Mountain torrent disaster early warning system based on deep learning and air-space-ground integrated perception

Abstract

The invention relates to the technical field of disaster early warning processing, in particular to a mountain torrent disaster early warning system based on deep learning and air-space-ground integrated perception. The system comprises the following modules: a space-air-ground sensing module, an edge calculation preprocessing module, a deep learning prediction module and an intelligent decision and early warning module. Wherein the air-space-ground sensing module is used for collecting multi-source data; wherein the multi-source data comprises earth surface deformation data obtained through a spaceborne synthetic aperture radar, high-precision topographic data obtained through an unmanned aerial vehicle laser radar, hydrological sensing data obtained through a ground internet-of-things distributed sensor, and meteorological numerical forecasting data obtained through a meteorological data interface. Through air-space-ground integrated perception, physical constraint deep learning and edge intelligence technologies, mountain torrent multi-source data collaborative acquisition and accurate mountain torrent disaster prediction are realized, so that the comprehensiveness, reliability and timeliness of mountain torrent disaster early warning are improved.

Description

Flash flood disaster early warning system based on deep learning and integrated air-space-ground perception Technical Field

[0001] This invention relates to the field of disaster early warning and processing technology, and in particular to a flash flood disaster early warning system based on deep learning and integrated air-space-ground perception. Background Technology

[0002] Existing flash flood early warning technologies suffer from three major shortcomings: First, the monitoring dimensions are limited to surface hydrological parameters, lacking the ability to perceive macroscopic deformations (such as landslides blocking rivers) and micro-topography (such as karst underground rivers), resulting in "monitoring blind spots." For example, traditional systems cannot identify the risk of river blockage caused by upstream landslide displacement. Second, data fusion often uses simple weighting methods, which cannot handle the spatiotemporal scale differences between SAR (low frequency, 12 days / time) and ground-based sensor (high frequency, 1Hz) data, resulting in insufficient fusion accuracy. For example, the matching error between NWP data (1km resolution) and ground grid (10m) reaches 50%. Third, the prediction models have poor adaptability to small sample watersheds. Purely data-driven models are prone to physical deviations (such as the contradiction between predicted peak flow and river roughness), and the forecast period is often less than 3 hours, making it difficult to meet emergency response needs. For example, in a watershed in the karst mountains of Southwest China (with only 4 sets of historical data), the prediction error using a traditional model reached 35%, with a false alarm rate of 15%. Summary of the Invention

[0003] Therefore, it is necessary to provide a flash flood disaster early warning system based on deep learning and integrated air-space-ground perception to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a flash flood disaster early warning system based on deep learning and integrated air-space-ground perception includes the following modules: an air-space-ground perception module for collecting multi-source data, including: surface deformation data acquired through spaceborne synthetic aperture radar, high-precision terrain data acquired through UAV lidar, hydrological sensing data acquired through ground-based IoT distributed sensors, and meteorological numerical forecast data acquired through a meteorological data interface; an edge computing preprocessing module for receiving the multi-source data output from the air-space-ground perception module, performing spatiotemporal registration, denoising, and missing data reconstruction operations, and outputting a standardized spatiotemporal tensor dataset; a deep learning prediction module for receiving the standardized spatiotemporal tensor dataset output from the edge computing preprocessing module, extracting data features from the standardized spatiotemporal tensor dataset through a spatiotemporal attention fusion model with embedded preset physical constraints, and predicting flash flood disaster parameters; and an intelligent decision-making and early warning module for receiving the flash flood disaster parameter prediction results output from the deep learning prediction module, constructing a two-dimensional early warning matrix based on the DS evidence theory, generating tiered early warning information, and linking the tiered early warning information with a geographic information system to generate refined evacuation routes.

[0005] The beneficial effects of this invention are as follows: First, the integrated air-space-ground sensing architecture adopted in this invention breaks through the limitations of the traditional single monitoring mode, realizes the collaborative collection of multi-dimensional and multi-scale data on surface deformation, topographic features, hydrological dynamics and meteorological changes, and builds a monitoring system with full coverage and multi-element linkage. It ensures the comprehensiveness and accuracy of flash flood disaster early warning from the data source and effectively avoids the problem of early warning deviation caused by data fragmentation and incomplete coverage in traditional monitoring.

[0006] Second, by embedding a spatiotemporal attention fusion model with physical constraints, the evolution law of watershed hydrology and physics is deeply integrated with deep learning algorithms. This not only fully explores the spatiotemporal correlation features in multi-source data, but also ensures that the prediction results conform to objective natural laws through physical constraint calibration. This solves the pain points of pure data-driven models being prone to being detached from real-world scenarios and having insufficient prediction credibility, and significantly improves the scientificity and reliability of flash flood disaster parameter prediction.

[0007] Third, by leveraging the collaborative design of edge computing and deep learning, the localized and efficient execution of data preprocessing and feature extraction was achieved, reducing data transmission latency and computing power consumption. Combined with the meta-learning strategy, the system has the ability to quickly adapt to small sample watersheds, breaking through the dependence of traditional early warning technology on massive historical data. This significantly improves the universality of the system in data-scarce mountainous watersheds and provides a flexible and feasible technical solution for flash flood early warning in different geographical environments.

[0008] Fourth, the fusion of multi-source prediction results and the construction of a two-dimensional early warning matrix based on DS evidence theory have enabled the accurate classification of early warning levels. Combined with the refined evacuation routes generated by the geographic information system, a closed-loop mechanism of "prediction-early warning-response" has been formed, which efficiently transforms technical prediction results into implementable emergency response plans, significantly improving the pertinence and timeliness of emergency response to flash floods. Attached Figure Description

[0009] Figure 1 is a schematic diagram of the execution flow of a flash flood disaster early warning system based on deep learning and integrated air-space-ground perception; Figure 2 is a schematic diagram of the modules of the flash flood disaster early warning system based on deep learning and integrated air-space-ground perception; Figure 3 is a schematic diagram of the spatiotemporal attention fusion model architecture; The realization of the purpose, functional characteristics and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0010] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0011] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0012] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0013] To achieve the above objectives, please refer to Figures 1 to 3. A flash flood disaster early warning system based on deep learning and integrated air-space-ground perception is described. The flash flood disaster early warning system 100 based on deep learning and integrated air-space-ground perception includes an air-space-ground perception module 101, an edge computing preprocessing module 102, a deep learning prediction module 103, and an intelligent decision-making and early warning module 104. The flash flood disaster early warning system 100 based on deep learning and integrated air-space-ground perception performs the following steps: S1: Collecting multi-source data; wherein, the multi-source data includes: surface deformation data acquired through spaceborne synthetic aperture radar, high-precision terrain data acquired through UAV lidar, and hydrological sensing data acquired through ground-based IoT distributed sensors. S1: Obtain meteorological numerical forecast data through the meteorological data interface; S2: Receive multi-source data output from the air-space-ground sensing module, perform spatiotemporal registration, denoising, and missing data reconstruction operations, and output a standardized spatiotemporal tensor dataset; S3: Receive the standardized spatiotemporal tensor dataset output from the edge computing preprocessing module, extract data features of the standardized spatiotemporal tensor dataset through a spatiotemporal attention fusion model with embedded preset physical constraints, and predict flash flood disaster parameters; S4: Receive the flash flood disaster parameter prediction results output from the deep learning prediction module, construct a two-dimensional early warning matrix based on the DS evidence theory and fuse the flash flood disaster parameter prediction results, and generate graded early warning information; link the graded early warning information with the geographic information system to generate refined evacuation routes.

[0014] Preferably, the air-space-ground sensing module is used to collect multi-source data. This multi-source data includes: surface deformation data acquired via spaceborne synthetic aperture radar (SAR), high-precision terrain data acquired via UAV lidar, hydrological sensing data acquired via ground-based IoT distributed sensors, and meteorological numerical forecast data acquired via a meteorological data interface. Optionally, when collecting multi-source data, the air-space-ground sensing module adopts a three-level collaborative mode involving space, aircraft, and ground. Specifically, the spaceborne SAR periodically observes the entire watershed, recording displacement changes at different locations on the surface to form continuous temporal data of surface deformation; the UAV lidar covers the watershed terrain along a preset route, generating high-precision terrain data including topographic undulations and valley distribution through high-density scanning; the ground-based IoT distributed sensors are deployed according to the watershed terrain characteristics and monitoring needs, simultaneously collecting hydrological sensing data composed of rainfall, soil moisture content, and microseismic signals; the meteorological data interface accesses regional meteorological forecast data through a standardized protocol, extracting rainfall-related meteorological numerical forecast data; and the data is associated and labeled according to the collection source, timestamp, and spatial coordinates to form a structured multi-source dataset.

[0015] In this embodiment, when the space-air-ground sensing module collects multi-source data, it adopts a three-level collaborative mode of space-air-ground system. Specifically, the spaceborne synthetic aperture radar uses the Sentinel-1 satellite IW mode to conduct periodic observations of the entire watershed. It acquires surface deformation data with a resolution of 0.5m and an average monthly accuracy of ±3mm through SBAS-InSAR technology, records the displacement changes at different locations on the surface, and forms continuous time-series data of surface deformation. The original observation interval of this time-series data is 12 days, which is converted to a 1-hour scale through cubic spline interpolation to match the time series of other data. The UAV is equipped with RIEGLVUX-1UAV lidar, which fully covers the watershed terrain according to a preset route and carries out operations with a high-density scanning mode of 50 points / ㎡ to generate a 1:500 scale digital elevation model and digital surface model. This model includes terrain information such as the amplitude of terrain undulation and the distribution and direction of valleys, constituting high-precision terrain data.

[0016] In one embodiment, the ground-based IoT distributed sensor uses distributed fiber optic grating sensing nodes, which are densely deployed according to the confluence nodes, slope areas, and key risk areas in the watershed topography, and evenly deployed in regular areas. The total number of nodes is 20-50. The sensing nodes include rainfall sensors, microseismic monitors, and soil moisture sensors. The rainfall sensors collect rainfall data at a sampling frequency of 1Hz, with a measurement range of 0-200mm / h and an accuracy of ±0.5%FS. The microseismic monitors capture regional microseismic signals at a monitoring frequency of 10-1000Hz and a resolution of 16 bits. The soil moisture sensors collect soil moisture content data, with a measurement range of 0-100%vol and an error controlled within 2%. The three types of sensors collect data synchronously and form hydrological sensing data. The meteorological data interface accesses ECMWF meteorological numerical forecast data through a standardized protocol. The original resolution of this data is 1km×1km, and the lead time covers 0-72 hours. Rainfall-related meteorological numerical forecast data is extracted from this data.

[0017] In another embodiment, the time-series data of surface deformation acquired by spaceborne synthetic aperture radar, the high-precision terrain data generated by UAV lidar, the hydrological sensing data collected by ground IoT distributed sensors, and the meteorological numerical forecast data extracted by the meteorological data interface are classified and labeled according to the source of collection, and the timestamps and spatial coordinates corresponding to each data are synchronously associated. The spatial coordinates are uniformly recorded using the WGS-84 coordinate system, and finally a structured multi-source dataset is formed.

[0018] Preferably, the edge computing preprocessing module is used to receive multi-source data output by the air-space-ground sensing module, perform spatiotemporal registration, denoising, and missing data reconstruction operations, and output a standardized spatiotemporal tensor dataset. Optionally, after receiving the multi-source data output by the air-space-ground sensing module, the edge computing preprocessing module performs processing operations according to the process of spatiotemporal unification, purification and supplementation, and format standardization, specifically as follows: aligning the timestamps of the multi-source data through time synchronization, unifying the spatial coordinates of data from different sources to the same coordinate system to complete spatiotemporal registration; separating the effective signal and noise components in the original sensor data through signal decomposition to remove noise interference; supplementing missing data through adjacent data correlation analysis; and regularizing the processed multi-source data according to a fixed time step and spatial grid scale, converting it into a standardized spatiotemporal tensor dataset with a unified format and outputting it.

[0019] In this embodiment, Kalman filtering technology is used to align the timestamps of multi-source data, unifying and calibrating the time references of spaceborne synthetic aperture radar surface deformation time series data, UAV lidar high-precision terrain data, ground IoT distributed sensor hydrological sensing data, and meteorological numerical forecast data, so that the timestamp error is controlled within 1 second; the spatial coordinates of all data are uniformly transformed to the 2000 National Geodetic Coordinate System through the base station-rover differential positioning method, so as to achieve spatial reference consistency of multi-source data and complete spatiotemporal registration.

[0020] In one embodiment, empirical mode decomposition (EMD) technology is used to decompose the raw spectral signals of ground-based IoT distributed sensors, separating the effective signals from noise components such as mountain vibrations, eliminating noise interference, and improving the data signal-to-noise ratio by 30 dB. By utilizing the correlation between data from adjacent time periods and adjacent areas, super-resolution reconstruction technology is used to supplement missing data, downscaling the meteorological numerical forecast data from the original resolution of 1 km × 1 km to a 10 m × 10 m grid, and controlling the data reconstruction error to within 5%.

[0021] In another embodiment, the multi-source data after spatiotemporal registration, denoising, and missing data supplementation are regularized according to a fixed time step of 1 hour and a spatial grid scale of 10m×10m. Data of different types and formats are converted into tensor forms with a unified structure to form a standardized spatiotemporal tensor dataset, which is then output. This dataset contains multi-dimensional integrated information on surface deformation, topography, hydrology, and meteorology, meeting the input requirements of the deep learning prediction module.

[0022] Preferably, the deep learning prediction module receives the standardized spatiotemporal tensor dataset output by the edge computing preprocessing module, extracts data features from the standardized spatiotemporal tensor dataset through a spatiotemporal attention fusion model with embedded preset physical constraints, and predicts flash flood disaster parameters. Optionally, after receiving the standardized spatiotemporal tensor dataset output by the edge computing preprocessing module, the deep learning prediction module extracts data features and predicts flash flood disaster parameters through a spatiotemporal attention fusion model with embedded preset physical constraints, specifically by: determining the feature extraction unit and parameter prediction unit in the spatiotemporal attention fusion model; performing hierarchical processing on the standardized spatiotemporal tensor dataset through the feature extraction unit, separating the single features of four types of data (surface deformation, topography, hydrology, and meteorology), and determining the cross-correlation features between different types of data; and constructing prediction logic based on the cross-correlation features through the parameter prediction unit, calibrating the output results by embedding preset physical constraints, and outputting preliminary flash flood disaster parameter prediction results.

[0023] In this embodiment, the spatiotemporal attention fusion model includes a feature extraction unit and a parameter prediction unit. The feature extraction unit consists of a Transformer global feature extraction layer, a spatiotemporal attention enhancement layer, and a related feature integration layer. The parameter prediction unit consists of a temporal modeling layer and a physical constraint calibration layer. Each layer performs processing operations sequentially according to the data flow order. The overall model adopts a hierarchical structure of 3 Transformer global feature extraction layers, 2 spatiotemporal attention enhancement layers, 1 related feature integration layer, 4 temporal modeling layers, and 1 physical constraint calibration layer.

[0024] In one embodiment, the standardized spatiotemporal tensor dataset is processed hierarchically by a feature extraction unit. First, the Transformer global feature extraction layer performs dimensional mapping on the data, converting the 10m×10m spatial grid and 1-hour time step standardized spatiotemporal tensor dataset into a feature matrix with consistent dimensions. Each Transformer global feature extraction layer has 64 feature dimensions. A self-attention mechanism is used to globally traverse the feature matrix, separating single features from four types of data: surface deformation, topography, hydrology, and meteorology. The topography single feature includes slope, curvature, and runoff accumulation; the hydrology single feature includes soil saturated water content, pore water pressure, and permeability coefficient; and the meteorology single feature includes rainfall return period and convective effective potential. The four types of single features are: energy, 1-hour rainfall intensity, and surface deformation single features including displacement amplitude and deformation rate. Then, the spatiotemporal attention enhancement layer assigns weights to the four types of single features. In the spatial dimension, features corresponding to confluence nodes and weak slope areas are assigned a weight coefficient of 0.7, and in the temporal dimension, features corresponding to short-term heavy rainfall periods of 0-3 hours and sudden changes in surface deformation are assigned a weight coefficient of 0.8, thereby strengthening the representation of key spatiotemporal features. Finally, the correlation feature integration layer calculates the Pearson correlation coefficient among the four types of single features. Based on the correlation coefficient, feature combinations with a correlation higher than 0.6 are selected, and cross-correlated features are generated through matrix multiplication. All single features and cross-correlated features are processed by max-min normalization to the range [0,1] to form a 32-dimensional feature vector.

[0025] In another embodiment, a prediction logic is constructed based on cross-correlation features through a parameter prediction unit. The temporal modeling layer adopts a gated loop structure, with 128 hidden units in each layer. Iterative calculations are performed on the 32-dimensional feature vector in chronological order to capture the evolution of the feature vector over time. Preset physical constraints are embedded during the iterative calculation process. These constraints include constraints related to Manning's formula and Darcy's law, which are incorporated into the calculation of computational loss at a weight of 20% of the total weight. The weight of the Manning's formula constraint term is dynamically adjusted according to the river channel roughness. When the river channel roughness is greater than 0.03, the weight is adjusted to 0.3. The Darcy's law constraint term sets the soil saturated hydraulic conductivity parameter according to the soil type, with a value of 10 for sandy soil. -3 m / s, for clay soil, the value is 10.-6 m / s; The output of the temporal modeling layer is verified by the physical constraint calibration layer. The deviation between the output and the theoretical value of the physical constraint is calculated. When the deviation exceeds 10%, the weight coefficient of the spatiotemporal attention enhancement layer and the hidden unit parameters of the temporal modeling layer are adjusted in reverse. Feature extraction and iterative calculation are re-executed. Finally, the preliminary flash flood disaster parameter prediction results containing the peak flow, inundation depth and probability of disaster occurrence in the next 0-6 hours are output. The peak flow prediction result is retained to two decimal places, the inundation depth prediction result is retained to one decimal place, and the probability of disaster occurrence is presented as a percentage and retained to an integer.

[0026] Optionally, when the spatiotemporal attention fusion model extracts data features, setting up dual attention channels includes: identifying the spatial locations of confluence nodes and weak slope areas within the watershed based on the spatiotemporal attention fusion model, and assigning priority extraction weights to the data features of spatial locations; identifying the time periods corresponding to short-term heavy rainfall and abrupt changes in surface deformation based on the spatiotemporal attention fusion model, and determining the impact factors of time periods on disaster prediction.

[0027] In this embodiment, when the spatiotemporal attention fusion model extracts data features, the dual attention channels set include a spatial attention channel and a temporal attention channel. The corresponding layer of the model is a spatiotemporal attention enhancement layer, which contains two parallel sub-layer structures, corresponding to the spatial attention calculation sub-layer and the temporal attention calculation sub-layer, respectively. Each sub-layer has 64 feature mapping dimensions and adopts a computational structure combining fully connected layers and activation functions. The specific operation is as follows: In one embodiment, the spatial attention calculation sub-layer first calls the high-precision terrain data of UAV lidar in the standardized spatiotemporal tensor dataset, calculates the global terrain slope value through the terrain slope algorithm, analyzes the terrain profile information through the valley cutting depth extraction algorithm, sets the valley cutting depth greater than 5m as the threshold for determining the confluence node, and the slope greater than 25° as the threshold for determining the weak slope area. Based on the thresholds, the spatial coordinates of the confluence node and the weak slope area in the watershed are selected. A binary spatial attention mask is generated based on spatial coordinates. In the mask, 1 corresponds to the confluence node and the weak area of ​​the slope, and 0 corresponds to the normal area. The mask is input into the fully connected layer and converted into a weight matrix. The data features corresponding to the mask coverage area are given a priority extraction weight of 0.7, and the data features of the normal area are given a basic extraction weight of 0.3. The spatial dimension weight is assigned by multiplying the weight matrix and the feature matrix element by element, which strengthens the representation of the data features of key spatial locations.

[0028] In another embodiment, the temporal attention computation sublayer extracts 1-hour rainfall intensity data from the hydrological sensing data of ground-based IoT distributed sensors and deformation rate data from the time-series surface deformation data of spaceborne synthetic aperture radar from the standardized spatiotemporal tensor dataset. A 1-hour rainfall intensity greater than 20 mm is set as the threshold for short-duration heavy rainfall, and a surface deformation rate greater than 5 mm / day is set as the threshold for abrupt surface deformation changes. A sliding window method with a sliding window size of 3 is used to traverse the data sequence to identify the time periods corresponding to short-duration heavy rainfall and abrupt surface deformation changes. The identified time periods are marked as critical time periods. An activation function is used to assign an influence weight of 0.8 to the data features of critical time periods and an influence weight of 0.2 to the data features of non-critical time periods. Based on the weight ratio and the variance of the data features, the influence factor of the time period on disaster prediction is calculated. After normalization, the influence factor is input into the feature fusion layer and integrated with the spatial attention-weighted features to form an enhanced data feature that takes into account key spatiotemporal characteristics.

[0029] Optionally, the preset physical constraints include constraints related to water flow patterns and soil infiltration patterns, specifically: embedding the preset physical constraints into the loss function of the spatiotemporal attention fusion model at a fixed ratio; automatically triggering a parameter correction mechanism when the prediction result deviates from the preset physical laws beyond a preset range; and adjusting the feature extraction weights and prediction logic of the spatiotemporal attention fusion model so that the prediction result conforms to the preset watershed hydrophysical evolution laws.

[0030] In this embodiment, the preset physical constraints include Manning's formula constraint corresponding to the water flow law and Darcy's law constraint corresponding to the soil infiltration law. The specific implementation is as follows: the Manning's formula constraint term and the Darcy's law constraint term are embedded into the loss function of the spatiotemporal attention fusion model at a fixed ratio of 20% of the total weight, wherein the weight of the Manning's formula constraint term accounts for 12% and the weight of the Darcy's law constraint term accounts for 8%. The loss function as a whole is composed of the weighted sum of the data fitting loss and the physical constraint loss.

[0031] In one embodiment, the Manning formula constraint term dynamically adjusts its local weight based on the river channel roughness. The river channel roughness is set according to the watershed underlying surface type: 0.035 for forest-covered areas, 0.025 for cultivated land-covered areas, and 0.015 for urban areas. When the river channel roughness is greater than 0.03, the local weight of the Manning formula constraint term is adjusted to 0.3. The Darcy's law constraint term sets the soil saturated hydraulic conductivity parameter according to soil type: 10 for sandy soil. -3 m / s, for clay soil, the value is 10. -6 m / s, soil value is 10 -4 m / s, this parameter is used to correct the infiltration flow calculation logic.

[0032] In another embodiment, a 10% threshold is set for the deviation between the predicted results and the theoretical calculations based on Manning's formula and Darcy's law. When the peak flow rate in the preliminary flash flood disaster parameter prediction results deviates by more than 10% from the theoretical calculations based on Manning's formula, or the inundation depth deviates by more than 10% from the water depth corresponding to the soil infiltration volume derived from Darcy's law, a parameter correction mechanism is automatically triggered. The feature extraction weights of the spatial attention channel and the temporal attention channel in the spatiotemporal attention fusion model are adjusted using a backpropagation algorithm. The adjustment range of the feature weights corresponding to confluence nodes and short-duration heavy rainfall periods is controlled within ±0.1. Simultaneously, the temporal modeling logic of the parameter prediction unit is corrected, and iterative calculations are performed again until the deviation between the predicted results and the preset watershed hydrophysical evolution law is reduced to within 10%.

[0033] Most importantly, before the deep learning prediction module extracts data features through the spatiotemporal attention fusion model with embedded preset physical constraints, it performs feature engineering processing on the standardized spatiotemporal tensor dataset, including: extracting topographic features such as terrain slope and runoff accumulation; extracting hydrological features such as soil moisture content and pore water pressure; extracting meteorological features such as rainfall intensity and rainfall duration; and normalizing all features to form a standardized feature vector that is input into the spatiotemporal attention fusion model.

[0034] In this embodiment, before the deep learning prediction module extracts data features through a spatiotemporal attention fusion model with embedded preset physical constraints, it performs feature engineering processing on the standardized spatiotemporal tensor dataset, specifically as follows: high-precision terrain data from UAV lidar is extracted from the standardized spatiotemporal tensor dataset, and the slope value of the entire grid cell is calculated based on the elevation difference of a 3×3 neighborhood window, with an output range of 0°-90°; based on the digital elevation model, after determining the direction of water flow, the upstream confluence area of ​​each grid cell is accumulated to obtain the confluence accumulation feature, with the unit being m².

[0035] In one embodiment, soil moisture content and pore water pressure data are extracted from the hydrological sensing data of the distributed sensor of the ground Internet of Things. The soil moisture content data is directly obtained from the raw data of 0-100%vol collected by the sensor. The pore water pressure data is calculated by converting the Bragg wavelength shift of the distributed fiber optic grating sensor with a conversion factor of 0.01 kPa / pm. The data unit is uniformly kPa. Rainfall intensity and rainfall duration data are extracted from the meteorological numerical forecast data. The rainfall intensity is statistically calculated in 1-hour time units, with the unit being mm / h. The rainfall duration is the cumulative duration of continuous rainfall periods, with the unit being h.

[0036] In another embodiment, additional features are extracted, including topographic curvature, soil saturated water content, permeability coefficient, rainfall return period, convective available potential energy, surface deformation displacement amplitude, and deformation rate. Topographic curvature is calculated using the second derivative, soil saturated water content is set to a fixed value according to soil type, permeability coefficient is determined with reference to the soil texture classification table, rainfall return period is obtained using the annual maximum value statistical method, and convective available potential energy is directly extracted from meteorological numerical forecast data. All extracted 32-dimensional features are processed using the maximum-minimum normalization method, mapping the feature values ​​to the [0,1] interval. The normalization formula is: Standardized feature value = (Original feature value - Minimum feature value) / (Maximum feature value - Minimum feature value). Finally, a 32-dimensional standardized feature vector is formed and input into the spatiotemporal attention fusion model.

[0037] Optionally, before the deep learning prediction module predicts flash flood disaster parameters through the spatiotemporal attention fusion model, the spatiotemporal attention fusion model is optimized through multiple rounds of iterative training. Specifically, the first round is based on the full feature vector for basic training. In each subsequent round of training, redundant features are eliminated through feature importance analysis, and the core features that affect disaster prediction are retained. The physical constraint weights and attention channel parameters are adjusted. After each round of training, the model accuracy is verified using a validation dataset until the model prediction error drops to a preset range.

[0038] Please refer to Figure 3. In this embodiment, before the deep learning prediction module predicts the parameters of flash flood disaster through the spatiotemporal attention fusion model, it performs multiple rounds of iterative training and optimization on the spatiotemporal attention fusion model. The feature extraction unit of the spatiotemporal attention fusion model includes 3 layers of Transformer global feature extraction layer, 2 layers of spatiotemporal attention enhancement layer, and 1 layer of associated feature integration layer. The parameter prediction unit of the spatiotemporal attention fusion model includes 4 layers of temporal modeling layer, 4 layers of physical constraint modeling layer, and 1 layer of physical constraint calibration layer. The feature dimension of each layer is set to 64 dimensions, and the number of hidden units in the temporal modeling layer is 128. Specifically, in one embodiment, the first round of training uses a 32-dimensional fully standardized feature vector as input, sets the number of iterations to 1000 rounds, and sets the basic learning rate to 0.001. During the training process, the total weight of physical constraints remains unchanged at 20%. The weight of key regions in the spatial attention channel is 0.7, and the weight of regular regions is 0.3. The weight of key time periods in the temporal attention channel is 0.8, and the weight of non-key time periods is 0.2. The basic training is completed and the initial prediction accuracy of the model is recorded.

[0039] In another embodiment, subsequent training rounds use feature contribution statistics to filter core features, eliminating redundant features with a contribution below 0.02, and retaining core features such as 1-hour rainfall intensity and previous soil moisture content index, whose cumulative contribution exceeds 60%. Simultaneously, physical constraint weights and attention channel parameters are adjusted. The Manning formula constraint weight is adjusted to 0.3 when the river channel roughness is greater than 0.03, and the Darcy's law constraint weight is adapted to the saturated hydraulic conductivity parameter according to soil type. The adjustment range of spatial and temporal attention weights is controlled within ±0.1. After each training round, an independent validation dataset is used to verify the model accuracy. The validation dataset contains historical monitoring data and disaster records from 10 typical watersheds. When the model's peak flow prediction error drops to within 8% and the inundation depth prediction error drops to within 0.3m, iterative training stops, and the final model parameters are determined.

[0040] Optionally, when the spatiotemporal attention fusion model predicts parameters for flash floods, it proceeds in layers according to the time dimension, including: first predicting the probability of flash flood occurrence and preliminary flood peak parameters during the 0-3 hour period; then predicting the disaster development trend and inundation range during the 3-6 hour period; forming a complete prediction chain based on the prediction results of different time periods during the 0-6 hour period, and outputting a flash flood prediction report containing multiple time periods and multiple parameters.

[0041] In this embodiment, when the spatiotemporal attention fusion model predicts flash flood disaster parameters, it advances in layers according to the time dimension. The model's temporal modeling layer is set as a short-term prediction sub-layer of 0-3 hours and a medium-term prediction sub-layer of 3-6 hours. Both layers contain a gated loop structure with 4 hidden units and 128 units. The feature input dimension is a 32-dimensional core feature vector, as follows: In one embodiment, for the prediction of the 0-3 hour period, the short-term prediction sub-layer uses a time step of 1 hour, receives the historical feature data of the most recent 6 hours from the standardized spatiotemporal tensor dataset, and combines the confluence node output by the spatial attention channel, the feature weight of the weak slope area, and the short-term heavy rainfall period influence factor output by the temporal attention channel. Through gated loop operation, it outputs the hourly probability of flash flood disaster occurrence and preliminary flood peak parameters. The flood peak parameters include the flood peak flow and the flood peak occurrence time. The flood peak flow is retained to two decimal places and the unit is m³ / s. The flood peak occurrence time is accurate to the minute. The probability of disaster occurrence is presented in the form of a percentage and retained to an integer place.

[0042] In another embodiment, for the 3-6 hour forecast, the intermediate forecast sublayer adjusts the iteration step size of the gated loop structure to 2 hours based on the evolution of the 0-3 hour forecast results and historical feature data, incorporating the dynamic changes of physical parameters such as river roughness and soil saturated hydraulic conductivity to predict the disaster development trend and inundation range. The inundation range is represented by the inundation depth of a 10m×10m grid, retaining one decimal place, with the unit being meters. The hourly forecast results of 0-3 hours and the segmented forecast results of 3-6 hours are integrated in chronological order, and the confidence interval of the forecast results for each time period is supplemented (the confidence level is set to 95%) to form a complete forecast chain covering 0-6 hours. Finally, a flash flood disaster forecast report containing the probability of disaster occurrence, flood peak parameters, disaster development trend, and inundation range for multiple time periods is output.

[0043] Preferably, the intelligent decision-making and early warning module is used to receive the flash flood disaster parameter prediction results output by the deep learning prediction module, construct a two-dimensional early warning matrix based on the DS evidence theory and fuse the flash flood disaster parameter prediction results, and generate hierarchical early warning information; and link the hierarchical early warning information with the geographic information system to generate refined evacuation routes.

[0044] Optionally, after receiving the flash flood disaster parameter prediction results output by the deep learning prediction module, the intelligent decision-making and early warning module constructs a two-dimensional early warning matrix based on the DS evidence theory by fusing the flash flood disaster parameter prediction results. Specifically, this involves: determining the multi-model fusion weights according to preset rules; performing credibility analysis on the prediction results of the spatiotemporal attention fusion model using DS evidence theory to obtain credibility data; fusing the prediction results into multi-source prediction results based on the credibility data; constructing an early warning matrix based on the multi-source prediction results and with the probability of disaster occurrence and the degree of disaster impact as the core dimensions; classifying early warning levels according to the threshold of the early warning matrix; and generating graded early warning information.

[0045] In this embodiment, after receiving the flash flood disaster parameter prediction results output by the deep learning prediction module, the intelligent decision-making and early warning module fuses the prediction results based on DS evidence theory and constructs a two-dimensional early warning matrix. The module includes an evidence processing unit, a multi-source fusion unit, and an early warning matrix construction unit. The evidence processing unit includes a credibility calculation subunit, a conflict identification subunit, and a conflict resolution subunit. The multi-source fusion unit has a built-in DS synthesis rule operation structure. Each unit executes according to the process of "evidence preprocessing - conflict handling - weighted synthesis", as follows: In one embodiment, the multi-model fusion weights are allocated according to preset rules. The prediction result weight of the spatiotemporal attention fusion model accounts for 60%, and the prediction results weights of the random forest model, XGBoost model, and CNN model account for 15%, 15%, and 10%, respectively. The weights are determined by back-calculating the accuracy of 100 sets of historical prediction data. For every 10% increase in accuracy, the corresponding model weight increases by 5%. The credibility calculation subunit maps the prediction results of each model to the basic probability allocation function in DS evidence theory. Taking flood peak flow prediction as an example, the deviation between the predicted value and the actual value is set as "high credibility proposition", 5%-10% as "medium credibility proposition", and >10% as "low credibility proposition", and basic probability values ​​of 0.7, 0.2 and 0.1 are assigned respectively, while retaining an uncertainty probability of 0.0. The conflict identification subunit calculates the evidence conflict coefficient between each model. The formula is K=Σ (product of the basic probability values ​​of each model for different propositions). When K>0.5, it is judged as high conflict, 0.3≤K≤0.5 is medium conflict, and K<0.3 is low conflict.

[0046] In another embodiment, corresponding resolution strategies are implemented for different conflict levels. For high conflict, a weighted correction method is used, which allocates the conflict coefficient K to the corresponding basic probability value according to the weight ratio of each model. The correction formula is m'(A)=m(A)+K×ω (ω is the model weight). For medium conflict, an evidence discount method is used, which multiplies the basic probability value of the conflict model by a discount factor of 0.8 to reduce its impact. For low conflict, the original basic probability value is directly retained. Based on the DS synthesis rule, the evidence of each model after resolution is synthesized, and the joint basic probability allocation function is calculated. The formula is m(A)=[Σ(product of the basic probability values ​​of each model for proposition A)] / (1-K), which yields the joint probability distribution of core parameters such as the probability of disaster occurrence and the degree of impact after multi-source fusion.

[0047] A two-dimensional early warning matrix is ​​constructed using the disaster occurrence probability (after multi-source fusion) as the horizontal axis and the disaster impact severity as the vertical axis. The disaster occurrence probability is divided into four intervals: 0-30%, 30%-50%, 50%-70%, and above 70%. The disaster impact severity is determined by the inundated area (≥10km² is severe, 5-10km² is relatively large, 1-5km² is moderate, and <1km² is minor), the number of affected people (≥10,000 is severe, 5,000-10,000 is relatively large, 1,000-5,000 is moderate, and <1,000 is minor), and the risk of infrastructure damage (damage to core facilities is severe, damage to important facilities is relatively large, damage to general facilities is moderate, and no facilities are damaged). The severity of the disaster is assessed using a comprehensive evaluation system, with corresponding impact values ​​of 0.2, 0.4, 0.6, and 0.8. A warning matrix threshold is set: a blue warning indicates a 30%-50% probability of occurrence with a minor impact; a yellow warning indicates a 50%-70% probability with a moderate impact; an orange warning indicates a 50%-70% probability with a significant impact or a 70% or higher probability with a moderate impact; and a red warning indicates a 70% or higher probability with a significant or higher impact. Based on the corresponding position of the multi-source fusion parameters in the matrix, the corresponding warning level is matched, generating tiered warning information that includes the warning level, the probability range of the disaster occurrence, the impact severity level, the core affected area, and key risk points.

[0048] Most importantly, when the intelligent decision-making and early warning module integrates the prediction results of flash flood disaster parameters based on DS evidence theory, it performs the following three steps: First, it converts the prediction results of each model into a standardized evidence body; second, it identifies and resolves conflicts between different pieces of evidence by calculating the conflict coefficient; third, it calculates the comprehensive credibility of the fused evidence by using evidence synthesis rules; and optimizes the threshold division of the early warning matrix based on the comprehensive credibility.

[0049] In this embodiment, the first step is to convert the prediction results of each model into a standardized evidence body: taking the flash flood disaster parameter prediction results of the spatiotemporal attention fusion model, random forest model, XGBoost model, and CNN model as the processing object, the prediction results include disaster occurrence probability, peak flow, inundation depth, and impact range data. Three core propositions are set: "Disaster level reaches red alert", "Disaster level reaches orange alert", and "Disaster level reaches yellow or below alert", and the prediction results of each model are mapped to the basic probability allocation value of the corresponding proposition. For the spatiotemporal attention fusion model, when the predicted probability of disaster occurrence is ≥70% and the inundation depth is ≥1.5m, the proposition "disaster level reaches red alert" is assigned a basic probability value of 0.7, the proposition "disaster level reaches orange alert" is assigned a basic probability value of 0.2, and the proposition "disaster level reaches yellow or below alert" is assigned a basic probability value of 0.1. The random forest model, XGBoost model, and CNN model are mapped to basic probability values ​​according to the same judgment criteria, while each model retains an uncertainty probability of 0.0, forming a standardized evidence body that includes a set of propositions, basic probability assignment values, and model weights (60% for the spatiotemporal attention fusion model, 15% for the random forest model, 15% for the XGBoost model, and 10% for the CNN model).

[0050] In one embodiment, the second step involves identifying and resolving conflicts between different pieces of evidence by calculating the conflict coefficient: using the formula for the conflict coefficient of evidence: (in For different propositions, Calculate the degree of conflict among all standardized evidence bodies (based on the basic probability assignment values ​​of each model for the corresponding proposition). When K > 0.5, it is considered a high conflict. In this case, a weighted correction method is used to resolve the conflict. The conflict coefficient K is allocated to the basic probability value of the corresponding proposition according to the weight proportion of each model. The correction formula is as follows: ( (For model weights); when 0.3≤K≤0.5, it is judged as a medium conflict, and the evidence discount method is adopted, multiplying the basic probability value of the conflicting evidence by a discount factor of 0.8; when K<0.3, it is judged as a low conflict, and the original basic probability allocation value is directly retained to complete the conflict resolution.

[0051] In another embodiment, the third step calculates the overall credibility of the fused evidence using evidence synthesis rules: based on the DS evidence synthesis rules, a joint operation is performed on the standardized evidence body after conflict resolution, and the synthesis formula is as follows: ;in The basic probability assignment values ​​for each model after resolution are given, where A represents the same proposition and 1-K is the normalization factor. The joint basic probability assignment values ​​for the three propositions "Disaster level reaches red alert," "Disaster level reaches orange alert," and "Disaster level reaches yellow or below alert" are calculated separately; these values ​​represent the comprehensive credibility of each proposition. The alert matrix thresholds are optimized based on the comprehensive credibility. When the comprehensive credibility of "Disaster level reaches red alert" is ≥0.6, the red alert trigger threshold in the alert matrix is ​​lowered by 5%; when the comprehensive credibility of "Disaster level reaches yellow or below alert" is ≥0.7, the yellow or below alert trigger threshold is increased by 5%, ensuring accurate matching between the alert matrix thresholds and the multi-source evidence fusion results.

[0052] Optionally, the dual-dimensional early warning matrix constructed by the intelligent decision-making and early warning module sets multi-level threshold intervals, including: each threshold interval corresponds to a unique early warning level; when the probability of disaster occurrence and the degree of impact of the fused multi-source prediction results fall into the corresponding interval, the corresponding early warning level and response measures are automatically matched; the response measures include personnel evacuation instructions and reservoir scheduling suggestion instructions, which are embedded in the hierarchical early warning information.

[0053] In this embodiment, the intelligent decision-making and early warning module constructs a dual-dimensional early warning matrix with the disaster occurrence probability of the fused multi-source prediction results as the horizontal axis and the disaster impact degree as the vertical axis, setting four threshold intervals, corresponding to four unique early warning levels: blue, yellow, orange, and red. Specifically, in one embodiment, the disaster occurrence probability on the horizontal axis is divided into four threshold intervals: 0-30%, 30%-50%, 50%-70%, and above 70%; the disaster impact degree on the vertical axis is determined comprehensively based on the inundated area, the number of affected people, and the risk of infrastructure damage, and is also divided into four threshold intervals. The impact severity values ​​are 0.2 (minor: flooded area <1km², affected population <0.1 million, no damage to core infrastructure), 0.4 (moderate: flooded area 1-5km², affected population 0.1-0.5 million, minor damage to general infrastructure), 0.6 (significant: flooded area 5-10km², affected population 0.5-1 million, partial damage to important infrastructure), and 0.8 (severe: flooded area ≥10km², affected population ≥10,000, severe damage to core infrastructure). Each threshold range is clearly defined by numerical boundaries, with no overlap or omissions.

[0054] In another embodiment, when the disaster occurrence probability and impact value of the fused multi-source prediction results fall into the corresponding threshold range, the system automatically matches the warning level: a probability of 30%-50% and an impact value of 0.2 is a blue warning, a probability of 50%-70% and an impact value of 0.4 is a yellow warning, a probability of 50%-70% and an impact value of 0.6 or above 70% and an impact value of 0.4 is an orange warning, and a probability of above 70% and an impact value of 0.6 or above is a red warning. Simultaneously, corresponding response measures are triggered. The personnel evacuation instructions specify the evacuation scope and time requirements according to the warning level. For a blue warning, evacuate personnel within 100m of the edge of the flooded area (to be completed within 1 hour); for a yellow warning, evacuate personnel within 200m of the flooded area and its surrounding area (to be completed within 40 minutes); for an orange warning, evacuate personnel within 500m of the flooded area and its surrounding area (to be completed within 30 minutes); and for a red warning, evacuate personnel within 1km of the flooded area and its surrounding area (to be completed within 20 minutes). The reservoir scheduling recommendation instructions specify the gate opening rate and pre-discharge flow. During a red warning, the gate opening rate is 2m / min, and the pre-discharge flow is controlled at 10% of the total reservoir capacity. The remaining warning levels decrease proportionally. All response measures are embedded in the graded warning information to form a complete warning content that includes the warning level, threshold range matching results, and response instruction details.

[0055] Most importantly, the intelligent decision-making and early warning module links the hierarchical early warning information with the geographic information system to generate refined evacuation routes. Specifically, this involves: extracting geographic element data of roads, bridges, and residential areas within the early warning area; analyzing road capacity and bridge load-bearing limitations to eliminate impassable road sections; constructing a traffic network model based on the remaining passable road sections; planning evacuation routes from residential areas to safe havens through shortest path planning; simulating and optimizing the planned evacuation routes based on disaster impact; marking evacuation duration and key safe haven nodes to form refined evacuation routes; and pushing hierarchical early warning information and evacuation routes to relevant parties through low-latency transmission.

[0056] In this embodiment, the intelligent decision-making and early warning module links the hierarchical early warning information with the geographic information system to generate refined evacuation routes. Specifically, the geographic information system loads vector map data of the early warning area at a scale of 1:5000, and extracts geographic element data of roads, bridges and settlements. The road data includes attributes such as road width, road material and design speed. The bridge data includes attributes such as span, design load-bearing capacity and construction year. The settlement data includes attributes such as location coordinates, population and building density. All geographic element data are associated with the 2000 National Geodetic Coordinate System.

[0057] In one embodiment, road capacity and bridge load-bearing limits are analyzed. Roads with low capacity are determined by the criteria of road width ≤ 3m and design speed ≤ 20km / h, and bridges with design load ≤ 5t are determined by the criteria of restricted traffic. Combining the flood range data in the fused multi-source prediction results, the traffic segments corresponding to flooded areas, low-capacity roads and restricted bridges are excluded. Roads with a road width ≥ 3m, design speed ≥ 30km / h and not flooded, and bridges with a design load ≥ 5t are retained as passable road segments.

[0058] In another embodiment, a traffic network model is constructed based on the remaining passable road segments. Starting from residential centroids and ending at pre-set refuge points (open areas at an altitude more than 2m above the historical highest inundation level and more than 500m away from the river), road segment weights are calculated according to the road design speed. The shortest path from each residential centroid to the nearest refuge point is planned through node traversal, avoiding road segments with a slope ≥15°. Disaster impact simulations are performed on the planned evacuation routes. Based on the time-series data of inundation depth for the next 0-6 hours, the inundation risk of the routes during the evacuation period is simulated. If a road segment has an inundation risk, the route is adjusted to an adjacent passable road segment. The total length, estimated evacuation time (calculated based on pedestrian walking speed of 5km / h and vehicle driving speed of 30km / h), and key refuge nodes along the route (one temporary refuge point every 2km) are labeled for each evacuation route, forming a refined evacuation path. By using 5G slicing technology, graded early warning information and refined evacuation routes are pushed to relevant targets, with the end-to-end transmission latency controlled within 100ms. The pushed content includes the early warning level, evacuation route vector data, estimated evacuation duration, and coordinates of key risk avoidance nodes.

[0059] Optionally, the ground IoT sensor array is powered by solar energy storage (12V / 100Ah lithium battery, continuous operation for ≥72 hours in cloudy and rainy weather), and the sensor nodes are networked through LoRaWAN protocol (communication distance ≥3km, node capacity ≥100). The raw data is transmitted after being encrypted with AES-128.

[0060] Optionally, the FPGA accelerator of the edge computing preprocessing layer adopts a pipeline architecture to demodulate the Bragg wavelength drift signal of the FBG sensor in real time (demodulation accuracy ±1pm), and realizes collaborative updating of edge nodes and cloud models through a federated learning mechanism (local model parameters are uploaded at 1-hour intervals, and privacy data is not uploaded to the cloud).

[0061] Optionally, the deep learning prediction layer introduces the meta-learning MAML algorithm. After pre-training the model on 10 "source watershed" data (1000 iterations, basic learning rate 0.001), the model converges within 7 days on the "target watershed" containing only 3-5 rainfall data (with a fine-tuned learning rate of 0.0001) through the "gradient descent-parameter fine-tuning" mechanism.

[0062] Preferably, the present invention also provides a deep learning prediction method for flash flood disasters based on integrated air-space-ground perception, comprising the following steps: Step 1: Multi-source data acquisition and preprocessing, synchronously acquiring SAR deformation, LiDAR terrain, FBG sensing and NWP data through the air-space-ground perception layer, and processing them into standardized spatiotemporal tensors through the edge computing layer, wherein the SAR data is converted from a 12-day interval to a 1-hour scale through cubic spline interpolation; Step 2: Feature engineering construction, extracting 32-dimensional feature vectors, including terrain features (slope, curvature, runoff accumulation), hydrological features (soil saturated water content, pore water pressure, permeability coefficient), and meteorological features (rainfall return period, convective effective potential energy CAPE, 1-hour rainfall intensity I1h), and processing them through max-min normalization (range [0,1]); Step 3: Model training and optimization, using the Adam optimizer (β1=0.9, β2=0.999) to train ST-Atte The ntionT-LSTM model introduces an early-stop mechanism (patience=20) to prevent overfitting. Key influencing factors (I1h and the weight of the previous soil moisture index API > 60%) are identified through Shapley value analysis, and physical constraint terms are embedded to correct the model parameters. Step 4: Prediction and early warning generation. The model outputs the probability distribution of the peak flow (error < 8%) and inundation depth (error < 0.3m) for the next 0-6 hours. The expected loss value is calculated by combining the vulnerability curve. When P (disaster) > 70% and the loss index > 0.6, a red warning is triggered and the emergency response plan is activated (including reservoir pre-discharge scheduling instructions and gate opening rate of 2m / min).

[0063] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0064] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A flash flood disaster early warning system based on deep learning and integrated air-space-ground perception, characterized in that, Includes the following modules: The system comprises a space-air-ground sensing module for collecting multi-source data, including: surface deformation data acquired via spaceborne synthetic aperture radar, high-precision terrain data acquired via UAV lidar, hydrological sensing data acquired via ground-based IoT distributed sensors, and meteorological numerical forecast data acquired via a meteorological data interface. An edge computing preprocessing module receives the multi-source data output from the space-air-ground sensing module, performs spatiotemporal registration, denoising, and missing data reconstruction, and outputs a standardized spatiotemporal tensor dataset. A deep learning prediction module receives the standardized spatiotemporal tensor dataset output from the edge computing preprocessing module, extracts data features from the standardized spatiotemporal tensor dataset using a spatiotemporal attention fusion model with embedded pre-defined physical constraints, and predicts flash flood parameters. An intelligent decision-making and early warning module receives the flash flood parameter prediction results output from the deep learning prediction module, fuses the flash flood parameter prediction results based on DS evidence theory to construct a two-dimensional early warning matrix, generates tiered early warning information, and links the tiered early warning information with the geographic information system to generate refined evacuation routes.

2. The flash flood disaster early warning system based on deep learning and integrated air-space-ground perception as described in claim 1, characterized in that, When collecting multi-source data, the aforementioned space-air-ground sensing module adopts a three-level collaborative mode involving spacecraft, aircraft, and ground. Specifically, it uses spaceborne synthetic aperture radar to periodically observe the entire watershed, recording displacement changes at different locations on the surface to form continuous temporal data of surface deformation; it uses UAV lidar to cover the watershed terrain along a preset route, generating high-precision terrain data including topographic undulations and valley distribution through high-density scanning; it uses distributed ground-based IoT sensors to collect hydrological sensing data composed of rainfall, soil moisture content, and microseismic signals, distributed according to the watershed terrain characteristics and monitoring needs; it uses a meteorological data interface to access regional meteorological forecast data through standardized protocols to extract rainfall-related numerical meteorological forecast data; and it associates and labels data according to the collection source, timestamp, and spatial coordinates to form a structured multi-source dataset.

3. The flash flood disaster early warning system based on deep learning and integrated air-space-ground perception as described in claim 1, characterized in that, After receiving multi-source data from the air-space-ground sensing module, the edge computing preprocessing module performs processing operations according to the process of spatiotemporal unification, purification and supplementation, and format standardization. Specifically, it aligns the timestamps of the multi-source data through time synchronization, unifies the spatial coordinates of data from different sources to the same coordinate system, and completes spatiotemporal registration; it separates the effective signal and noise components in the original sensor data through signal decomposition to remove noise interference; it supplements missing data through adjacent data correlation analysis; and it regularizes the processed multi-source data according to a fixed time step and spatial grid scale, converts it into a standardized spatiotemporal tensor dataset with a unified format, and outputs it.

4. The flash flood disaster early warning system based on deep learning and integrated air-space-ground perception as described in claim 1, characterized in that, After receiving the standardized spatiotemporal tensor dataset output by the edge computing preprocessing module, the deep learning prediction module extracts data features and predicts flash flood disaster parameters through a spatiotemporal attention fusion model with embedded preset physical constraints. Specifically, it determines the feature extraction unit and parameter prediction unit in the spatiotemporal attention fusion model; performs hierarchical processing on the standardized spatiotemporal tensor dataset through the feature extraction unit, separates the single features of four types of data (surface deformation, topography, hydrology, and meteorology), and determines the cross-correlation features between different types of data. The parameter prediction unit is used to construct prediction logic based on cross-correlation features. Preset physical constraints are embedded to calibrate the output results, and preliminary mountain flood disaster parameter prediction results are output.

5. The flash flood disaster early warning system based on deep learning and integrated air-space-ground perception as described in claim 3, characterized in that, When the spatiotemporal attention fusion model extracts data features, it sets up a dual attention channel, including: identifying the spatial location of confluence nodes and weak slope areas within the watershed based on the spatiotemporal attention fusion model, and assigning priority extraction weights to the data features of spatial locations; and identifying the time periods corresponding to short-term heavy rainfall and sudden changes in surface deformation based on the spatiotemporal attention fusion model, and determining the influence factors of the time periods on disaster prediction.

6. The flash flood disaster early warning system based on deep learning and integrated air-space-ground perception according to claim 4, characterized in that, The preset physical constraints include constraints related to water flow patterns and soil infiltration patterns. Specifically, the preset physical constraints are embedded into the loss function of the spatiotemporal attention fusion model at a fixed ratio. When the prediction result deviates from the preset physical laws by more than a preset range, a parameter correction mechanism is automatically triggered. The feature extraction weights and prediction logic of the spatiotemporal attention fusion model are adjusted so that the prediction result conforms to the preset watershed hydrophysical evolution laws.

7. The flash flood disaster early warning system based on deep learning and integrated air-space-ground perception as described in claim 5, characterized in that, Before the deep learning prediction module predicts flash flood disaster parameters using the spatiotemporal attention fusion model, it performs multiple rounds of iterative training and optimization on the spatiotemporal attention fusion model. Specifically, the first round is based on the full feature vector for basic training. In each subsequent round of training, redundant features are eliminated through feature importance analysis, and the core features that affect disaster prediction are retained. The physical constraint weights and attention channel parameters are adjusted. After each round of training, the model accuracy is verified using a validation dataset until the model prediction error drops to a preset range.

8. The flash flood disaster early warning system based on deep learning and integrated air-space-ground perception according to claim 4, characterized in that, When the spatiotemporal attention fusion model predicts parameters for flash floods, it proceeds in layers according to the time dimension, including: first predicting the probability of flash flood occurrence and preliminary flood peak parameters during the 0-3 hour period; then predicting the disaster development trend and inundation range during the 3-6 hour period; and forming a complete prediction chain based on the prediction results of different time periods during the 0-6 hour period, outputting a flash flood prediction report containing multiple time periods and multiple parameters.

9. The flash flood disaster early warning system based on deep learning and integrated air-space-ground perception as described in claim 1, characterized in that, After receiving the flash flood disaster parameter prediction results output by the deep learning prediction module, the intelligent decision-making and early warning module constructs a two-dimensional early warning matrix based on the DS evidence theory by fusing the flash flood disaster parameter prediction results. Specifically, the multi-model fusion weights are determined according to preset rules, and the credibility analysis of the prediction results of the spatiotemporal attention fusion model is performed through the DS evidence theory to obtain credibility data. The prediction results are fused into multi-source prediction results based on the credibility data; Based on multi-source prediction results, an early warning matrix is ​​constructed with the probability of disaster occurrence and the degree of disaster impact as the core dimensions. Early warning levels are divided according to the threshold of the early warning matrix, and graded early warning information is generated.

10. The flash flood disaster early warning system based on deep learning and integrated air-space-ground perception according to claim 9, characterized in that, The intelligent decision-making and early warning module constructs a dual-dimensional early warning matrix with multi-level threshold intervals, including: each threshold interval corresponds to a unique early warning level; when the probability of disaster occurrence and the degree of impact of the fused multi-source prediction results fall into the corresponding interval, the corresponding early warning level and response measures are automatically matched; the response measures include personnel evacuation instructions and reservoir scheduling suggestion instructions, which are embedded in the hierarchical early warning information.

Images