A flood and debris flow risk discrimination method based on fusion mechanism, remote sensing and deep learning

Abstract

The application discloses a mountain torrent and debris flow risk discrimination method fusing mechanism, remote sensing and deep learning, comprising a data fusion and standardization module, a mechanism simulation module, a remote sensing identification module, a deep learning module considering physical constraints and an explainability analysis and closed-loop free module; the method realizes deep fusion and complementary advantages of three types of technologies of mechanism, remote sensing and deep learning by constructing a unified data interface, a unified data model and a physical-data dual driving system and a closed-loop self-learning structure, effectively solves technical problems of data heterogeneity, data intercommunication difficulty, low efficiency, weak model generalization ability, poor explainability and the like in traditional disaster discrimination methods, constructs a self-learning and self-adaptive risk intelligent identification system, and further improves the accuracy, real-time performance and practicality of mountain torrent and debris flow risk identification and discrimination, so as to provide scientific support for risk decision of management departments.

Description

Technical Field

[0001] This invention relates to the field of disaster risk identification and prevention technology, and in particular to a method for identifying flash flood and debris flow risks by integrating mechanisms, remote sensing and deep learning. Background Technology

[0002] Flash floods and debris flows are characterized by their suddenness, destructive power, and wide-ranging impact, seriously threatening people's lives and property and the stability of the ecological environment. Research on disaster risk identification, susceptibility assessment, and risk mapping technologies is crucial for achieving precise disaster prevention, scientific early warning, and efficient response. Currently, relevant technologies both domestically and internationally can be broadly categorized into three types: mechanism simulation, remote sensing identification, and data-driven / deep learning. While each type has played an important role in engineering practice and scientific research, there are still issues of insufficient integration and poor synergy. An integrated intelligent discrimination method is urgently needed to overcome the current technological bottlenecks.

[0003] Mechanism simulation models, centered on watershed hydrology, channel hydraulics, and shallow-water equations for solid-liquid two-phase flow, simulate runoff, sediment transport, and erosion / deposition processes, outputting the impact range of flash floods and debris flows, as well as intensity indicators such as flow velocity, depth, and impact force. Remote sensing identification models, based on multi-temporal optical / SAR change detection, coherence / radiative characteristic analysis, and DEM differencing, achieve post-disaster trace extraction and active area identification, used for historical event cataloging and susceptibility map updates. Data-driven / deep learning models employ random forests, XGboost, U-Net / DeepLab, and ConvLSTM / Transformer to learn the mapping relationship between "topography-rainfall-land cover-historical disaster sites" and risk / occurrence probability, supporting large-scale susceptibility and real-time risk identification.

[0004] In engineering applications, two fusion modes are often used: one is "mechanism → feature → machine learning", which uses intermediate quantities of the mechanism model as prior features and inputs them into the learner in combination with data such as terrain and rainfall; the other is "remote sensing → sample → machine learning", which directly extracts labels from remote sensing data to train the learner.

[0005] Existing technologies and integration models have significant shortcomings that limit application effectiveness: 1. Mechanism simulations are difficult to obtain parameters, highly sensitive to rainfall and physical property errors, and limited in characterizing complex processes such as phase transition triggering and rapid pore pressure rise. Their calibration-verification-iteration chain with remote sensing and big data is weak, making them difficult to adapt to dynamically changing scenarios; 2. Remote sensing identification cannot invert the time-series process of pre-disaster triggering and evolution, is severely affected by cloud and fog, geometric distortion, and coherence, and can only support post-disaster mapping and long-term susceptibility updates, unable to provide pre-disaster probability and real-time early warning; 3. Data... According to the analysis, the interpretability of the driven / deep learning model is lacking, the physical consistency constraint is weak, and it is sensitive to sample bias, imbalance and label noise. It cannot answer the core questions of "why it is triggered" and "how to optimize the mechanism model", and lacks practical guidance value; 4. Fusion mode one is a one-time forward pipeline without feedback loop, and cannot feed the learning results back to the mechanism model parameters and structure, and the fusion depth is insufficient; 5. Fusion mode two focuses on post-event evidence and does not incorporate the temporal triggering process and hydrological conditions, making it difficult to build a short-term identification and early warning link and unable to achieve full-process risk prevention and control. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a method for identifying flash flood and debris flow risks that integrates mechanisms, remote sensing, and deep learning. By constructing a unified data interface, a physical-data dual-driven system, and a closed-loop self-learning structure, it achieves deep integration and complementary advantages of the three technologies. This solves the problems of difficult data interoperability, low efficiency, poor interpretability, and weak generalization ability of traditional technologies, thereby improving the accuracy, real-time performance, and practicality of flash flood and debris flow risk identification and providing scientific support for risk decision-making by management departments.

[0007] A method for assessing the risk of flash floods and debris flows that integrates mechanisms, remote sensing, and deep learning includes the following steps:

[0008] Module 1, Data Fusion and Standardization Module:

[0009] Collect multi-source data, including topographic data, underlying surface data, meteorological and hydrological data, and remote sensing data;

[0010] Unify the spatiotemporal reference, unify the spatial reference of all raster data to the same coordinate system, and unify the temporal reference of all time series data to the same time axis;

[0011] A unified spatiotemporal cube data carrier object, RiskCube, is constructed as a data exchange and fusion carrier, including: a static layer, a dynamic layer, a mechanism output layer, a metadata layer, and a tag layer;

[0012] A spatiotemporal coupling coding rule combining grid coding and time step is adopted to generate a unique spatiotemporal unit ID. All standardized and normalized data are filled into the RiskCube according to the structure of each layer. A data validity mask is generated, including missing measurement areas and quality anomaly areas. For missing measurement areas, the weights are reset to zero. For cloud and fog obstruction, low coherence of synthetic aperture radar SAR, and radar obstruction areas, the quality weights are reduced. After completing the spatiotemporal alignment, unified coding and quality labeling of the full data, the standardized RiskCube is output.

[0013] Module 2, Mechanism Simulation Module:

[0014] The evolution of flash floods and debris flows is simulated through a physical mechanism model. Physical quantities are generated and transformed into weak labels required for deep learning. The input data is based on standardized RiskCube data, including: DEM data and derived topographic parameters, rainfall field, initial water content or pore pressure proxy and physical property parameter set. The output data physical quantities include: flow depth, flow velocity, equivalent driving force index, erosion-related index.

[0015] The set of physical property parameters includes fixed parameters and calibrable parameters; the fixed parameters include at least density parameters, and the calibrable parameters include at least cohesion and yield stress parameters;

[0016] Module 3, Remote Sensing Identification Module:

[0017] The input data is based on standardized RiskCube data, including multi-temporal optical remote sensing data and SAR remote sensing data; optical data is used to extract land cover changes, and SAR data is used to overcome the limitations of cloud and rain weather and capture changes in the post-disaster surface backscattering coefficient; strong labels and corresponding confidence scores are output.

[0018] Based on the standardized remote sensing data, a change feature vector is calculated, and a preliminary active region mask is obtained according to the change feature vector. The preliminary mask is then subjected to morphological filtering, connected component analysis, and channel direction connectivity enhancement to obtain an enhanced active region mask.

[0019] The enhanced active area mask is used as a strong label, which is used to describe the remote sensing post-disaster trace or active area. The strong label confidence score is calculated based on the cloud mask quantization value, SAR coherence value, observation angle, time deviation quality-related parameters, and is used to characterize the reliability of the strong label. The texture features and spectral features of the remote sensing image are extracted to form the surface feature vector, which is written into the feature layer of RiskCube along with the strong label and the strong label confidence score to provide input for the subsequent deep learning module.

[0020] Module 4, Deep Learning Module Considering Physical Constraints:

[0021] The input data is the fused RiskCube feature raster data, including static layer data, dynamic feature sequences, and feature sequences output by mechanism simulation;

[0022] The model is trained by combining the mechanism-constrained loss function, and the output includes a risk probability map and predicted flow depth, predicted flow velocity, and predicted erosion rate.

[0023] Module 5, Interpretability Analysis and Closed-Loop Liberalization Module:

[0024] The interpretability analysis of the prediction results of the learning model is performed, and the feature importance ranking and attribution map are calculated. The feature importance ranking is the ranking result of the importance of each input feature, and the attribution map is the distribution map of the contribution of each input feature to the prediction result.

[0025] Based on the preset feature-model parameter mapping relationship, update the calibrable parameters of the model; rerun the mechanism module under the updated model parameters to obtain new weak labels and verify whether the indicators have improved; if they have improved, save the parameter update; if they have not improved, revert to the previous state and reduce the learning rate; repeat the above steps for multiple iterations, or stop the iteration to generate prediction results.

[0026] Furthermore, in module 1, unifying the spatial reference of all raster data to the same coordinate system specifically includes: the unified spatial reference is determined by the grid range and spatial resolution, the grid range defines the spatial boundary of the computational region, and a unified projected coordinate system is used; the spatial resolution is the size of the regular grid cell; the study region is divided into a regular grid of M rows × N columns according to the grid range and spatial resolution, and continuous value raster data uses bilinear interpolation or cubic convolution interpolation, while discrete category raster data uses nearest neighbor interpolation;

[0027] The process of unifying the time reference of all time series data to the same time axis includes: the unified time reference is determined by the unified time axis and the time step; the time axis is a discrete time series; a fixed time step is used between adjacent moments; rainfall data is accumulated to the time step using area weight or time weight; remote sensing data images are time-aligned to the nearest time series point, and the time deviation is recorded as part of the quality weight.

[0028] Furthermore, in module 1, the set of physical property parameters includes solid density, water density, porosity, internal friction angle, cohesion, yield stress, friction coefficient, saturated hydraulic conductivity, and elastic modulus; the above parameters are based on measured data, observation data, or reference data, and are allowed to be interval values ​​or prior values.

[0029] Furthermore, in module 2, the output data physical quantities also include stacking thickness, the influence range mask of thresholding processing, and the hazard intensity index.

[0030] Furthermore, in module 2, the weak label is calculated using the following formula:

[0031]

[0032] Where H is the flow depth and U is the flow velocity. E is the equivalent driving force index, and E is the erosion-related index. Both b and are calibrable coefficients, and their initial values ​​are empirical values; This is the Sigmoid function.

[0033] Furthermore, in module 3, the change feature vector includes band ratio, multi-temporal shortwave infrared band difference, coherence change, and texture contrast change value.

[0034] Among them, the band ratio is obtained by calculating the band ratio of multispectral surface reflectance, the coherence change is obtained by calculating the correlation coefficient of multi-temporal images, and the texture feature change is obtained by calculating one of the entropy, contrast or homogeneity of the gray-level co-occurrence matrix.

[0035] Furthermore, in module 4, the model training using the mechanistic constraint loss function includes:

[0036] The dual-label hybrid supervision loss is based on the strong label confidence and data validity mask. The binary cross-entropy loss of the strong label region and the weak label region are calculated separately, and the hybrid total loss is obtained by weighted summation. The weight coefficient of the strong label is greater than or equal to the weight coefficient of the weak label.

[0037] Physical consistency constraint loss includes: monotonicity constraint loss, used to constrain the risk probability to increase monotonically with the driving force index; nonnegativity constraint loss, used to constrain the predicted flow depth and predicted erosion rate to be nonnegative; and channel prior constraint loss, used to constrain the risk probability in non-channel areas.

[0038] The total loss function is the weighted sum of the mixed total loss and the losses of each physical consistency constraint.

[0039] Furthermore, in module 5, the updated model calibrable parameters include:

[0040] Calculate the difference between the mean risk probability of strongly labeled regions and the mean risk probability of non-strongly labeled regions;

[0041] The parameter adjustment direction and amplitude are determined based on the difference and the preset direction function;

[0042] Update the calibrable parameters according to the adjustment direction and preset learning rate;

[0043] The updated calibrable parameters are truncated to keep them within the preset parameter threshold range.

[0044] Furthermore, in module 5, the preset feature-model parameter mapping relationship includes: when the feature importance ranking result shows that the equivalent driving force index contributes significantly, and the mechanism model prediction value is lower than the value corresponding to the remote sensing strong label, the friction-related parameter is increased or the roughness term is decreased.

[0045] Furthermore, module 5 also includes: using the flow depth, flow velocity, and erosion rate output by the mechanism module as teacher signals to train the network to learn an approximate mechanism model, forming a mechanism proxy network, which is used to quickly approximate the mechanism model output online, replacing part of the time-consuming numerical solution process.

[0046] The beneficial effects of this invention are as follows: By constructing a unified data interface, a physical-data dual-drive system, and a closed-loop self-learning structure, this invention achieves deep integration and complementary advantages of three types of technologies, solving problems such as difficult data interoperability, low efficiency, poor interpretability, and weak generalization ability of traditional technologies. It improves the accuracy, real-time performance, and practicality of identifying flash flood and debris flow risks, providing scientific support for risk decision-making by management departments, and comprehensively breaking through existing technological bottlenecks, as detailed below:

[0047] 1. Unified data and model interface standard: Construct a standard data interface between mechanism models, remote sensing images, and AI algorithms to completely solve the problems of heterogeneous multi-source data, independent processing, and inefficient interoperability in traditional technologies, realize standard data exchange between the three types of technologies, and lay a solid foundation for deep integration of the entire process.

[0048] 2. Physical-Data Dual-Driven System: Based on mechanistic models such as the shallow water equation, a hybrid identification system is constructed by combining deep learning technology to form a hybrid identification system that combines interpretability and speed, effectively improving the accuracy of risk identification and accelerating response speed.

[0049] 3. Closed-loop self-learning structure: A two-way feedback link between interpretable analysis and mechanistic model is established, and the analysis results are directly fed back to the mechanistic model to realize automatic parameter correction and structural optimization, forming a complete closed loop of interpretation-parameter modification-verification, and continuously improving the rationality and regional adaptability of the model.

[0050] 4. Dynamic Risk Identification and Early Warning Capabilities: Based on real-time meteorological and multi-source remote sensing data updates, minute-level risk probability maps are generated to replace the traditional static, long-cycle input mode; through mechanistic constraints and transfer learning, problems such as poor AI generalization ability and label noise are solved, upgrading remote sensing from post-observation to a dynamic calibration tool, supporting short-term early warning and cross-basin applications, improving prediction accuracy, improving the efficiency of mechanistic simulation, shortening parameter tuning time, and providing interpretable reports that can directly serve management department decision-making and model optimization.

[0051] The present invention will be explained in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0052] Figure 1 This is a flowchart of the flash flood and debris flow risk assessment method that integrates mechanism, remote sensing, and deep learning according to the present invention. Detailed Implementation

[0053] Example 1:

[0054] This embodiment presents a method for identifying the risk of flash floods and debris flows that integrates mechanisms, remote sensing, and deep learning. Figure 1 As shown, by using a unified data model, physical constraint learning, and interpretable closed-loop optimization, a self-learning and adaptive intelligent risk identification system is constructed. This effectively solves the technical problems of data heterogeneity, weak model generalization ability, and poor interpretability in traditional disaster identification, improving the accuracy and real-time performance of flash flood and debris flow disaster risk identification. The specific technical solution is as follows:

[0055] This method consists of five core modules, and the specific implementation of each module is as follows:

[0056] Module 1, Data Fusion and Standardization Module:

[0057] This module is used to achieve unified alignment and standardization of multi-source heterogeneous data, providing a consistent data input basis for all subsequent modules. Specifically, it includes:

[0058] Step 1.1 Collect multi-source input data:

[0059] The collected multi-source input data includes topographic data, underlying surface data, meteorological and hydrological data, and remote sensing data. The topographic data includes digital elevation model (DEM), slope, curvature, runoff accumulation, and channel masking. The underlying surface data includes geological type, soil type, and land cover. The meteorological and hydrological data includes rainfall raster, soil moisture, water level, and flow rate. The rainfall raster can be obtained through radar, station network interpolation, and NWP. The remote sensing data includes multi-temporal optical data and synthetic aperture radar (SAR) feature raster.

[0060] Step 1.2 Unifying the spatiotemporal reference and defining the RiskCube hierarchy:

[0061] A unified spatiotemporal cube data carrier object, RiskCube, is established as the sole data exchange carrier between remote sensing inversion, hydrodynamic mechanism calculation, and machine learning prediction. A unified spatiotemporal benchmark is established as the sole spatial and temporal constraint condition for this data carrier object. All data layers are read, written, interacted, and calculated based on the same set of spatial benchmarks, temporal benchmarks, and data formats, achieving seamless data interoperability among multiple technologies.

[0062] 1. Unified spatiotemporal reference:

[0063] 1.1 Unified Spatial Reference System: Unified Projected Coordinate System EPSG (European Petroleum Survey Group), Resolution R, Grid Range bbox;

[0064] The unified spatial reference is determined by the grid range (bbox) and the spatial resolution (R): the grid range (bbox) defines the spatial boundary of the computational region and adopts a unified projected coordinate system; the spatial resolution (R) is the size of the regular grid cell; the study region is divided into a regular grid of M rows × N columns based on the bbox and R; the RiskCube layer uses this grid to achieve a one-to-one correspondence between spatial positions; the study region is divided into a regular grid based on the bbox and R, and the spatial positions of all data layers correspond one-to-one.

[0065] 1.2 Unified Time Base:

[0066] A unified time base consists of a unified time axis and a time step. Commonly determined: The time axis is a discrete time series. A fixed time step Δt is used between adjacent moments; all dynamic data is stored and interacted synchronously along this time axis; excluding static layers in the RiskCube hierarchy;

[0067] 2. RiskCube hierarchy definition (based on a unified spatiotemporal benchmark):

[0068] 2.1 Static Layer:

[0069] This includes: DEM, slope, curvature, runoff accumulation, channel mask, geological / soil type, land cover, etc.

[0070] 2.2 Dynamic Observation Layer:

[0071] This includes: rainfall gratings (radar / station network interpolation / NWP), soil moisture, water level / flow rate, multi-temporal optical / SAR feature gratings, etc.

[0072] 2.3 Mechanism Output Layer:

[0073] This includes: flow depth, flow velocity, shear force, erosion rate, deposition thickness, and influence range mask.

[0074] 2.4 Metadata Layer:

[0075] This includes: data source, confidence level, cloud mask / coherence mask, missing test markers, quality weights, etc.

[0076] 2.5 Label Layers:

[0077] Including: strong tags (Remote sensing post-disaster traces / activity areas), weak labels (Mechanism risk / intensity classification).

[0078] Step 1.3, Spatiotemporal registration and resampling process:

[0079] 1. Determine the reference grid With reference time axis ;

[0080] 2. All raster data are projected onto EPSG. Continuous raster data (such as DEM, slope, curvature, runoff accumulation, rainfall raster, soil moisture, water level / flow, multi-temporal optical / SAR feature raster) are interpolated using bilinear interpolation or cubic convolution interpolation, while discrete raster data (such as channel masks, geological / soil types, land cover, etc.) are interpolated using nearest neighbor interpolation.

[0081] 3. Align all time series data to :

[0082] Rainfall: accumulated to t (accumulated using area weight or time weight);

[0083] Remote sensing: Image time relative alignment to nearest And record the time deviation. As part of the quality weight;

[0084] Step 1.4, Feature normalization and data quality control:

[0085] The collected data undergoes feature standardization, and continuous variables are standardized using Min-Max normalization or Z-score normalization methods. For discrete category data (such as masks, land use, soil texture, and geological type), their original coding form is preserved, and no normalization is performed.

[0086] Step 1.5, Output:

[0087] A unified spatiotemporal cube data carrier object, RiskCube, is constructed as the data exchange and fusion carrier. A spatiotemporal coupling coding rule combining grid coding and time step is adopted to generate a unique spatiotemporal unit ID. All standardized and normalized data are then filled into the RiskCube according to the structure of each layer. A data validity mask is generated. The data is categorized into two types: areas with missing data (data does not exist) and areas with abnormal quality (data exists but is unreliable). For areas with missing data, the weight is reset to zero; for areas obscured by clouds or fog, with low SAR coherence, or with radar obstruction, the quality weight is reduced.

[0088] For example: Calculate the quality weight ω based on the quality score S (0≤S≤1) for each region: ( (As the baseline weight, with a value of 1.0); where the cloud-obscured area S is directly assigned a value of 0.1, and the SAR coherence C is below the threshold. (e.g., 0.3) (Linear mapping), the radar blocking area S is assigned a value of 0;

[0089] After completing the spatiotemporal alignment, unified encoding, and quality labeling of the full dataset, the final output is a standardized RiskCube, which provides a unified data input carrier for subsequent modules.

[0090] Module 2, Mechanism Simulation Module:

[0091] This module is used to simulate the evolution process of flash floods and debris flows through physical mechanism models, generate physical quantities, and further construct weak labels required for deep learning. Mechanism models suitable for flash flood and debris flow simulation are selected. The mechanism models are based on multiphase flow shallow water equations and sediment-carrying-erosion models or open-source / commercial models (such as EDDA 2.0, FLO-2D, DEBRIS-2D). The core governing equations are based on numerical models of multiphase flow shallow water equations and sediment-carrying-erosion equations, which are used to describe the multiphase flow, sediment transport, erosion and deposition processes of flash floods and debris flows.

[0092] 1. Module inputs and outputs:

[0093] The input data is based on RiskCube data output from the data fusion and standardization module, including:

[0094] 1) DEM and derived topography (slope, confluence, channel masking);

[0095] 2) Rainfall field (Radar / site network / NWP are all acceptable);

[0096] 3) Initial moisture content / pore pressure agent (using empirical initialization);

[0097] 4) Set of physical property parameters (Can be a constant or spatial raster distribution parameter, and can be interval / prior).

[0098] This set of physical property parameters Data sources include measured, observed, or reference data, including solid density. water density The measured data, porosity internal friction angle Cohesion c, Yield stress coefficient of friction saturated hydraulic conductivity elastic modulus Measured data / reference data from literature, etc.

[0099] The output is:

[0100] Current depth; Flow rate; Equivalent driving force index (based on bed shear force); Erosion-related indicators (erosion rate or erosion depth increment); Stacking thickness; Influence range mask (thresholding); Hazard intensity indicators (such as impact force, momentum flux, etc.).

[0101] Among them, parameters Divided into fixed parameters Such as density, and learnable and calibrable parameters. Such as corrosion coefficient, friction, yield stress, roughness, etc. And can be stated in advance. Scope:

[0102]

[0103] 2. Physical weak tag generation rules:

[0104] The risk probability is weakly labeled (0-1), and the calculation formula is as follows:

[0105]

[0106] in, Both b and are calibrable coefficients, and their initial values ​​are empirical values; The sigmoid function maps any real number to the range 0-1, facilitating the conversion of "danger intensity" into "risk probability". A common definition is:

[0107]

[0108] Module 3, Remote Sensing Identification Module:

[0109] This module is used to extract post-disaster traces and activity areas from remote sensing data, and to generate strong labels and surface feature vectors, specifically including:

[0110] Remote sensing strong label generation rules:

[0111] From the RiskCube output by the data fusion and standardization module, multi-temporal optical remote sensing data, SAR remote sensing data, and DEM data that have undergone unified projection and resampling are extracted as inputs. Among them, optical data is used to extract land cover changes, SAR data is used to reduce the impact of cloud and rain obstruction and capture changes in the backscattering coefficient of the land surface after disasters, and DEM data is used to extract the direction of channels to guide subsequent directional morphological processing.

[0112] Input: Multi-temporal optical / SAR and DEM;

[0113] The output is: strong tags +Confidence ;

[0114] By performing radiometric normalization / speckle removal on pre- and post-disaster images, the change feature vector is calculated for the standardized remote sensing images. The calculation formula is as follows:

[0115]

[0116] in, This refers to the band ratio; The difference between multiple temporal shortwave infrared bands; This is a coherence change; This represents the change in texture contrast.

[0117] The band ratio is calculated using multispectral surface reflectance bands. Coherence variations were calculated using correlation coefficients from multi-temporal images. Texture feature changes are calculated using one of the following: entropy, contrast, or homogeneity, obtained from the gray-level co-occurrence matrix.

[0118] Based on the changing feature vectors, a preliminary mask of the active region is obtained through threshold segmentation, machine learning classifiers, or semantic segmentation networks;

[0119] Threshold segmentation can be achieved using fixed thresholds, adaptive thresholds, or Otsu automatic thresholds. Machine learning classifiers include random forests, SVMs, or XGBoost, and semantic segmentation networks include U-Net or DeepLab series networks.

[0120] Morphological opening operations are used to remove isolated noise points and small patches, and connected component analysis is performed to eliminate small noise regions with an area smaller than a set threshold (preferably 50 pixels).

[0121] Based on the spatial extension direction of the channels in the study area (determined by the DEM channel network or the principal axis direction of image PCA), directional linear structural elements are constructed along the channel direction. directional expansion or closure operations are then performed on the active region mask to achieve enhanced connectivity along the channel direction, resulting in the enhanced active region mask. .

[0122] With enhanced active area mask As a strong tag, output (Remote sensing post-disaster traces / activity areas); Calculated based on quality-related parameters such as cloud mask, SAR coherence, observation angle, and time deviation, outputting strong label confidence scores. This is used to characterize the reliability of strong labels; texture and spectral features of remote sensing images are extracted to form surface feature vectors, which are then compared with... It is written into the feature layer of RiskCube to provide input for subsequent deep learning modules.

[0123] Strong label confidence The specific calculation method is as follows:

[0124] First, normalize each parameter to the [0,1] interval, and then use a weighted summation algorithm to calculate... The calculation formula is as follows:

[0125] =0.25× +0.35× +0.2× +0.2×

[0126] The cloud mask quantization value is obtained by processing the original optical image through a cloud detection algorithm (such as the threshold-based Fmask algorithm) to obtain a cloud mask binary image. The value is 1 when the pixel is "cloudless" and 0 when it is "cloudy".

[0127] Time Deviation Quality Quantification Value ,in This represents the absolute number of days between the image acquisition date and the reference date; the closer it is to 1, the smaller the deviation.

[0128] The SAR coherence variation is used to characterize changes in surface structure. This parameter is directly extracted from SAR remote sensing images and obtained after normalization.

[0129] The observation angle quantization value is converted according to a preset threshold. For example, let the sensor side viewing angle be... ,like ,but ;like ,but ;like ,but .

[0130] The value ranges from [0,1], and the higher the value, the stronger the reliability of the strong label.

[0131] Module 4, Deep Learning Module Considering Physical Constraints:

[0132] This module takes the fused feature grid as input and combines it with a mechanism-constrained loss function for model training to achieve real-time risk probability prediction. Specifically, it includes:

[0133] Define the network input and output. The input feature X consists of static features, an observation feature sequence, and a mechanistic feature sequence. The calculation formula is as follows:

[0134]

[0135] Where L is the length of the history window. This represents static layer data; Representing dynamic layer features, it is a time series of length L, containing a sequence of dynamic features obtained from observation or inversion. Representing the mechanistic layer characteristics, within the same time window, the feature sequence is based on the output of the mechanistic simulation module;

[0136] Output risk probability graph and subsequent predicted flow depth Predicting flow velocity Predicting erosion rate It can be used for model distillation or physical constraints.

[0137] Dual-label hybrid supervision loss:

[0138] Strongly labeled supervision (increased confidence level) is calculated using the following formula:

[0139]

[0140] in, For strong label confidence, For strong labels, Cross-entropy, a binary classification term, measures how far the "predicted probability" differs from the "true label." This is the index of the time corresponding to the event.

[0141] Weak label supervision (data validity weighting) is calculated using the following formula:

[0142]

[0143] in, As a data validity mask, This is a weak label for risk probability.

[0144] The total mixed loss (adaptive weights) is calculated using the following formula:

[0145]

[0146] in Both are weighting coefficients, indicating that strong labels have greater weight.

[0147] Physical consistency constraint loss:

[0148] Provide the computable physical constraints, including:

[0149] Monotonicity constraint: The probability of risk increases monotonically with the driving force index, and the calculation formula is as follows:

[0150]

[0151] in, The loss is due to monotonicity constraints. This is a risk probability diagram; As a driving force indicator; The change in risk that violates monotonicity; i represents the change in risk according to the driving force index. The sample index sorted from smallest to largest;

[0152] The output of the proxy physical quantity must be subject to a non-negativity constraint, and the calculation formula is as follows:

[0153]

[0154] in, To predict flow depth, To predict the erosion rate:

[0155] Prior constraints for the channel (risk is more concentrated in the channel / confluence area), the calculation formula is as follows:

[0156]

[0157] in, For trench masking, it means that risks in non-trench areas are penalized. A smaller weighting coefficient is used to prevent overlooking landslides on slopes.

[0158] Total loss function The calculation formula is as follows:

[0159]

[0160] in, For the total mixed loss, The loss is due to monotonicity constraints. Non-negative constraint loss; For the prior constraint loss of the channel, These are the weighting coefficients.

[0161] Module 5, Interpretability Analysis and Closed-Loop Liberalization Module:

[0162] This module is used to analyze the interpretability of model predictions, identify dominant triggering factors, and provide feedback to correct the mechanism model, achieving closed-loop iterative optimization. Specifically, it includes:

[0163] Calculate the importance ranking of all input features (Rank(feature)) to determine the relative importance of each feature to the model's predictions.

[0164] Calculate the contribution of each feature to the prediction result and generate a feature attribution map to intuitively determine the direction and magnitude of feature contribution.

[0165] The interpretability results are compared with the mechanism simulation values ​​and remote sensing inversion values; if the feature contribution is significant but there is a significant deviation in the mechanism or inversion results, the parameter calibration process is triggered.

[0166] For example, when the model interpretive analysis shows that the equivalent driving force index (shear force) τ contributes significantly, but the τ obtained from the mechanism simulation is too small and inconsistent with the remote sensing inversion results, parameter calibration can be performed by increasing the friction-related parameters or decreasing the roughness term.

[0167] The general form of updating parameters is:

[0168]

[0169]

[0170] Wherein, direction function Used to determine the direction of parameter adjustment; the positive or negative sign of the direction determines the increase or decrease. This represents the difference in predicted risk probabilities between strongly labeled and non-strongly labeled regions, used to measure the consistency between model predictions and actual disaster distribution. Parameters that can be calibrated for the model to be updated. The learning rate; For strong labels, This is a risk probability diagram;

[0171] Specifically, a mechanism simulation is performed to obtain weak labels. Physical quantities output by the mechanism module Strong labels obtained by the remote sensing module and its corresponding confidence level Train and update the learning model to obtain The system calculates and interprets the importance ranking of the features in the results, along with an attribution graph. Rank(feature) represents the importance ranking of each input feature, and the attribution graph shows the contribution distribution of each feature to the prediction result. The model parameters are then updated based on parameter mapping and update rules to calibrate the parameters. The mechanism module was re-executed with the updated parameters, resulting in new weak labels. If the verification metric improves, save the parameter update; if it does not improve, revert to the previous state and reset the learning rate. The model is scaled down according to a preset ratio. The above process is repeated for multiple iterations. When the number of iterations reaches the maximum number of iterations N, or when the validation metric shows no improvement for K consecutive rounds, the iteration is terminated, the final model parameters are output, and the model enters the online inference stage to generate real-time risk prediction results.

[0172] The mechanism module performs only computation, while the AI ​​(a deep learning module that considers physical constraints) handles high-frequency inference. Alternatively, the H, U, and E outputs from the mechanism can be used as teacher signals to allow the network to quickly learn approximations, forming a "mechanism proxy network" that can replace some numerical solutions online.

[0173] The above description is only intended to illustrate the technical solutions of the present invention and is not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention (such as the application of various formulas, the order of steps / modules, etc.) without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying the risk of flash floods and debris flows that integrates mechanisms, remote sensing, and deep learning, characterized in that, This method includes the following modules: Module 1, Data Fusion and Standardization Module: Collect multi-source data, including topographic data, underlying surface data, meteorological and hydrological data, and remote sensing data; Unify the spatiotemporal reference, unify the spatial reference of all raster data to the same coordinate system, and unify the temporal reference of all time series data to the same time axis; A unified spatiotemporal cube data carrier object, RiskCube, is constructed as a data exchange and fusion carrier, including: a static layer, a dynamic layer, a mechanism output layer, a metadata layer, and a tag layer; A spatiotemporal coupling coding rule combining grid coding and time step is adopted to generate a unique spatiotemporal unit ID. All standardized and normalized data are filled into the RiskCube according to the structure of each layer. A data validity mask is generated, including missing measurement areas and quality anomaly areas. For missing measurement areas, the weights are reset to zero. For cloud and fog obstruction, low coherence of synthetic aperture radar SAR, and radar obstruction areas, the quality weights are reduced. After completing the spatiotemporal alignment, unified coding and quality labeling of the full data, the standardized RiskCube is output. Module 2, Mechanism Simulation Module: The evolution of flash floods and debris flows is simulated through a physical mechanism model. Physical quantities are generated and transformed into weak labels required for deep learning. The input data is based on standardized RiskCube data, including: DEM data and derived topographic parameters, rainfall field, initial water content or pore pressure proxy and physical property parameter set. The output data physical quantities include: flow depth, flow velocity, equivalent driving force index, erosion-related index. The set of physical property parameters includes fixed parameters and calibrable parameters; the fixed parameters include at least density parameters, and the calibrable parameters include at least cohesion and yield stress parameters; Module 3, Remote Sensing Identification Module: The input data is based on standardized RiskCube data, including multi-temporal optical remote sensing data and SAR remote sensing data; optical data is used to extract land cover changes, and SAR data is used to overcome the limitations of cloud and rain weather and capture changes in the post-disaster surface backscattering coefficient; strong labels and corresponding confidence scores are output. Based on the standardized remote sensing data, a change feature vector is calculated, and a preliminary active region mask is obtained according to the change feature vector. The preliminary mask is then subjected to morphological filtering, connected component analysis, and channel direction connectivity enhancement to obtain an enhanced active region mask. The enhanced active area mask is used as a strong label, which is used to describe the remote sensing post-disaster trace or active area. The strong label confidence score is calculated based on the cloud mask quantization value, SAR coherence value, observation angle, time deviation quality-related parameters, and is used to characterize the reliability of the strong label. The texture features and spectral features of the remote sensing image are extracted to form the surface feature vector, which is written into the feature layer of RiskCube along with the strong label and the strong label confidence score to provide input for the subsequent deep learning module. Module 4, Deep Learning Module Considering Physical Constraints: The input data is the fused RiskCube feature raster data, including static layer data, dynamic feature sequences, and feature sequences output by mechanism simulation; The model is trained by combining the mechanism-constrained loss function, and the output includes a risk probability map and predicted flow depth, predicted flow velocity, and predicted erosion rate. Module 5, Interpretability Analysis and Closed-Loop Liberalization Module: The interpretability analysis of the prediction results of the learning model is performed, and the feature importance ranking and attribution map are calculated. The feature importance ranking is the ranking result of the importance of each input feature, and the attribution map is the distribution map of the contribution of each input feature to the prediction result. Based on the preset feature-model parameter mapping relationship, update the calibrable parameters of the model; rerun the mechanism module under the updated model parameters to obtain new weak labels and verify whether the indicators have improved; if they have improved, save the parameter update; if they have not improved, revert to the previous state and reduce the learning rate; repeat the above steps for multiple iterations, or stop the iteration to generate prediction results.

2. The method for identifying flash flood and debris flow risk by integrating mechanisms, remote sensing, and deep learning as described in claim 1, is characterized in that... In Module 1, unifying the spatial reference of all raster data to the same coordinate system specifically includes: the unified spatial reference is determined by the grid range and spatial resolution, the grid range defines the spatial boundary of the computational region, and a unified projected coordinate system is used; the spatial resolution is the size of the regular grid cell; the study area is divided into a regular grid of M rows × N columns according to the grid range and spatial resolution, continuous value raster data uses bilinear interpolation or cubic convolution interpolation, and discrete category raster data uses nearest neighbor interpolation; The process of unifying the time reference of all time series data to the same time axis includes: the unified time reference is determined by the unified time axis and the time step; the time axis is a discrete time series; a fixed time step is used between adjacent moments; rainfall data is accumulated to the time step using area weight or time weight; remote sensing data images are time-aligned to the nearest time series point, and the time deviation is recorded as part of the quality weight.

3. The method for identifying flash flood and debris flow risk by integrating mechanisms, remote sensing, and deep learning as described in claim 1, is characterized in that... In Module 2, the set of physical property parameters includes solid density, water density, porosity, internal friction angle, cohesion, yield stress, friction coefficient, saturated hydraulic conductivity, and elastic modulus; the above parameters are based on measured data, observation data, or reference data, and are allowed to be interval values ​​or prior values.

4. The method for identifying flash flood and debris flow risk by integrating mechanisms, remote sensing, and deep learning as described in claim 1, is characterized in that... In module 2, the output data physical quantities also include stacking thickness, the influence range mask of thresholding processing, and the hazard intensity index.

5. The method for identifying flash flood and debris flow risk by integrating mechanisms, remote sensing, and deep learning as described in claim 1, characterized in that, In module 2, the weak tag is calculated using the following formula: Where H is the flow depth and U is the flow velocity. E is the equivalent driving force index, and E is the erosion-related index. Both b and are calibrable coefficients, and their initial values ​​are empirical values; This is the Sigmoid function.

6. The method for determining the risk of flash floods and debris flows by integrating mechanisms, remote sensing, and deep learning as described in claim 1, is characterized in that... In module 3, the change feature vector includes band ratio, multi-temporal shortwave infrared band difference, coherence change, and texture contrast change value; Among them, the band ratio is obtained by calculating the band ratio of multispectral surface reflectance, the coherence change is obtained by calculating the correlation coefficient of multi-temporal images, and the texture feature change is obtained by calculating one of the entropy, contrast or homogeneity of the gray-level co-occurrence matrix.

7. The method for determining the risk of flash floods and debris flows by integrating mechanisms, remote sensing, and deep learning as described in claim 1, is characterized in that... In module 4, the model training using the mechanistic constraint loss function includes: The dual-label hybrid supervision loss is based on the strong label confidence and data validity mask. The binary cross-entropy loss of the strong label region and the weak label region are calculated separately, and the hybrid total loss is obtained by weighted summation. The weight coefficient of the strong label is greater than or equal to the weight coefficient of the weak label. Physical consistency constraint loss includes: monotonicity constraint loss, used to constrain the risk probability to increase monotonically with the driving force index; nonnegativity constraint loss, used to constrain the predicted flow depth and predicted erosion rate to be nonnegative; and channel prior constraint loss, used to constrain the risk probability in non-channel areas. The total loss function is the weighted sum of the mixed total loss and the losses of each physical consistency constraint.

8. The method for determining the risk of flash floods and debris flows by integrating mechanisms, remote sensing, and deep learning as described in claim 1, is characterized in that... In module 5, the updated model calibrable parameters include: Calculate the difference between the mean risk probability of strongly labeled regions and the mean risk probability of non-strongly labeled regions; The parameter adjustment direction and amplitude are determined based on the difference and the preset direction function; Update the calibrable parameters according to the adjustment direction and preset learning rate; The updated calibrable parameters are truncated to keep them within the preset parameter threshold range.

9. The method for determining the risk of flash floods and debris flows by integrating mechanisms, remote sensing, and deep learning as described in claim 1, is characterized in that... In module 5, the preset feature-model parameter mapping relationship includes: when the feature importance ranking result shows that the equivalent driving force index contributes significantly and the mechanism model prediction value is lower than the value corresponding to the strong remote sensing label, the friction-related parameter is increased or the roughness term is decreased.

10. The method for determining the risk of flash floods and debris flows by integrating mechanisms, remote sensing, and deep learning as described in claim 1, characterized in that, Module 5 also includes: using the flow depth, flow velocity, and erosion rate output by the mechanism module as teacher signals to train the network to learn an approximate mechanism model, forming a mechanism proxy network, which is used to quickly approximate the mechanism model output online, replacing part of the time-consuming numerical solution process.

Images