost model, and train the fused model based on the combined feature matrix and disaster binary classification data to obtain a cascaded dominant learning model. Input the second information into the cascaded dominant learning model to obtain the assessment result of the susceptibility probability of mountain torrents and debris flows. The combined feature matrix is obtained by combining the prediction results of the RFECV-RF model and the screening data.

[0011] Furthermore, the hydrological element data includes basic precipitation data, basic temperature data, and basic data of snow depth and snow density. The material element data includes source volume data, source location data, and basic data of material erodibility. The topographic energy element data includes erosion terrain data and basin water convergence data.

[0012] Furthermore, preprocessing the first information to obtain preprocessed data includes:

[0013] Classify and integrate the first information;

[0014] Calibrate the data of the first information.

[0015] Furthermore, calculating index factors based on the preprocessed data includes:

[0016] Calculate the average annual precipitation, the average number of Q95 extreme precipitation days in 10 years, the number of Q95 extreme precipitation times for 2 consecutive days in 10 years, the average Q95 extreme precipitation amount factor in 10 years, the Q95 extreme precipitation amount threshold in 10 years, the number of extreme precipitation days in a certain year, and the number of events where the precipitation for 2 consecutive days exceeds the Q95 extreme precipitation amount threshold in 10 years based on the said basic precipitation data, where Q95 represents the 95% quantile of the data;

[0017] Calculate the average daily temperature, the Q95 extreme high temperature threshold in summer in 10 years, the number of events where the temperature exceeds the Q95 extreme high temperature threshold for 3 consecutive days or more in a certain year, the average number of Q95 extreme high temperature events lasting for 3 consecutive days or more in 10 years, and the average Q95 extreme high temperature factor in 10 years based on the said basic temperature data;

[0018] Calculate the water density, degree-day factor, daily meltwater rate, the extreme meltwater rate caused by extreme high temperature in 10 years on average per day, the number of extreme meltwater times in a certain year, and the average number of extreme meltwater times in 10 years based on the said basic data of snow depth and snow density;

[0019] Calculate the volume of slope-deposited materials caused by weathering and erosion based on the source material volume data;

[0020] Calculate the geodesic distances from the source to the gully and the gully mouth based on the source location data;

[0021] Calculate the soil erodibility based on the basic data of the ease of erosion of the material;

[0022] Calculate the slope, elevation interval, terrain undulation degree, and surface cutting depth based on the erosion terrain data;

[0023] Calculate the terrain humidity index, drainage density, and longitudinal gradient of the main gully based on the catchment area data of the basin.

[0024] Furthermore, standardize and basinize the said index factors to obtain the processed index factors, including:

[0025] Standardize and classify all the said index factors within the value range of 1 - 10;

[0026] Assign values to the standardized and classified index factors according to the average value of all pixel values within the basin.

[0027] Furthermore, eliminate the factors with strong correlation in the processed index factors based on the Pearson correlation index and preset conditions, including:

[0028] Convert the processed index factors into a pandas data structure;

[0029] Use the corr() method of the pandas library to efficiently calculate the Pearson correlation coefficients between all factors;

[0030] Eliminate the highly correlated factors with coefficients > 0.85 in the Pearson correlation coefficient matrix and appear 3 times or more until the lower limit of the number of factors under each factor category is triggered.

[0031] Furthermore, model training includes:

[0032] Construct a random forest model and an XGBoost model;

[0033] Use Bayesian optimization to tune the hyperparameters of the random forest model and the XGBoost model respectively to obtain the tuned random forest model and XGBoost model;

[0034] Process the screened data through pandas and numpy, and perform cross-validation splitting using the cross-validation method to obtain a training set;

[0035] Train the tuned random forest model and XGBoost model respectively according to the training set. The loss function used for training is the binary cross-entropy loss function. Based on the recursive feature elimination-cross-validation model, retain the indicator factors that contribute during the training process to obtain a trained random forest classifier. Embed the random forest classifier into the RFECV framework to obtain the RFECV-RF model;

[0036] Also train the tuned XGBoost model according to the prediction results of the RFECV-RF model to obtain a trained XGBoost model.

[0037] Furthermore, the model generalization formula is as follows:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] .

[0044] It is expressed as the probability assessment result of the susceptibility of flash floods and debris flows. X is the vector of all factors. WS, M, and TP are the hydrological element data, material element data, and terrain energy element data, respectively. X=[{WS1, WS2, …, WSm},{M1, M2, …, Ml},{TP1, TP2, …, TPn}], where m, l, and n are the number of different factors of the three elements. is the predicted value of X after the RFECV screening factor is selected by a single tree in the random forest model. is the final prediction result of the random forest, N is the total number of decision trees, is the prediction result of the RFECV-RF model, K is the number of decision trees in the XGBoost model, It means that the k-th decision tree in XGBoost has a new feature vector The predicted value of Then the sample The weight of the leaf node mapped to the kth tree, where is the mapping function, The result is the weighted sum of all tree subnodes. is the Sigmoid function, where e is the natural logarithm base ≈ 2.718, z and the weight vector And the product of the feature vector X, finally, the prediction model , the prediction results of all decision trees k in XGBoost for the input feature vector X and the global weight of each tree After weighted summation, pass the Sigmoid function Convert to probability prediction values.

[0045] The beneficial effects of the present invention are:

[0046] The probabilistic assessment method for the susceptibility of flash floods and debris flows based on a cascade learning model proposed in the present invention not only constructs a dual-driven model architecture of "physical mechanism constraints + data feature mining", but also provides a more intelligent, low-cost, fast-calculation, watershed unitization, full-domain batching, repeatable operation, extreme climate sensitivity, and consideration of the spatial heterogeneity of source terrain and future climate change solutions. It overcomes the bottlenecks of traditional methods in parameter heterogeneity characterization, monitoring blind spots, insufficient mechanism-data fusion, and insufficient algorithm stability. It solves the problem of model generalization under the conditions of scarce monitoring samples and complex environmental conditions in remote arid mountainous areas, and realizes the long-term scale susceptibility assessment of flash floods and debris flows at the watershed scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the overall process in an embodiment of the present invention.

[0048] Figure 2Schematic diagram of data composition and acquisition in the embodiments of the present invention.

[0049] Figure 3 Flowchart of correlation factor screening in the embodiments of the present invention.

[0050] Figure 4 Flowchart of multi-model construction and training in the embodiments of the present invention.

[0051] Figure 5 Cascade architecture diagram of the advantage learning model in the embodiments of the present invention.

[0052] Figure 6 Flowchart of disaster susceptibility assessment in the embodiments of the present invention. Detailed implementation manners

[0053] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. Generally, the components of the embodiments of the present invention described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations.

[0054] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0055] It should be noted that similar reference numerals and letters denote 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 subsequent drawings.

[0056] "StratifiedKFold" represents the stratified K-fold cross-validation method, the "corr() method" is a method of the data table object in the Pandas library, "numpy" is an open-source numerical calculation extension of Python, and "GPM" and "NCAR" are two important data sources in global-scale hydrological and meteorological research. "pandas.concat" is a function in the Pandas data processing library in Python for splicing multiple data tables, and "ROC-AUC" is one of the classification model evaluation metrics. "Extreme high temperature", "extreme precipitation", and "extreme meltwater" are all specific terms in meteorology, representing air temperature, precipitation, and meltwater that exceed historical extremes or statistical thresholds. "iloc dynamic extraction" means dynamically extracting a data subset according to the numerical positions of rows and columns. "ArcGIS" is a geographic information system software platform.

[0057] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0058] As Figure 1 shown, the overall evaluation process of the present invention includes five main steps: 1. Data acquisition; 2. Data processing and standardization; 3. Relevance factor screening; 4. Machine learning model construction and training; 5. Cascade-type advantage learning model vulnerability assessment.

[0059] A method for evaluating the susceptibility probability of mountain torrents and debris flows based on a cascade learning model according to the present invention includes:

[0060] S1: Obtain first information and second information, where the first information includes historical hydrological element data, historical material element data, historical topographic energy element data, and disaster binary classification data, and the second information includes a dataset to be evaluated, where the dataset to be evaluated includes future hydrological element data, future material element data, and future topographic energy element data;

[0061] The data used in this application includes hydrological, material, and topographic energy environment element data, and the content involved is related to the disaster-forming and disaster-causing conditions of mountain torrents and debris flows, mainly composed of water sources, material sources, and topography that facilitates rapid water convergence and easy erosion. Therefore, obtaining such data is the basis for subsequent mining of the relationships between factors and the environment-disaster relationship. Since the model established later can be used migratably, only the regional environmental data needs to be replaced to be applied to other regions or periods, or the model can be further optimized together with the disaster binary classification data. As Figure 2 shown, specifically refer to the following:

[0062] (1.1) Hydrological element data sources and acquisition: The original basic hydrological element data are all obtained by batch downloading through Python code according to the network protocol requirements through API or FTP. Among them, the basic precipitation data comes from the Global Precipitation Measurement Integrated Multi-Satellite Final Precipitation V07 daily dataset, with a data spatial resolution of 0.1°, level 3B, and the unit is mm. The basic temperature data comes from the reanalysis temperature daily 3-hour and 6-hour datasets of the Meteorological Research Center, including the maximum temperature and the average temperature, with a data spatial resolution of 0.56° and the 6-hourly data with the unit of Kelvin. The basic data of snow depth and snow density come from the dataset of the ERA 5th generation single-layer daily reanalysis processed data, with a spatial resolution of 0.25°, sampled every 6 hours, and the units are m and kg / m3 respectively. This type of data is used for meltwater runoff calculation. Using the above data, relevant parameters such as precipitation type, temperature type, and extreme meltwater type can be calculated subsequently. Similarly, future hydrological data uses the predicted data of daily precipitation, temperature, snow depth, and surface snow amount with a spatial resolution of 0.25° after downscaling and bias correction by the Earth System Model to estimate the meltwater rate.

[0063] (1.2) Data sources and acquisition of material elements: Material elements include three categories: the volume of material sources, the location of material sources, and the erodibility of materials. The basic data is obtained from the network. The material volume data consists of landslide data and slope weathering and erosion data, including the landslide volume and distribution dataset calculated based on empirical formulas and the annual remote sensing image synthesis data extracted through the cloud platform. Due to the dependence on the digital elevation model, the spatial resolution of the digital elevation model data for the material location data is 30m. The basic data of the material erodibility comes from the plateau engineering geological rock group dataset.

[0064] (1.3) Data sources and acquisition of topographic energy elements: It includes two categories: erosional topography and catchment area. The subsequent specific parameters are all extracted from the digital elevation model data and the derived river system basin data obtained above. The river system basin data is obtained through the hydrological analysis tools of the geographic information system, mainly including filling depressions, calculating flow directions, calculating flow accumulation, generating and Strahler-classifying river networks, generating basins, and vectorizing river systems and basins. The main gully is superimposed on the digital elevation model to extract the main gully slope drop and slope gradient, and the unique ID of the basin is set in sequence. At the same time, the terrain-gully mouth extraction method is used to batch extract gully mouths from terrain data and manually correct them.

[0065] (1.4) Disaster binary classification data: Use earth visualization and geospatial analysis tools for visual interpretation and identification of flash flood and debris flow disasters, and save them in the form of ".shp" files as "0 - no", "1 - yes".

[0066] S2: Preprocess the first information to obtain preprocessed data, calculate index factors based on the preprocessed data. The index factors include hydrological index factors, material index factors, and topographic energy index factors. Standardize and basinize the index factors to obtain the processed index factors;

[0067] The data used in this application includes 21 index factors under three major elements. In the present invention, the hydrological element data includes 9 index factors, fully considering the impact of extreme meteorological events on disaster gestation; the material element consists of 5 index factors, considering the volume, location, and potential erosion of material sources; the topographic energy element has 7 index factors. In addition to conventional topographic indexes, new indexes related to the characteristics of channels and basins where flash flood and debris flow disasters are densely developed are added, so as to fully reflect the impact of factors on disasters. Based on the characteristics of multi-source heterogeneous data, the preprocessed and standardized data is particularly important for scientifically and accurately mining the non-linear relationships between factors and reflecting the relationship between disaster-forming factors and disasters. The following mainly elaborates on the data processing and factor construction steps according to the element and factor categories to which the data belongs:

[0068] (2.1) Classify and integrate valid data. Use Python code to classify all hydrological basic data into separate ".parquet" files by looping and pandas.concat according to precipitation, temperature, and snowmelt. This reduces storage costs and speeds up I / O read queries through columnar storage format and efficient compression algorithms. Invalid data such as -69999 and other unrecorded log data are converted to NaN data. If the temperature data is in K, subtract 273.15 and convert it to °C. The temporal resolution of the data is unified to daily units, and the spatial resolution is resampled to 0.05°.

[0069] (2.2) Data calibration. Considering the large spatial differences in mountainous meteorology and the coarse spatial resolution of global data, Python will be used to target the effective data files in (2.1) for precipitation and temperature data. Data from eight national stations in and around the experimental area will be used. Vectorization and parallel computing methods will be used to match regional coordinates of the GPM and NCAR data with lower resolutions, perform inverse distance weighted interpolation to obtain calibration errors, and use the principle of multi-station proximity to make local corrections. After correction, the negative precipitation values will be assigned a value of 0, and the precipitation assignment data in the CMIP6 model will be corrected to 0. The temperature will remain the calibrated result. This will improve the calibration rate and the spatial accuracy and reliability of multi-source data.

[0070] (2.3) Differentiated pipeline calculation of hydrological indicators.

[0071] ① Precipitation calculation: It includes annual average precipitation (PA), 10-year annual average Q95 (95% quantile of data) extreme precipitation days (Q95PYrC), 10-year consecutive 2-day Q95 extreme precipitation number (Q95P2C), and 10-year average Q95 extreme precipitation (Q95PD) factor. To calculate extreme precipitation parameters, it is necessary to first calculate the 10-year Q95 extreme precipitation threshold (Q95Pc) and return the corresponding data row, and then perform detailed calculations on related indicator factors. Among them, P i,y represents the precipitation on the i-th day of the y-th year, dy is the total number of days in the y-th year, and Y tot is the total number of years, PA is in mm / year, N is all precipitation data, is an indicator function and is skipped if it is 0 and is 1 if it is 1. dCounty is the number of days with extreme precipitation in year y. P2County is the number of days with precipitation exceeding Q95P for two consecutive days in year y. c The number of events.

[0072] The details can be calculated in sequence through the following formula:

[0073] ;

[0074] ;

[0075] When it is 1, otherwise it is 0;

[0076] ;

[0077] ;

[0078] ;

[0079] .

[0080] ② Temperature - related calculations: Annual average daily temperature (TAD), number of Q95 summer extreme high - temperature occurrences lasting 3 days or more in a year (Q95T3YC), and average temperature of extreme high - temperature in 10 years (Q95TAgD) factors are set. To calculate extreme high - temperature parameters, it is necessary to first calculate the Q95 summer extreme high - temperature threshold (Q95Tc - summer) in 10 years and return the corresponding data rows, and then perform detailed calculations of related index factors. Among them, Ta (°C) represents the sum of all daily average temperature values, Ms is all the data of the highest temperature in summer, The indicator function is the same as above, T3County (times) is the number of events with temperatures exceeding the Q95 threshold for 3 consecutive days or more in the y - th year, Tmax i…i+2 , and y is the highest temperature (°C) from the i - th to the (i + 2)-th day in the y - th year.

[0081] Specifically, it can be calculated sequentially through the following formulas:

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] .

[0087] ③ Extreme melt - water - related calculations: Annual average extreme high - temperature - induced extreme melt - water rate in 10 years (Q95SRAg) and annual average number of extreme melt - water occurrences in 10 years (Q95SRAC) are set. For the snow - density data in the basic data, data correction is performed according to the snow - depth data. The validity of the snow - density data is determined by the cumulative sum of the snow - depth data of the previous year of the daily snow - density data > 0, and the data that is not > 0 is assigned as NaN. Combining the temperature data corresponding to the matched Q95 extreme high - temperature data, the water density is calculated according to the absolute water - density empirical formula (kg / m 3), and then calculate the degree-day factor (DDF) in units of cm·°C (kg / m 3 ) and the water density data. Combine with the Tmax ‒1 ·day ‒1 data to calculate the daily meltwater rate M. Among them, the constant i…i+2 =999.97358, =3.9818, A = 7.0134e-8, B = 7.926504e-6, C = -7.575677e-8, D = 7.314894e-10, E = -3.596458e-12, =0°C; is the average daily temperature (°C). Then, judge whether the snowmelt rate is effective according to Ta>0°C, otherwise assign M to 0. Then, group by longitude and latitude, calculate the average value of the extreme high-temperature-induced extreme meltwater rate in the past 10 years and count the average annual number of times. Among them, the indicator function is the same as above, Mi,y is the meltwater rate on the i-th day of the y-th year (cm·day ‒1 ), dMCount y is the number of extreme meltwater times in the y-th year (times). Since the future climate data does not include the daily snow density-related data, it is necessary to calculate the snow density by using the available future surface snow cover (kg·m -2 ) and snow depth (m) ratio, so as to complete the calculation of indicators similar to historical data. Specifically, it can be calculated in sequence through the following formulas:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] , when is 1, otherwise it is 0; .

[0093] (4) Calculation of the material index factor for remote sensing terrain fusion.

[0094] ① Calculation of sediment source volume: Production of weathering and erosion data through zonal band extraction and conditional elimination. The sediment source produces weathering and erosion data using TerrSet software and remote sensing data, separates them into individual raster files, imports and converts the data into the ".rst" format, and then uses a classifier to classify the data into three categories of erosion areas, non-erosion areas, and glacier areas based on band characteristics. Through multi-zone differential operations, the weathering and erosion raster data of blocks with obvious spectral differences are obtained and saved as raster files. In ArcGIS, the data is integrated through mosaicking and splicing, and the data is binary classified and the obtained original erosion data is converted into a vector file. By eliminating cultivated land, channels, landslides, and only retaining the slope condition (>20°) and the area exceeding 100m 2 of weathering and erosion data, the final weathering and erosion slope deposit source area is obtained, and the volume V of slope deposits caused by weathering and erosion is calculated using a similar empirical formula erosion , where A is the surface weathering and erosion area. Subsequently, in the factor standardization step, the landslide volume and the weathering and erosion volume will be combined into one, that is, SourceV.

[0095] .

[0096] ② Calculation of sediment source location: Aggregate the spatial distribution vector data of landslides and slope surface weathering and erosion, use python code to create a polygon segmentation tool, input the basin and sediment source data, and segment the data according to the basin boundary to ensure that the disaster contribution sediment source only comes from the directly affiliated basin. Then use the Euclidean distance operation tool in ArcGIS, input the digital elevation model, water system channels, gully mouths, and the processed sediment source vector data, and calculate the geodesic distances from the sediment source to the channels and gully mouths respectively. The smaller the result, the closer the sediment source is to the channel or gully mouth, and the greater the possibility of flowing into the channel to become a potential sediment source. Therefore, in subsequent grading, the larger value will be taken in reverse assignment.

[0097] ③ Inspection of material erodibility data: Consider the erodibility of soil and rock masses respectively to consider the erodibility of potential sediment sources. The calculation of soil erodibility K uses the nomograph formula to obtain factor values, which helps to reflect the sensitivity of soil to erosion and the soil loss rate, and is extremely important for the material supply and transportation of mountain torrents and debris flows. Among them, high values correspond to highly erodible areas, and low values correspond to difficult-to-erode types. The obtained engineering geological rock group data is converted from text to numerical classification, and is assigned high values according to the loose category and low values according to the hard rock type. Check and supplement the data, and use neighborhood values to assign null values to the data to ensure the integrity and effectiveness of the data.

[0098] (5) Calculation of topographic energy index factors based on digital elevation model.

[0099] ① Erosion terrain calculation: Based on the obtained digital elevation model data, use the ArcGIS terrain analysis tool to calculate the slope and elevation interval factors. Through the slope, further extract the effective area ratio, that is, the erosion and disaster-sensitive slope range. Calculate the terrain undulation factor according to a circular moving window with a radius of 16 pixels and a diameter of 33 pixels, which indicates the vulnerability of the terrain to erosion. Use the raster calculator to calculate the average elevation and the lowest elevation difference within the area, and extract the surface cutting depth factor.

[0100] ② Watershed water collection capacity calculation: mainly includes the topographic wetness index (abbreviated as TWI), drainage density (abbreviated as Dd), and longitudinal gradient of the main gully (abbreviated as CG). Among them, TWI is related to the upstream catchment area (A) and slope ( ), reflecting the humidity distribution of the terrain and soil, and the larger the value, the wetter it is. Dd is the ratio of the total cumulative length of the channels (Lu) in the watershed to the watershed area. CG is the average longitudinal gradient (%) of the main gully within the watershed unit, that is, the average of the sum of the ratios of the unit length to the unit drop of the s-curve section of the main gully, which is the average change rate of the elevation per unit length of the gully, and this variable affects the available flow velocity and dynamic potential energy of the substances and liquids in the gully. In summary, the water collection capacity of the watershed can be demonstrated and analyzed from the aspects of watershed humidity distribution, gully gradient, gully distribution, and potential energy, and it is closely related to the hydrological process of mountain torrents and debris flows.

[0101] (6) Factor standardization and basinization. Standardize and classify all the results of the quantified factors according to the value range of 1-10. Classify some of the index factors under substances, terrain energy, and hydrological elements according to the classical classification method. Assign the results of the standardized index factors to the average value of all pixel values in the watershed through ArcGIS, that is, complete the basinization of the index factors. Match and integrate all the factors and the disaster binary classification data according to the unique ID of the watershed and store them as a ".Shp" file for machine learning to call.

[0102] S3: Based on the Pearson correlation index and preset conditions, eliminate the factors with strong correlation in the processed index factors to obtain the screening data;

[0103] Before machine learning in the present invention, due to the large number of factors and the large number of derivative data of the digital elevation model data, it is necessary to pre-eliminate the factors with strong correlation and retain independent factors to ensure that the factors used in the learning model are representative and feature-independent. Therefore, the Pearson correlation index is used to complete efficient screening. As Figure 3 shown, the specific steps are as follows:

[0104] (3.1)Data Preparation. The input data are various factors, presented as a two-dimensional array. Through Python, the original data is converted into a pandas data structure, that is, into a two-dimensional labeled data structure to retain the data structuring characteristics, support flexible data processing, and dynamic factor column name input.

[0105] (3.2)Calculate the correlation coefficient matrix and eliminate highly cross-correlated factors. Use the corr() method of the pandas library to efficiently calculate the Pearson correlation coefficients between all factors. Use the Count function to count, and dynamically extract and eliminate the highly correlated factors with a coefficient > 0.85 and occurring 3 times or more in the Pearson correlation coefficient matrix until the factor quantity lower limit under each element category is triggered (for example, a retention factor threshold of 5 is set for each major category under 22 factors). This reduces the impact of data repeatability interference on machine learning and ensures the comprehensiveness of the data. This method has strong practicality and flexibility. It pre-reduces data redundancy before starting machine learning and can flexibly adjust the input factors, ensuring the reliability, experimental nature, and scalability of the data.

[0106] (3.3)Retain relatively independent factors. The factor data with strong independence that meet the conditions in step (3.2) are centrally stored as a temporary data frame. After checking and verifying without errors, the temporary data is used as the input for the subsequent machine learning model, ensuring a high-quality data foundation for the training data.

[0107] S4: The model training includes training a preset random forest model based on the screened data, retaining the indicator factors that contribute during the training process based on the recursive feature elimination-cross-validation model to obtain a trained random forest classifier, embedding the random forest classifier into the RFECV framework to obtain the RFECV-RF model, and also includes training a preset XGBoost model based on the prediction results of the RFECV-RF model to obtain a trained XGBoost model;

[0108] The present invention adopts machine learning models of multiple architectures, including three types: classical learning models, gradient boosting models, and deep learning models, to ensure the diversity of learning models and obtain the optimal model through comparison. The recursive feature elimination-cross-validation (abbreviated as RFECV) method is combined to rank the factor contributions, so as to screen out non-dominant factors and explore the influence degree of various factors on disasters. Although the learning models used are all relatively mature technologies, in the field of flash flood and debris flow susceptibility assessment, the results obtained by existing research are partially insufficient in stability, the publicly available ROC-AUC values are generally low, and the understanding of factor contributions is weak. In addition, in mountain hydrological process-related disasters, especially in the application of debris flow disaster susceptibility assessment, there are few cases, and there is often a lack of comprehensive consideration of the three major disaster-forming factors or extreme hydrological processes. Therefore, in view of these problems, the present invention uses Python to start from the three major factors and extreme precipitation, high temperature, and meltwater, and uses the means of machine learning with strong self-learning performance and high efficiency, combined with the RFECV algorithm and hyperparameter optimization, to comprehensively and multi-dimensionally evaluate the relationship between each relatively independent factor and flash flood and debris flow in arid areas, make up for the deficiencies of existing technologies, improve the model evaluation accuracy, and make it more suitable for disaster assessment under extreme hydrological conditions of climate change. As Figure 4 shown, the specific initial evaluation model and related technical solutions are as follows:

[0109] (4.1) Import the basic machine learning models. The models imported using Python include random forest (abbreviated as RF), logistic regression (abbreviated as LR) models; the gradient boosting tree models (abbreviated as GBDT) include extreme gradient boosting (abbreviated as XGBoost) and light gradient boosting machine (abbreviated as LightGBM) models; the deep learning model or neural network model uses the deep binary classification optimization model (abbreviated as KBDL-DBCO) and the variable-length long short-term memory network (LSTM) model of the recurrent neural network.

[0110] (4.2) Data input and distribution. Input the result data of feature engineering, process the data through pandas and numpy, and use the cross-validation method such as StratifiedKFold (Fold = 10) for cross-validation splitting to ensure the stable distribution of the training set and the validation set (the splitting ratio of 7:3). The innovation of this step lies in the flexible data conversion for different models, so as to improve the adaptability of the models.

[0111] (4.3)Hyperparameter Optimization. Bayesian Optimization (BO for short) is used to optimize the model hyperparameters, replacing the traditional grid search or random search methods to improve the parameter tuning efficiency. Appropriate hyperparameter ranges are defined for each model, and the parameter performance is evaluated through cross-validation. For deep learning models, the number of hidden layer neurons, learning rate, dropout rate are determined by BO, or the optimal model weights are obtained by setting a dynamic adjustment of the number of training rounds. This method not only maximizes the global search ability of BO but also can find a near-optimal hyperparameter combination in a short time.

[0112] (4.4)Model Training. Different models are trained separately. The training and validation set labels are used for supervised learning, the binary cross-entropy loss (BCE for short) function or the Gini impurity (GI for short) splitting criterion, and cross-validation are used to prevent overfitting and improve the model robustness. This step trains the model through the ".fit()" module. Supervised learning and loss functions are designed for classical and gradient boosting models to prevent model overfitting and enhance model stability. Specific model architectures and training processes are designed for deep learning models (such as KBDL-DBCO and LSTM): for example, the model is designed with a feature input layer, and a hidden layer with two fully connected layers or an LSTM layer, the loss function is set using the Adam optimizer, and an early stopping mechanism is adopted to stop training when the loss reaches the set condition to avoid overfitting. To improve the training efficiency, the module of building a high-performance and reusable data input pipeline in the TensorFlow framework is also used for data loading and batch processing. This step retains the advantages of traditional machine learning models and deep learning models, and is optimized for the disaster-prone probability assessment task, improving the model generalization ability and training efficiency. The model training formula is as follows:

[0113] ;

[0114] where Q is the total number of samples, u v is the true label (0 or 1) of the v-th sample, and ûv is the probability that the model predicts the v-th sample belongs to the positive class (the probability range is 0~1).

[0115] (4.5)Feature Selection. The RFECV algorithm in the sklearn library is used to screen high-dimensional features, only preserving the features that contribute more during the model training and prediction process, dynamically adjusting the number of features and evaluating the model performance. This step closely combines feature selection with model training, significantly reducing data redundancy and improving the model generalization ability and operation efficiency.

[0116] (4.6)Model evaluation and visualization. Evaluate the model performance based on cross-validation and an additional 30% of the test set data, and calculate metrics such as POD, POFD, EI, ROC-AUC, accuracy, and F1 score. Output the model ROC curve, learning curve, and histogram of feature importance rankings. Record the training duration, memory consumption, and various model performance evaluation metrics and save them in a ".xlsx" file. This step introduces multi-dimensional evaluation metrics and uses visualization graphs to display the model results. Especially in the field of flash flood and debris flow susceptibility assessment, these metrics can more comprehensively reflect the performance of the obtained model and intuitively reflect the factor contributions.

[0117] (4.7)Model storage. Save the trained model in the ".joblib" format for direct subsequent invocation and deployment. At the same time, record the space used for storing the model file to reflect the size of the trained model itself. Here, model compression and fast loading technologies are utilized to retain the optimal model weights obtained after the model BO operation, ensuring the efficient storage and fast, accurate deployment of the model.

[0118] S5: Flash flood and debris flow susceptibility assessment of the cascaded dominant learning model: Integrate the RFECV-RF model and the XGBoost model, and train the integrated model based on the combined feature matrix and disaster binary classification data to obtain the cascaded dominant learning model. Input the second information into the cascaded dominant learning model to obtain the flash flood and debris flow susceptibility probability assessment result. The combined feature matrix is obtained by merging the prediction results of the RFECV-RF model and the screened data.

[0119] The present invention automatically screens and returns a dominant learning model based on the machine learning model performance results and uses it to construct a cascaded, relatively stable in operation, and algorithm-optimized integrated model. The architecture of the integrated model is as Figure 5 shown. In view of the uniqueness of the arid area environment, including the increasing trend of the ratio of disasters triggered by extreme meteorological events, the spatial heterogeneity of the source material migration potential, and the terrain erosion and water convergence ability, the present invention for the first time establishes a flash flood and debris flow disaster probability prediction model and susceptibility evaluation system based on factor contribution degree applicable to the arid area environment. In addition, in combination with the diversity of climate change scenarios and the characteristics of temporal changes, through the updated hydrometeorological data, the long-term scale including historical and future disaster susceptibility assessments are completed. As Figure 6 shown, the specific steps are as follows:

[0120] (5.1)ROC-AUC model result comparison. Query and sort the AUC values in the saved model evaluation result ".xlsx" file. After judging the advantages and disadvantages of the model results of the model with features removed by combining RFECV and the basic model results, select the best two groups of models for the next model integration and optimization.

[0121] (5.2) Construction of Cascade Advantage Learning Model. Deploy the trained RFECV-RF and XGBoost models. In a cascaded manner, the prediction results of the former after feature engineering elimination are loaded into the data as new features, and the latter is used to train and predict the prediction results of the RFECV-RF model, completing the feature selection, model training, combined model evaluation, and storage steps described in 4. This method maximally retains the advantages of the dominant machine learning models through a feature adaptive elimination mechanism and a data cascade extraction strategy, enabling efficient fusion training of the models. By means of conjugate optimization and adaptive adjustment methods, a disaster susceptibility assessment result with stable and excellent prediction accuracy can be obtained. This method breaks through the problems of low model accuracy, uneven results of various indicators, and unstable models in disaster susceptibility assessment by traditional machine learning models due to redundant feature selection and insufficient generalization ability of single models. The model generalization formula is as follows. Among them, HSP is the predicted result of susceptibility probability, represented by the function f(X), X is the vector of all factors, WS, M, and TP are hydrological, material, and topographic energy factors respectively, and m, l, and n in X = [{WS1, WS2, …, WSm}, {M1, M2, …, Ml}, {TP1, TP2, …, TPn}] are the numbers of different factors of the three factors. is tree i The predicted value of is the predicted value of a single tree in the random forest for X after RFECV screening factors, is the final prediction result of the random forest, and N is the total number of decision trees; is the newly added feature, and K is the number of decision trees in the XGBoost model, then represents the prediction value of the k-th decision tree in XGBoost for the new feature vector ; then maps the sample to the weight of the leaf node of the k-th tree, where is the mapping function; The result of is the weighted sum of all tree leaf nodes. By internally using the Sigmoid function, the predicted value is converted into a probability value between 0 and 1 as the final prediction result, thereby explaining the probability of the sample belonging to the positive class. Among them, e is the base of the natural logarithm ≈ 2.718, and z is related to the product of the weight vector and the feature vector X. Finally, the prediction model weights and sums the prediction results of all decision trees k in XGBoost for the input feature vector X with the global weight of each tree and then converts it into a probability prediction value through the Sigmoid function

[0122] ;

[0123] ;

[0124] ;

[0125] ;

[0126] ;

[0127] 。

[0128] (5.3) Susceptibility assessment and vector output. Use the trained cascaded dominance learning model to predict the susceptibility probability of mountain flood and debris flow for the assessment dataset, and realize the long-term susceptibility evaluation of future disasters in the mountain flood and debris flow basins in arid areas by replacing the hydrological element factor data in the future scenario. Output the obtained probability results to the "Prob" column and save them in the vector general file format for further analysis using any geographic information system software. This method simplifies the model training and data replacement steps, and combines the prediction results with geospatial data and the attributes of mountain flood and debris flow basin units, facilitating the rapid generation of a visual susceptibility distribution map, providing an intuitive, scientifically spatial and practical evaluation system for the susceptibility assessment of mountain flood and debris flow disasters with the basin as the basic evaluation unit.

[0129] In summary, the present invention is driven by the three disaster-causing factors of water source, material and terrain energy and integrates the multi-source heterogeneous data fusion cascade advantage basic machine learning algorithm. The evaluation system of mountain torrents and debris flows in arid areas not only constructs a dual-driven model architecture of "physical mechanism constraint + data feature mining", but also provides a more intelligent, low-cost, fast calculation, basin unitization, full-domain batching, repeatable operation, extreme climate sensitivity, and consideration of spatial heterogeneity of material source terrain and future climate change. The solution overcomes the bottlenecks of traditional methods in parameter heterogeneity characterization, monitoring blind spots, insufficient mechanism-data fusion, and insufficient algorithm stability, solves the problem of model generalization under the conditions of scarce monitoring samples and complex environment in remote arid mountainous areas, and realizes the long-term scale susceptibility assessment of mountain torrents and debris flows at the basin scale. The running speed of this method only takes 2s in an area of 27,000 square meters, which is an order of magnitude higher than the traditional method, providing a scientific basis for the quantitative assessment of the susceptibility of mountain disasters and prevention and control planning. The present invention is driven by the three major disaster-causing factors of water source, material, and terrain energy, and integrates the cascade-based advantage machine learning algorithm of multi-source heterogeneous data fusion. It constructs the autocorrelation dynamic feature disambiguation and factor elimination mechanism based on Pearson correlation and the three-factor multi-factor importance ranking algorithm driven by machine learning. It establishes a multi-dimensional disaster-pregnancy and disaster-causing main control factor system based on the three major factors of water source, material, and terrain energy potential, supports future long-term disaster prediction and evaluation, and realizes intelligent modeling of flash flood and mudslide susceptibility based on contribution priority. The evaluation system can quickly process a large amount of high-dimensional data, transfer learning across regions, reduce the subjective dependence of manual modeling, fully explore the inherent laws of massive data, break through the limitations of single-factor analysis, save manpower and contact monitoring costs, and has self-learning ability, strong model adaptability, and stable output results, providing a universal solution for the long-term evaluation of the susceptibility of flash flood and mudslide disasters in arid areas.

[0130] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for evaluating the probability of flash flood and debris flow susceptibility based on a cascaded learning model, characterized in that Including: Obtain the first information and the second information, where the first information includes historical hydrological element data, historical material element data, historical topographic energy element data, and disaster binary classification data, and the second information includes the dataset to be evaluated; Preprocess the first information to obtain preprocessed data, calculate index factors based on the preprocessed data, where the index factors include hydrological index factors, material index factors, and topographic energy index factors, and standardize and basinize the index factors to obtain processed index factors; Based on the Pearson correlation index and preset conditions, eliminate the highly correlated factors in the processed index factors to obtain screening data; Model training, where the model training includes training a preset random forest model according to the screening data, retaining the contributing index factors in this training process based on the recursive feature elimination-cross-validation model to obtain a trained random forest classifier, embedding the random forest classifier into the RFECV framework to obtain the RFECV-RF model, and also includes training a preset XGBoost model according to the prediction results of the RFECV-RF model to obtain a trained XGBoost model; Fuse the RFECV-RF model and the XGBoost model, and train the fused model based on the combined feature matrix and the disaster binary classification data to obtain a cascaded dominant learning model, input the second information into the cascaded dominant learning model to obtain the evaluation result of the susceptibility probability of mountain torrents and debris flows, and the combined feature matrix is obtained by combining the prediction results of the RFECV-RF model and the screening data; Based on the Pearson correlation index and preset conditions, eliminating the highly correlated factors in the processed index factors includes: Convert the processed index factors into a pandas data structure; Use the corr() method of the pandas library to efficiently calculate the Pearson correlation coefficients between all factors; Eliminate the highly correlated factors with a coefficient > 0.85 and appearing 3 times or more in the Pearson correlation coefficient matrix until the lower limit of the factor quantity under each element category is triggered; The model training includes: Construct a random forest model and an XGBoost model; Use Bayesian optimization to optimize the hyperparameters of the random forest model and the hyperparameters of the XGBoost model respectively to obtain optimized random forest and XGBoost models; Process the screening data through pandas and numpy, and perform cross-validation splitting using the cross-validation method to obtain a training set; Train the optimized random forest model according to the training set, and the loss function used in the training is the binary cross-entropy loss function. Retain the contributing index factors in this training process based on the recursive feature elimination-cross-validation model to obtain a trained random forest classifier, and embed the random forest classifier into the RFECV framework to obtain the RFECV-RF model; The optimized XGBoost model is further trained according to the prediction result of the RFECV-RF model to obtain a trained XGBoost model.

2. The method for evaluating the probability of mountain flood and debris flow susceptibility based on a cascaded learning model according to claim 1, wherein The hydrological element data includes basic precipitation data, basic temperature data, and basic snow depth and snow density data. The material element data includes source volume data, source location data, and basic material erodibility data. The topographic energy element data includes erosion terrain data and basin catchment data.

3. A method for evaluating the probability of mountain flood and debris flow susceptibility based on a cascaded learning model according to claim 1 or 2, characterized in that, Preprocess the first information to obtain preprocessed data, including: Classify and integrate the first information; Calibrate the data of the first information.

4. A method for evaluating the probability of flash flood and debris flow susceptibility based on a cascaded learning model according to claim 2, characterized in that Calculate index factors according to the preprocessed data, including: Calculate the annual average precipitation, the number of Q95 extreme precipitation days in 10 years, the number of consecutive 2-day Q95 extreme precipitation events in 10 years, the 10-year average Q95 extreme precipitation amount factor, the Q95 extreme precipitation amount threshold in 10 years, the number of extreme precipitation days in a certain year, and the number of events where the precipitation for 2 consecutive days exceeds the Q95 extreme precipitation amount threshold in 10 years based on the basic precipitation data, where Q95 represents the 95% quantile of the data; Calculate the annual average daily temperature, the Q95 extreme high temperature threshold in summer in 10 years, the number of events where the temperature exceeds the Q95 extreme high temperature threshold for 3 consecutive days or more in a certain year, the number of Q95 extreme high temperature events lasting for 3 consecutive days or more in 10 years, and the 10-year extreme high temperature average temperature factor based on the basic temperature data; Calculate the water density, degree-day factor, daily meltwater rate, the 10-year average extreme meltwater rate caused by extreme high temperature, the number of extreme meltwater events in a certain year, and the 10-year annual average number of extreme meltwater events based on the basic snow depth and snow density data; Calculate the volume of slope-deposited material caused by weathering and erosion according to the source volume data; Calculate the geodesic distance from the source to the gully and the gully mouth according to the source location data; Calculate the soil erodibility according to the basic material erodibility data; Calculate the slope, elevation interval, terrain undulation degree, and surface cutting depth according to the erosion terrain data; Calculate the topographic wetness index, drainage density, and longitudinal gradient of the main gully according to the basin catchment data.

5. A method for evaluating the probability of mountain flood and debris flow susceptibility based on a cascaded learning model according to claim 1, characterized in that, Standardize and basinize the index factors to obtain processed index factors, including: Standardize and grade all the index factors within the value range of 1-10; Assign values to the standardized and graded index factors according to the average value of all pixel values within the basin.

6. The method for evaluating the probability of mountain flood and debris flow susceptibility based on a cascaded learning model according to claim 1, wherein The generalization formula of the cascaded dominant learning model is as follows: ； ； ； ； ； ； Denoted as the susceptibility probability assessment result of the mountain flood and debris flow, X is the vector of all factors, WS, M, and TP are hydrological element data, material element data, and terrain energy element data respectively, X = [{WS1, WS2, …, WSm}, {M1, M2, …, Ml}, {TP1, TP2, …, Tpn}], where m, l, n are the numbers of different factors of the three elements. is the predicted value of X after a single tree in the random forest model filters the factors by RFECV. is the final prediction result of the random forest, and N is the total number of decision trees. is the prediction result of the RFECV - RF model, and K is the number of decision trees in the XGBoost model. Then it represents the prediction value of the k - th decision tree in XGBoost for the new feature vector . Then the sample is mapped to the weight of the leaf node of the k - th tree, where is the mapping function. The result of is the weighted sum of all tree leaf nodes. is the Sigmoid function, where e is the base of the natural logarithm ≈ 2.718, z is related to the product of the weight vector and the feature vector X. Finally, the prediction model , after weighted summing the prediction results of all decision trees k in XGBoost for the input feature vector X and the global weight of each tree, is converted into a probability prediction value through the Sigmoid function .

Images