Mountain torrent automatic identification and real-time early warning system based on deep learning

Abstract

The invention discloses a mountain torrent automatic identification and real-time early warning system based on deep learning, and the system comprises a data collection and quality control module which is used for collecting rainfall and topographic data, and generating basic observation data; the rainfall correction and fusion module is used for generating a rasterized rainfall field and uncertainty measurement; the drainage basin topology construction module is used for carrying out concave-pit filling and flat-pit communication on the topographic data and constructing a drainage basin topology; the spatial-temporal feature generation module is used for performing topology alignment on the rasterized precipitation field and the drainage basin to obtain a spatial-temporal feature sequence; the depth prediction module is used for inputting the spatial-temporal feature sequence into a spatial-temporal deep learning model to form a depth prediction result; the hydrodynamic simulation module is used for setting initial and boundary conditions on the unstructured grid and outputting a hydrodynamic simulation result; and the coupling correction and early warning issuing module is used for updating prediction and boundaries, generating graded early warning and completing archiving. According to the invention, automatic mountain torrent identification and real-time early warning are realized.

Description

Technical Field

[0001] This invention relates to the field of emergency early warning management technology, and in particular to a flash flood automatic identification and real-time early warning system based on deep learning. Background Technology

[0002] Currently, most flash flood early warning systems rely on traditional hydrological and hydraulic models and monitoring networks, with data primarily sourced from rain gauges, river level stations, and some remote sensing information. While these systems have achieved some degree of prediction and warning for basin-wide floods, they often face challenges such as data sparsity, low computational efficiency, and delayed response in flash flood scenarios characterized by their suddenness and uneven spatial and temporal distribution. Existing systems typically employ a unidirectional model, calculating runoff from rainfall, then flow and water level, lacking a data-backward correction mechanism, making it difficult to correct prediction errors in real time during operation.

[0003] At the data processing level, existing technologies have significant shortcomings in precipitation fusion and uncertainty quantification. Due to differences in temporal resolution, spatial coverage, and error characteristics among rain gauge, radar, and satellite data, simple linear interpolation or weighted averaging often fails to identify localized abnormal precipitation in a timely manner, thus affecting the accuracy of subsequent hydrological response simulations. Furthermore, some systems use fixed thresholds for flash flood triggering, ignoring the spatiotemporal differences across different watersheds and climatic conditions. This results in warnings that are either overly sensitive and produce false alarms, or delay responses and miss optimal intervention opportunities.

[0004] In hydrodynamic calculations, traditional methods for solving two-dimensional shallow water equations are computationally intensive, and most employ explicit schemes or simple iterative methods, making it difficult to balance accuracy and stability. Furthermore, at the model coupling level, existing technologies typically lack a dynamic fusion mechanism for deep learning model predictions and hydrodynamic simulation results, leading to accumulated biases and boundary inconsistencies between predictions and simulations, and preventing the formation of a feedback correction loop.

[0005] Therefore, how to provide a deep learning-based automatic flash flood identification and real-time early warning system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based automatic flash flood identification and real-time early warning system. This invention constructs a complete link from multi-source data acquisition to intelligent prediction and dynamic early warning, realizing the automation and real-time processing of precipitation information fusion, flash flood evolution process simulation and risk classification release, thereby improving the accuracy of flash flood monitoring and the efficiency of early warning response.

[0007] According to an embodiment of the present invention, a flash flood automatic identification and real-time early warning system based on deep learning includes:

[0008] The data acquisition and quality control module is used to collect precipitation and topographic data, align them with the spatial grid at a uniform time step, and generate basic observation data with quality control markers.

[0009] The precipitation correction and fusion module is used to correct the time-step deviation of multi-source precipitation and generate a rasterized precipitation field with fixed resolution and uncertainty measurement.

[0010] The watershed topology construction module is used to fill depressions and connect flat areas in terrain data, and generate a list of nodes and a list of upstream and downstream connections to form a watershed topology.

[0011] The spatiotemporal feature generation module is used to align the rasterized precipitation field with the watershed topology, summarize precipitation by node coverage area and fill in missing measurements to obtain the spatiotemporal feature sequence of nodes and edges.

[0012] The deep prediction module is used to input spatiotemporal feature sequences into a spatiotemporal deep learning model, train it according to quantile loss and classification loss, and generate deep prediction results.

[0013] The hydrodynamic simulation module is used to set initial and boundary conditions on an unstructured mesh, solve the two-dimensional shallow water equations using finite volume discretization and double conjugate gradient stabilization iteratively, and output hydrodynamic simulation results.

[0014] The coupling correction and early warning release module is used to update the prediction and boundary to obtain the coupling correction results, generate hierarchical early warnings, and complete the archiving.

[0015] Optionally, modules can be integrated using the following methods:

[0016] Multi-source precipitation and topographic data were collected, and quality control and integrity verification were performed to obtain quality-controlled basic observation data.

[0017] Bias correction and spatiotemporal fusion are performed on precipitation data to generate a rasterized precipitation field with fixed spatiotemporal resolution and uncertainty measurement.

[0018] Based on topographic data, watershed partitions and upstream-downstream connectivity are established, and watershed topology is constructed.

[0019] Align the rasterized precipitation field with the watershed topology and generate spatiotemporal feature sequences of nodes and edges at a unified time step;

[0020] Input the spatiotemporal feature sequence into the spatiotemporal deep learning model and output the deep prediction results for the future forecast period;

[0021] Driven by depth prediction results and gridded precipitation fields, a two-dimensional shallow water hydrodynamic simulation is performed in the target area to obtain hydrodynamic simulation results including water depth, flow velocity, peak arrival time, maximum water level and inundation range that evolve over time.

[0022] Dual-drive coupled correction is implemented based on depth prediction and hydrodynamic simulation results as input. The deviations in time and space are calculated and correction values ​​are generated. The depth prediction and simulation boundaries are updated based on the correction values, and the coupled correction results are output.

[0023] The dynamic triggering threshold is calculated based on uncertainty measurement and spatiotemporal characteristic sequence, and compared with the coupling correction result to form a triggering judgment quantity. The triggering probability is reliably calibrated and the decision threshold is determined according to the preset cost ratio. A graded early warning is generated, and the results of continuous time steps are deduplicated and throttled. Early warning information is issued and archived.

[0024] Optionally, obtaining the quality-controlled basic observation data specifically includes:

[0025] Establish a unified time step and spatial grid, determine spatial reference coordinates and target grid resolution, compile a data collection list and create tasks, acquire precipitation data from rain gauges, radar, and satellites, and simultaneously acquire topographic data including basic terrain and river boundaries, recording the source, timestamp, coordinates and integrity markers for each data entry;

[0026] Numerical range and temporal variation range checks were performed on precipitation and topographic data. Consistency checks and multi-source comparison checks were conducted. Anomalies were marked using sliding window statistics. Missing data caused by terrain obstruction or communication failures were weighted and interpolated according to the principles of temporal proximity, spatial proximity, and terrain similarity. Quality control marks were generated for each data point. Numerical range checks, temporal variation checks, and consistency checks were recorded to form a basic data set.

[0027] Precipitation data is converted to a unified coordinate system and resampled to the target grid with a unified time step. Station observations are mapped to the target grid based on distance and elevation differences. Topographic data is aligned to the same spatial grid and output quality-controlled basic observation data.

[0028] Optionally, the formation of the gridded precipitation field specifically includes:

[0029] The system calls up the quality-controlled basic observation data, reads and aligns it to the same timestamp and target grid in sequence according to time steps, performs bias correction on the precipitation data separately, calculates the stability correction coefficient for each source on each grid, obtains the stability correction coefficient by using a sliding window on an event-by-event basis and robust statistics, performs multiplicative correction on the original values ​​step by step, outputs the corrected precipitation data, and retains the correction coefficient and time window information.

[0030] Using corrected precipitation data as input, spatiotemporal fusion is performed on a pixel-by-pixel basis. By taking into account source reliability, observation error level and neighborhood consistency, a rasterized precipitation field with fixed spatial resolution and fixed time step is generated, and the uncertainty metric for each pixel is given.

[0031] Optionally, the construction of the watershed topology specifically includes:

[0032] On a unified spatial grid, load terrain data and align it to the target coordinates and grid resolution. Perform depression filling and flat depression connectivity processing to eliminate local depressions without drainage and false steps, and generate a hydrologically consistent terrain grid.

[0033] Based on the terrain grid, the direction of water flow and upstream source are determined grid by grid. The cumulative amount of confluence is calculated in the order from upstream to downstream. Channel pixels are extracted based on the cumulative amount and continuously tracked to form the river line. The confluence points and outflow points of the river are identified to obtain the river skeleton and the set of key control points.

[0034] Constrained by the river channel framework and key control points, the grid is uniquely assigned to the corresponding control point according to the confluence of the river mouth, and sub-basin partitions are obtained. With sub-basins as nodes and upstream and downstream connections as directions, a node list and an upstream and downstream connection list are generated to construct a sortable watershed topology from upstream to downstream.

[0035] Optionally, the process of generating the spatiotemporal feature sequence specifically includes:

[0036] Using the rasterized precipitation field and uncertainty metric as input and the watershed topology as constraint, spatial alignment is performed on the target raster at a unified time step, establishing the grid-to-node assignment relationship, generating the node coverage area and edge adjacency list, and mapping the quality control marker and uncertainty metric to the corresponding node and edge.

[0037] The coverage area of ​​each node is traversed by time step, and the precipitation values ​​in the area are collected according to the quality control mark and uncertainty measure. Missing locations are filled in according to the principle of time proximity and spatial proximity, and node-level uncertainty is obtained simultaneously, forming a node feature record containing precipitation, state quantity and uncertainty.

[0038] For each directed edge, static attributes including river segment length, elevation difference, and upstream-downstream connectivity are collected. These attributes are then combined with node features from adjacent time steps to generate temporal features of the edge. The sliding window length and step size are set, and the node and edge features from consecutive time steps are packaged into a spatiotemporal feature sequence in a topological order from upstream to downstream.

[0039] Optionally, the output process of the depth prediction result specifically includes:

[0040] Set the step size and number of steps of the sliding window, align the spatiotemporal feature sequence with the nodes and upstream and downstream connection list of the watershed topology, package them into the model input sample set, and record the timestamp, coverage node range and corresponding quality control mark for each sample.

[0041] The model is input into the spatiotemporal deep learning model using the input sample set. Training is performed in batches. The quantile loss function is used to constrain the fitting error of the quantile flow, and the binary cross-entropy loss is used to constrain the classification error of the trigger probability. Iterative updates are performed according to the division of the training set, validation set, and test set until the convergence condition is met on the validation set, and the convergence parameters and model version are obtained.

[0042] At each current time step, the node and edge feature sequences are extracted as model inputs. The convergence parameters are called for inference, and the quantile flow and trigger probability of each node in the future forecast period are output. The results are organized into a depth prediction result according to the time step and node index.

[0043] Optionally, obtaining the hydrodynamic simulation results specifically includes:

[0044] A two-dimensional computational domain is established in the target area and an unstructured triangular mesh is generated. Velocity and water depth are used as basic state variables to construct the governing equations of the two-dimensional hydrodynamic model. The gridded precipitation field is mapped to the surface rainfall drive within the mesh. The depth prediction results are mapped to the upstream inflow boundary time series according to the watershed topology. A fixed water level is set in the downstream area to complete the initial water depth and boundary condition configuration.

[0045] The finite volume discretization method is used to solve the governing equations of a two-dimensional hydrodynamic model. Combined with a biconjugate gradient stabilization iterative method, the linear equations for each time step are solved. By traversing all grid cells and boundaries, the water exchange at each boundary and the source term contribution of each cell are calculated based on the water depth and velocity of the previous time layer. These are then assembled into a sparse coefficient matrix and right-hand side terms. The water depth and velocity of the previous time layer are used as the initial solution. The initial residuals are calculated and the residual norm is recorded. A convergence threshold, a maximum number of iterations, and a divergence protection value are set. In each iteration, two sparse matrix and vector multiplication operations are performed, the residuals are updated, and stabilization correction is executed. If the residuals increase or become unstable, a stabilization step is triggered and the search direction is reset. After each iteration, the magnitude of the residual decrease and the magnitude of the solution vector change are immediately evaluated. If the threshold is met, convergence is determined; otherwise, iteration continues. After convergence, the water depth and velocity of the current time layer are written back to the grid, completing the update for the current time step. The biconjugate gradient stabilization iterative method belongs to the Krylov subspace iterative method and is used for the efficient solution of implicitly discretized two-dimensional shallow water equations.

[0046] After the time progression ends, the water depth sequence of each unit changes over time, the moment when the water depth reaches the extreme value is extracted as the peak arrival time, the maximum water depth of the entire period is recorded as the maximum water level, the flooding range is defined by the area where the water depth exceeds the set threshold, and the hydrodynamic simulation results including the water depth, flow velocity, peak arrival time, maximum water level and flooding range that evolve over time are output.

[0047] Optionally, the output of the coupling correction result specifically includes:

[0048] On a unified time step and spatial grid, the depth prediction results are aligned with the hydrodynamic simulation results. The flow and water depth information of the two at the same time and spatial location are extracted. According to the watershed topology, the water depth field and velocity field in the hydrodynamic simulation are mapped to the node region. The area average of the water depth in each node region is taken, and the difference sequence with the corresponding depth prediction results is calculated. The difference sequence is compared in time order to determine the deviation trend of each node in time, and the time offset of the peak arrival time is identified to generate the initial value of the node-level time deviation.

[0049] Based on the uncertainty measurement results, weight coefficients are assigned to each node and grid point. Time deviation and spatial deviation are weighted to form time correction and spatial correction. The time correction is used to adjust the amplitude and time position of the flow prediction curve of each node. The corrected flow is mapped to the corrected inflow boundary along the upstream and downstream connectivity. The spatial correction is used to locally smooth and adjust the water depth distribution in the hydrodynamic simulation, so that the depth prediction is consistent with the simulation field at the same time step. The trigger probability is finely adjusted according to the direction of the corrected flow change.

[0050] The stopping condition is set to the maximum value of both node time deviation and spatial deviation being lower than the set threshold or reaching the maximum number of iterations. The correction calculation and simulation boundary update are performed iteratively, and the process stops when the stopping condition is met. The coupling correction result is then output.

[0051] Optionally, the process of issuing and archiving early warning information specifically includes:

[0052] The dynamic trigger threshold of each node is calculated by combining the spatiotemporal feature sequence and corrected according to the current uncertainty to form a threshold table;

[0053] The coupling correction results are compared with the dynamic trigger thresholds in the threshold table node by node to generate a trigger judgment quantity. The judgment quantity and trigger probability of each node are comprehensively evaluated, and reliability calibration is performed to maintain the original level of high confidence nodes and reduce the level of low confidence nodes. The node triggering results are classified into different levels and a graded early warning result is generated.

[0054] The system performs deduplication and throttling on the graded early warning results, retaining the time steps with significant changes and limiting the minimum release interval. The filtered graded early warning results are organized into records according to time sequence and spatial location. The system outputs structured information including trigger nodes, early warning levels, confidence levels, and impact ranges, and writes it synchronously into the archived database, thus completing the early warning release process.

[0055] The beneficial effects of this invention are:

[0056] 1. This invention establishes a highly consistent and accurate basic data system through unified acquisition and quality control of multi-source precipitation data and topographic data, which solves the problems of heterogeneous data sources, large accuracy differences and serious missing data in the existing system, and provides reliable input for subsequent modeling.

[0057] 2. This invention introduces a deviation correction and spatiotemporal fusion mechanism in the precipitation processing stage, which can dynamically adjust the error weights between different data sources, realize the quantification of uncertainty in the gridded precipitation field, and make the input data more consistent with the actual rainfall distribution characteristics, thereby improving the realism and spatial precision of rainfall-driven processes.

[0058] 3. This invention, by constructing a watershed topology and node connection relationship based on terrain, realizes automatic identification of water flow paths and sub-watershed division under complex mountainous terrain conditions, providing structured support for the dynamic tracking of flood propagation paths.

[0059] 4. This invention innovatively proposes a dual-drive coupling correction mechanism for depth prediction and hydrodynamic simulation, realizing the bidirectional integration of data-driven and physical-driven approaches. By correcting temporal and spatial deviations in real time, the prediction results are dynamically self-corrected during continuous operation, ensuring the continuity and accuracy of the forecast. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of a deep learning-based automatic flash flood identification and real-time early warning system proposed in this invention;

[0062] Figure 2 This is a schematic diagram of the algorithm structure of a deep learning-based automatic flash flood identification and real-time early warning system proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figures 1-2 A deep learning-based automatic flash flood identification and real-time early warning system includes:

[0065] The data acquisition and quality control module is used to collect precipitation and topographic data, align them with the spatial grid at a uniform time step, and generate basic observation data with quality control markers.

[0066] The precipitation correction and fusion module is used to correct the time-step deviation of multi-source precipitation and generate a rasterized precipitation field with fixed resolution and uncertainty measurement.

[0067] The watershed topology construction module is used to fill depressions and connect flat areas in terrain data, and generate a list of nodes and a list of upstream and downstream connections to form a watershed topology.

[0068] The spatiotemporal feature generation module is used to align the rasterized precipitation field with the watershed topology, summarize precipitation by node coverage area and fill in missing measurements to obtain the spatiotemporal feature sequence of nodes and edges.

[0069] The deep prediction module is used to input spatiotemporal feature sequences into a spatiotemporal deep learning model, train it according to quantile loss and classification loss, and generate deep prediction results.

[0070] The hydrodynamic simulation module is used to set initial and boundary conditions on an unstructured mesh, solve the two-dimensional shallow water equations using finite volume discretization and double conjugate gradient stabilization iteratively, and output hydrodynamic simulation results.

[0071] The coupling correction and early warning release module is used to update the prediction and boundary to obtain the coupling correction results, generate hierarchical early warnings, and complete the archiving.

[0072] This invention constructs an automatic flash flood identification and real-time early warning system that integrates deep learning and hydrodynamic simulation, forming a fully automated chain from data acquisition, precipitation correction, watershed topology construction to prediction, simulation and early warning release. It realizes intelligent modeling and dynamic correction of flood evolution under multi-source precipitation information and complex terrain conditions, significantly improving the accuracy and timeliness of flash flood monitoring. Furthermore, it achieves high reliability and interpretability of early warning results through a dual-drive coupling correction and uncertainty threshold determination mechanism.

[0073] In this embodiment, the modules are interconnected using the following method:

[0074] Multi-source precipitation and topographic data were collected, and quality control and integrity verification were performed to obtain quality-controlled basic observation data.

[0075] Bias correction and spatiotemporal fusion are performed on precipitation data to generate a rasterized precipitation field with fixed spatiotemporal resolution and uncertainty measurement.

[0076] Based on topographic data, watershed partitions and upstream-downstream connectivity are established, and watershed topology is constructed.

[0077] Align the rasterized precipitation field with the watershed topology and generate spatiotemporal feature sequences of nodes and edges at a unified time step;

[0078] Input the spatiotemporal feature sequence into the spatiotemporal deep learning model and output the deep prediction results for the future forecast period;

[0079] Driven by depth prediction results and gridded precipitation fields, a two-dimensional shallow water hydrodynamic simulation is performed in the target area to obtain hydrodynamic simulation results including water depth, flow velocity, peak arrival time, maximum water level and inundation range that evolve over time.

[0080] Dual-drive coupled correction is implemented based on depth prediction and hydrodynamic simulation results as input. The deviations in time and space are calculated and correction values ​​are generated. The depth prediction and simulation boundaries are updated based on the correction values, and the coupled correction results are output.

[0081] The dynamic triggering threshold is calculated based on uncertainty measurement and spatiotemporal characteristic sequence, and compared with the coupling correction result to form a triggering judgment quantity. The triggering probability is reliably calibrated and the decision threshold is determined according to the preset cost ratio. A graded early warning is generated, and the results of continuous time steps are deduplicated and throttled. Early warning information is issued and archived.

[0082] This invention constructs a deep learning-based automatic flash flood identification and real-time early warning system, establishing a fully intelligent process from data acquisition, precipitation correction, watershed modeling to prediction, simulation, and early warning dissemination. It achieves the organic integration of data-driven and physical modeling, enabling dynamic identification of flood formation and propagation processes under complex terrain and multi-source precipitation conditions. This effectively reduces false alarm and missed alarm rates, improves the timeliness and accuracy of flash flood forecasts, and significantly enhances the intelligent monitoring and emergency response capabilities for mountain flood disasters.

[0083] In this embodiment, obtaining the quality-controlled basic observation data specifically includes:

[0084] Establish a unified time step and spatial grid, determine spatial reference coordinates and target grid resolution, compile a data collection list and create tasks, acquire precipitation data from rain gauges, radar, and satellites, and simultaneously acquire topographic data including basic terrain and river boundaries, recording the source, timestamp, coordinates and integrity markers for each data entry;

[0085] Numerical range and temporal variation range checks were performed on precipitation and topographic data. Consistency checks and multi-source comparison checks were conducted. Anomalies were marked using sliding window statistics. Missing data caused by terrain obstruction or communication failures were weighted and interpolated according to the principles of temporal proximity, spatial proximity, and terrain similarity. Quality control marks were generated for each data point. Numerical range checks, temporal variation checks, and consistency checks were recorded to form a basic data set.

[0086] Precipitation data is converted to a unified coordinate system and resampled to the target grid with a unified time step. Station observations are mapped to the target grid based on distance and elevation differences. Topographic data is aligned to the same spatial grid and output quality-controlled basic observation data.

[0087] This invention establishes a unified time step and spatial grid, enabling standardized integration and quality control of multi-source precipitation and topographic data. It effectively solves the problems of spatiotemporal resolution differences and missing data between different data sources, ensuring the integrity and consistency of input data and providing reliable basic data support for subsequent modeling and simulation.

[0088] In this embodiment, the formation of the gridded precipitation field specifically includes:

[0089] The system calls up the quality-controlled basic observation data, reads and aligns it to the same timestamp and target grid in sequence according to time steps, performs bias correction on the precipitation data separately, calculates the stability correction coefficient for each source on each grid, obtains the stability correction coefficient by using a sliding window on an event-by-event basis and robust statistics, performs multiplicative correction on the original values ​​step by step, outputs the corrected precipitation data, and retains the correction coefficient and time window information.

[0090] Using corrected precipitation data as input, spatiotemporal fusion is performed on a pixel-by-pixel basis. By taking into account source reliability, observation error level and neighborhood consistency, a rasterized precipitation field with fixed spatial resolution and fixed time step is generated, and the uncertainty metric for each pixel is given.

[0091] This invention implements bias correction and spatiotemporal fusion on multi-source precipitation data, integrates the credibility and observation errors of different data sources, effectively eliminates systematic biases and improves data consistency, generates high-precision rasterized precipitation fields and uncertainty quantification results, thereby enhancing the realism and spatial precision of rainfall-driven events.

[0092] In this embodiment, the construction of the watershed topology specifically includes:

[0093] On a unified spatial grid, load terrain data and align it to the target coordinates and grid resolution. Perform depression filling and flat depression connectivity processing to eliminate local depressions without drainage and false steps, and generate a hydrologically consistent terrain grid.

[0094] Based on the terrain grid, the direction of water flow and upstream source are determined grid by grid. The cumulative amount of confluence is calculated in the order from upstream to downstream. Channel pixels are extracted based on the cumulative amount and continuously tracked to form the river line. The confluence points and outflow points of the river are identified to obtain the river skeleton and the set of key control points.

[0095] Constrained by the river channel framework and key control points, the grid is uniquely assigned to the corresponding control point according to the confluence of the river mouth, and sub-basin partitions are obtained. With sub-basins as nodes and upstream and downstream connections as directions, a node list and an upstream and downstream connection list are generated to construct a sortable watershed topology from upstream to downstream.

[0096] This invention achieves hydrological consistency of the terrain raster by filling depressions and connecting flat areas in the terrain data, automatically identifies the river skeleton, confluence points and watershed boundaries, and constructs a top-down sortable watershed topology structure that can accurately describe the water flow convergence path and upstream-downstream relationships.

[0097] In this embodiment, the process of generating the spatiotemporal feature sequence specifically includes:

[0098] Using the rasterized precipitation field and uncertainty metric as input and the watershed topology as constraint, spatial alignment is performed on the target raster at a unified time step, establishing the grid-to-node assignment relationship, generating the node coverage area and edge adjacency list, and mapping the quality control marker and uncertainty metric to the corresponding node and edge.

[0099] The coverage area of ​​each node is traversed by time step, and the precipitation values ​​in the area are collected according to the quality control mark and uncertainty measure. Missing locations are filled in according to the principle of time proximity and spatial proximity, and node-level uncertainty is obtained simultaneously, forming a node feature record containing precipitation, state quantity and uncertainty.

[0100] For each directed edge, static attributes including river segment length, elevation difference, and upstream-downstream connectivity are collected. These attributes are then combined with node features from adjacent time steps to generate temporal features of the edge. The sliding window length and step size are set, and the node and edge features from consecutive time steps are packaged into a spatiotemporal feature sequence in a topological order from upstream to downstream.

[0101] This invention aligns the rasterized precipitation field with the watershed topology, establishes the correspondence between nodes and edges, and generates a spatiotemporal feature sequence. This enables a unified expression of rainfall, topography, and water flow connection information, and can fully characterize the watershed hydrological response process in both time and space dimensions. It provides high-quality dynamic input data for deep learning models, thereby improving the spatiotemporal correlation and accuracy of flood evolution prediction.

[0102] In this embodiment, the output process of the depth prediction result specifically includes:

[0103] Set the step size and number of steps of the sliding window, align the spatiotemporal feature sequence with the nodes and upstream and downstream connection list of the watershed topology, package them into the model input sample set, and record the timestamp, coverage node range and corresponding quality control mark for each sample.

[0104] The model is input into the spatiotemporal deep learning model using the input sample set. Training is performed in batches. The quantile loss function is used to constrain the fitting error of the quantile flow, and the binary cross-entropy loss is used to constrain the classification error of the trigger probability. Iterative updates are performed according to the division of the training set, validation set, and test set until the convergence condition is met on the validation set, and the convergence parameters and model version are obtained.

[0105] At each current time step, the node and edge feature sequences are extracted as model inputs. The convergence parameters are called for inference, and the quantile flow and trigger probability of each node in the future forecast period are output. The results are organized into a depth prediction result according to the time step and node index.

[0106] This invention constructs a spatiotemporal deep learning model that organically combines watershed topology and node features. By utilizing a joint optimization strategy of quantile loss and classification loss, it achieves accurate modeling of flash flood evolution and multi-quantile prediction of future flow. This effectively characterizes nonlinear spatiotemporal dependencies and improves the model's generalization and convergence performance, thereby enhancing the accuracy of flood trigger probability judgment and forecast results.

[0107] In this embodiment, obtaining the hydrodynamic simulation results specifically includes:

[0108] A two-dimensional computational domain is established in the target area and an unstructured triangular mesh is generated. Velocity and water depth are used as basic state variables to construct the governing equations of the two-dimensional hydrodynamic model. The gridded precipitation field is mapped to the surface rainfall drive within the mesh. The depth prediction results are mapped to the upstream inflow boundary time series according to the watershed topology. A fixed water level is set in the downstream area to complete the initial water depth and boundary condition configuration.

[0109] The finite volume discretization method is used to solve the governing equations of a two-dimensional hydrodynamic model. Combined with a biconjugate gradient stabilization iterative method, the linear equations for each time step are solved. By traversing all grid cells and boundaries, the water exchange at each boundary and the source term contribution of each cell are calculated based on the water depth and velocity of the previous time layer. These are then assembled into a sparse coefficient matrix and right-hand side terms. The water depth and velocity of the previous time layer are used as the initial solution. The initial residuals are calculated and the residual norm is recorded. A convergence threshold, a maximum number of iterations, and a divergence protection value are set. In each iteration, two sparse matrix and vector multiplication operations are performed, the residuals are updated, and stabilization correction is executed. If the residuals increase or become unstable, a stabilization step is triggered and the search direction is reset. After each iteration, the magnitude of the residual decrease and the magnitude of the solution vector change are immediately evaluated. If the threshold is met, convergence is determined; otherwise, iteration continues. After convergence, the water depth and velocity of the current time layer are written back to the grid, completing the update for the current time step. The biconjugate gradient stabilization iterative method belongs to the Krylov subspace iterative method and is used for the efficient solution of implicitly discretized two-dimensional shallow water equations.

[0110] After the time progression ends, the water depth sequence of each unit changes over time, the moment when the water depth reaches the extreme value is extracted as the peak arrival time, the maximum water depth of the entire period is recorded as the maximum water level, the flooding range is defined by the area where the water depth exceeds the set threshold, and the hydrodynamic simulation results including the water depth, flow velocity, peak arrival time, maximum water level and flooding range that evolve over time are output.

[0111] This invention achieves efficient simulation of mountain flood dynamics under complex terrain conditions by establishing an unstructured grid in the target area and using a finite volume discretization and biconjugate gradient stabilization iterative method to solve the two-dimensional shallow water equations. It can stably solve large-scale sparse equation sets while maintaining computational accuracy, and quickly generate water depth, flow velocity and inundation range results that evolve over time, thereby improving the stability and real-time performance of mountain flood evolution simulation.

[0112] In this embodiment, the output of the coupling correction result specifically includes:

[0113] On a unified time step and spatial grid, the depth prediction results are aligned with the hydrodynamic simulation results. The flow and water depth information of the two at the same time and spatial location are extracted. According to the watershed topology, the water depth field and velocity field in the hydrodynamic simulation are mapped to the node region. The area average of the water depth in each node region is taken, and the difference sequence with the corresponding depth prediction results is calculated. The difference sequence is compared in time order to determine the deviation trend of each node in time, and the time offset of the peak arrival time is identified to generate the initial value of the node-level time deviation.

[0114] Based on the uncertainty measurement results, weight coefficients are assigned to each node and grid point. Time deviation and spatial deviation are weighted to form time correction and spatial correction. The time correction is used to adjust the amplitude and time position of the flow prediction curve of each node. The corrected flow is mapped to the corrected inflow boundary along the upstream and downstream connectivity. The spatial correction is used to locally smooth and adjust the water depth distribution in the hydrodynamic simulation, so that the depth prediction is consistent with the simulation field at the same time step. The trigger probability is finely adjusted according to the direction of the corrected flow change.

[0115] The stopping condition is set to the maximum value of both node time deviation and spatial deviation being lower than the set threshold or reaching the maximum number of iterations. The correction calculation and simulation boundary update are performed iteratively, and the process stops when the stopping condition is met. The coupling correction result is then output.

[0116] This invention introduces a dual-drive coupling correction mechanism based on depth prediction and hydrodynamic simulation results, which can dynamically calculate and correct deviations in the time and space dimensions, realize the real-time fusion of data-driven models and physical simulation models, effectively eliminate inconsistencies between prediction and simulation, improve the stability and accuracy of results, and enhance the adaptability and reliability of flash flood process simulation.

[0117] In this embodiment, the process of issuing and archiving early warning information specifically includes:

[0118] The dynamic trigger threshold of each node is calculated by combining the spatiotemporal feature sequence and corrected according to the current uncertainty to form a threshold table;

[0119] The coupling correction results are compared with the dynamic trigger thresholds in the threshold table node by node to generate a trigger judgment quantity. The judgment quantity and trigger probability of each node are comprehensively evaluated, and reliability calibration is performed to maintain the original level of high confidence nodes and reduce the level of low confidence nodes. The node triggering results are classified into different levels and a graded early warning result is generated.

[0120] The system performs deduplication and throttling on the graded early warning results, retaining the time steps with significant changes and limiting the minimum release interval. The filtered graded early warning results are organized into records according to time sequence and spatial location. The system outputs structured information including trigger nodes, early warning levels, confidence levels, and impact ranges, and writes it synchronously into the archived database, thus completing the early warning release process.

[0121] This invention introduces a dynamic threshold calculation and reliability calibration mechanism based on uncertainty measurement, which can automatically adjust the triggering conditions for risk changes in different watershed nodes and achieve multi-dimensional comprehensive judgment by combining the coupling correction results. This improves the accuracy and stability of flash flood warning classification. At the same time, by deduplicating the results and optimizing the information release frequency through flow control, it ensures that the warning response is efficient, reliable and traceable.

[0122] Example 1:

[0123] To verify the feasibility of this invention in practice, it was applied to a flood control monitoring and early warning project in a typical flash flood-prone watershed. This watershed has a highly undulating topography, deep valleys, concentrated rainfall, and low surface infiltration, making it extremely prone to short-duration, high-intensity flash floods. Traditional threshold methods rely on experience to set fixed rainfall or flow thresholds, triggering an early warning when monitoring data exceeds these thresholds. While this method is simple in structure, it lacks dynamic adaptability to topography, spatial differences in rainfall, and watershed response characteristics, often leading to false alarms or missed alarms under severe convective weather or localized torrential rain conditions.

[0124] During the application process, precipitation information from multiple sources, including regional rain gauges, weather radar, and satellite precipitation data, was collected and integrated, while high-resolution topographic data was acquired simultaneously. After quality control and consistency checks, a basic dataset with a unified time step and spatial resolution was formed.

[0125] Based on unified topographic data, the river network and sub-regional boundaries are automatically extracted through flow direction calculation and confluence accumulation analysis, forming a watershed topology covering 28 sub-basins and containing 122 nodes, providing physical support for subsequent modeling and simulation. Subsequently, the precipitation field is aligned with the topology to generate spatiotemporal feature sequences of nodes and edges, which are then input into a spatiotemporal deep learning model for training and prediction. This model employs a structure combining graph convolution and time series networks, effectively capturing the propagation effects of rainfall in the spatial and temporal dimensions of the watershed.

[0126] To further improve the physical plausibility of the prediction results, the depth prediction output and the precipitation driving field were used as inputs to perform a two-dimensional shallow water equation hydrodynamic simulation. The simulation was solved using the finite volume discretization method and the double conjugate gradient stabilization iterative algorithm. The computational domain was divided into an irregular triangular mesh, totaling approximately 68,000 elements.

[0127] During model operation, depth prediction results and hydrodynamic simulation results are fused through a coupled correction module. The system automatically calculates temporal and spatial deviations and generates correction values, synchronously updating the prediction and simulation boundaries. The system automatically calculates dynamic trigger thresholds using uncertainty metrics and determines the optimal warning level based on the trigger probability-cost ratio. To prevent frequent fluctuations in warning information, the system performs deduplication and throttling processes between consecutive time steps, outputting updated results only during critical changes. Specific experimental comparison data are shown in Table 1.

[0128] Table 1 Comparative Experiment Results

[0129]

[0130] As shown in Table 1, the coupling mechanism between deep learning and the hydrodynamic model significantly improves the accuracy of flow and depth prediction, increasing the Nash efficiency coefficient for peak flow prediction from 0.79 to 0.91 and reducing the peak time error from 55 minutes to 33 minutes. Regarding spatial simulation accuracy, the root mean square error of depth decreased from 0.38 meters to 0.22 meters, and the mean absolute error decreased from 12.6 m³ / s to 7.8 m³ / s. Computational efficiency also improved significantly, with the single-step simulation time reduced from 4.6 seconds to 2.6 seconds. In terms of early warning performance, the system's average advance warning time increased to 65 minutes, approximately 160% higher than the traditional empirical method, while the false alarm rate and missed alarm rate decreased to 12% and 8%, respectively, indicating that this invention possesses higher practicality and stability under complex watershed conditions.

[0131] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based automatic flash flood identification and real-time early warning system, characterized in that, include: The data acquisition and quality control module is used to collect precipitation and topographic data, align them with the spatial grid at a uniform time step, and generate basic observation data with quality control markers. The precipitation correction and fusion module is used to correct the time-step deviation of multi-source precipitation and generate a rasterized precipitation field with fixed resolution and uncertainty measurement. The watershed topology construction module is used to fill depressions and connect flat areas in terrain data, and generate a list of nodes and a list of upstream and downstream connections to form a watershed topology. The spatiotemporal feature generation module is used to align the rasterized precipitation field with the watershed topology, summarize precipitation by node coverage area and fill in missing measurements to obtain the spatiotemporal feature sequence of nodes and edges. The deep prediction module is used to input spatiotemporal feature sequences into a spatiotemporal deep learning model, train it according to quantile loss and classification loss, and generate deep prediction results. The hydrodynamic simulation module is used to set initial and boundary conditions on an unstructured mesh, solve the two-dimensional shallow water equations using finite volume discretization and double conjugate gradient stabilization iteratively, and output hydrodynamic simulation results. The coupling correction and early warning release module is used to update the prediction and boundary to obtain the coupling correction results, generate hierarchical early warnings, and complete the archiving.

2. The flash flood automatic identification and real-time early warning system based on deep learning according to claim 1, characterized in that, The modules are connected in the following way: Multi-source precipitation and topographic data were collected, and quality control and integrity verification were performed to obtain quality-controlled basic observation data. Bias correction and spatiotemporal fusion are performed on precipitation data to generate a rasterized precipitation field that includes uncertainty metrics; Based on topographic data, watershed partitions and upstream-downstream connectivity are established, and watershed topology is constructed. Align the rasterized precipitation field with the watershed topology and generate spatiotemporal feature sequences of nodes and edges at a unified time step; Input the spatiotemporal feature sequence into the spatiotemporal deep learning model and output the deep prediction results for the future forecast period; Driven by depth prediction results and gridded precipitation fields, a two-dimensional shallow water hydrodynamic simulation is performed in the target area to obtain hydrodynamic simulation results. Based on the depth prediction and hydrodynamic simulation results as input, a dual-drive coupled correction is implemented, the deviation in time and space is calculated and the correction amount is generated, the depth prediction and simulation boundary is updated, and the coupled correction result is output. The dynamic trigger threshold is calculated based on uncertainty measurement and spatiotemporal characteristic sequence, reliability calibration is performed, graded early warning is generated, early warning information is issued and archived.

3. The flash flood automatic identification and real-time early warning system based on deep learning according to claim 2, characterized in that, The acquisition of the quality-controlled basic observation data specifically includes: Establish a unified time step and spatial grid, determine spatial reference coordinates and target grid resolution, compile a data collection list and create tasks, acquire precipitation data from rain gauges, radar, and satellites, and acquire topographic data at the same time, recording the source, timestamp, coordinates and integrity markers for each data entry; We performed numerical range and temporal variation checks on precipitation and topographic data, conducted source consistency and multi-source comparison checks, used sliding window statistics to mark anomalies, and generated quality control tags for each data point to form a basic dataset. Precipitation data is converted to a unified coordinate system and resampled to the target grid with a unified time step. Topographic data is aligned to the same spatial grid and output quality-controlled basic observation data.

4. The flash flood automatic identification and real-time early warning system based on deep learning according to claim 2, characterized in that, The formation of the gridded precipitation field specifically includes: The system retrieves the quality-controlled basic observation data, reads and aligns it to the same timestamp and target raster in sequence according to time steps, performs bias correction on the precipitation data, and outputs the corrected precipitation data. Using corrected precipitation data as input, spatiotemporal fusion is performed on a pixel-by-pixel basis. By taking into account source reliability, observation error level and neighborhood consistency, a rasterized precipitation field with fixed spatial resolution and fixed time step is generated, and the uncertainty metric for each pixel is given.

5. A flash flood automatic identification and real-time early warning system based on deep learning according to claim 2, characterized in that, The construction of the watershed topology specifically includes: On a unified spatial grid, load terrain data and align it to the target coordinates and grid resolution. Perform depression filling and flat depression connectivity processing to eliminate local depressions without drainage and false steps, and generate a hydrologically consistent terrain grid. Based on the terrain grid, the direction of water flow and upstream source are determined grid by grid. The cumulative amount of confluence is calculated in the order from upstream to downstream. Channel pixels are extracted based on the cumulative amount and continuously tracked to form the river line. The confluence points and outflow points of the river are identified to obtain the river skeleton and the set of key control points. Constrained by the river channel framework and key control points, the grid is uniquely assigned to the corresponding control point according to the confluence of the river mouth, and sub-basin partitions are obtained. With sub-basins as nodes and upstream and downstream connections as directions, a node list and an upstream and downstream connection list are generated to construct the watershed topology.

6. The flash flood automatic identification and real-time early warning system based on deep learning according to claim 2, characterized in that, The generation process of the spatiotemporal feature sequence specifically includes: Using the rasterized precipitation field and uncertainty metric as input and the watershed topology as constraint, spatial alignment is performed on the target raster at a unified time step, establishing the grid-to-node assignment relationship, generating the node coverage area and edge adjacency list, and mapping the quality control marker and uncertainty metric to the corresponding node and edge. The coverage area of ​​each node is traversed by time step, and the precipitation values ​​in the area are collected according to the quality control mark and uncertainty measure. Missing locations are filled in according to the principle of time proximity and spatial proximity, and node-level uncertainty is obtained simultaneously, forming a node feature record containing precipitation, state quantity and uncertainty. Static attributes are collected for each directed edge, and temporal features of the edge are generated by combining the node features of adjacent time steps. The sliding window length and step size are set, and the node and edge features of consecutive time steps are packaged into a spatiotemporal feature sequence in topological order from upstream to downstream.

7. A flash flood automatic identification and real-time early warning system based on deep learning according to claim 2, characterized in that, The specific process of outputting the depth prediction results includes: Set the step size and number of steps of the sliding window, align the spatiotemporal feature sequence with the nodes and upstream and downstream connection list of the watershed topology, package them into the model input sample set, and record the timestamp, coverage node range and corresponding quality control mark for each sample. The model is input into the spatiotemporal deep learning model using the input sample set. Training is performed in batches. The quantile loss function is used to constrain the fitting error of the quantile flow, and the binary cross-entropy loss is used to constrain the classification error of the trigger probability. Iterative updates are performed according to the division of the training set, validation set, and test set until the convergence condition is met on the validation set, and the convergence parameters and model version are obtained. At each current time step, the node and edge feature sequences are extracted as model inputs. The convergence parameters are called for inference, and the quantile flow and trigger probability of each node in the future forecast period are output. The results are organized into a depth prediction result according to the time step and node index.

8. A flash flood automatic identification and real-time early warning system based on deep learning according to claim 2, characterized in that, The hydrodynamic simulation results are obtained specifically in the following ways: A two-dimensional computational domain is established and an unstructured triangular mesh is generated. The two-dimensional hydrodynamic model control equations are constructed using velocity and water depth as basic state variables. The gridded precipitation field is mapped to the surface rainfall drive within the mesh. The depth prediction results are mapped to the upstream inflow boundary time series according to the watershed topology. A fixed water level is set in the downstream area to complete the initial water depth and boundary condition configuration. The finite volume discretization method is used to solve the governing equations of the two-dimensional hydrodynamic model. Combined with the double conjugate gradient stabilization iterative method, the linear equations for each time step are solved. By traversing all grid cells and boundaries, the water exchange of each boundary and the source term contribution of each cell are calculated according to the water depth and velocity of the previous time layer. The sparse coefficient matrix and right-hand side are assembled. The water depth and velocity of the previous time layer are used as the initial solution. The initial residual is calculated and the residual norm is recorded. The convergence threshold, the maximum number of iterations and the divergence protection value are set. In each iteration, two sparse matrix and vector multiplication operations are performed, the residual is updated and the stabilization correction is performed. The magnitude of residual decrease and the magnitude of solution vector change are evaluated immediately after each iteration. After convergence, the water depth and velocity of the current time layer are written back to the grid to complete the update of the current time step. After the time progression ends, the water depth sequence of each unit changes over time, the moment when the water depth reaches the extreme value is extracted as the peak arrival time, the maximum water depth of the entire period is recorded as the maximum water level, the flooding range is defined by the area where the water depth exceeds the set threshold, and the hydrodynamic simulation results are output.

9. A flash flood automatic identification and real-time early warning system based on deep learning according to claim 2, characterized in that, The output of the coupling correction result specifically includes: On a unified time step and spatial grid, the depth prediction results are aligned with the hydrodynamic simulation results. According to the watershed topology, the water depth field and velocity field in the hydrodynamic simulation are mapped to the node region. The average area of ​​the water depth in each node region is taken, and the difference sequence with the corresponding depth prediction results is calculated. The difference sequence is compared in time order to generate the initial value of the node-level time deviation. Based on the uncertainty measurement results, weight coefficients are assigned to each node and grid point. Time deviation and spatial deviation are weighted to form time correction and spatial correction. The time correction is used to adjust the amplitude and time position of the flow prediction curve of each node. The corrected flow is mapped to the corrected inflow boundary along the upstream and downstream connectivity. The spatial correction is used to locally smooth and adjust the water depth distribution in the hydrodynamic simulation. The trigger probability is fine-tuned according to the direction of the corrected flow change. Set a stopping condition, iteratively perform the calculation of the correction amount and the update of the simulation boundary, and stop when the stopping condition is met, and output the coupling correction result.

10. A flash flood automatic identification and real-time early warning system based on deep learning according to claim 2, characterized in that, The process of issuing and archiving early warning information specifically includes: The dynamic trigger threshold of each node is calculated by combining the spatiotemporal feature sequence and corrected according to the current uncertainty to form a threshold table; The coupling correction results are compared with the dynamic trigger thresholds in the threshold table node by node to generate a trigger judgment quantity. The judgment quantity and trigger probability of each node are comprehensively evaluated, reliability calibration is performed, the node trigger results are classified into levels, and a level warning result is generated. The results of the graded early warning are deduplicated and throttled. The filtered graded early warning results are organized into records according to time sequence and spatial location. The output includes structured information containing trigger nodes, early warning levels, confidence levels and impact ranges, and is synchronously written to the archived database.

Images