Mountain torrent disaster assessment method and system based on unmanned aerial vehicle and brain-like pulse network

Abstract

The invention relates to a mountain torrent disaster assessment method and system based on an unmanned aerial vehicle and a brain-like pulse network, and belongs to the technical field of disaster prediction. The method comprises the following steps: collecting multi-source remote sensing data, infrared soil humidity data and rainfall data of a research area, preprocessing the multi-source data, constructing a brain-like pulse network SNN, and training an SNN model by using the collected historical unmanned aerial vehicle remote sensing data and field disaster survey data and taking the spatial-temporal characteristics of preprocessing the historical data as input and the mountain torrent disaster risk level as output. Inputting the preprocessed real-time multi-unmanned aerial vehicle data into the trained SNN model, and outputting the risk level of each space grid at the moment t by the model; and eliminating isolated abnormal values of the risk levels of the adjacent grids by adopting spatial autocorrelation analysis to obtain a mountain torrent risk spatial distribution diagram of the research area. The method realizes wide coverage, low energy consumption and high precision (the accuracy is greater than or equal to 92%) of mountain torrent assessment, and is suitable for disaster prevention and reduction decision-making in remote mountainous areas.

Description

Technical Field

[0001] This invention relates to a method and system for assessing flash flood disasters based on unmanned aerial vehicles and brain-like pulse networks, belonging to the field of disaster prediction technology. Background Technology

[0002] Flash floods are characterized by their suddenness, uneven spatial distribution, and great destructive power; therefore, the accuracy and real-time nature of their assessment directly determine the efficiency of disaster prevention and mitigation decisions. Existing flash flood assessment methods have the following key limitations:

[0003] 1. Insufficient data acquisition coverage: Traditional methods rely on fixed ground hydrological stations (such as water level stations and rain gauge stations) and satellite remote sensing data. The former has a low deployment density in mountainous and remote areas with complex terrain, making it difficult to capture local micro-scale hydrological features (such as local steep slope confluence and small valley water accumulation); the latter is affected by revisit period (such as Landsat satellite revisit period of 16 days) and cloud cover, and cannot meet the real-time assessment needs of flash floods at the "minute-hour" level.

[0004] 2. Low efficiency of multi-source data processing: Existing assessments often use a single data source (such as only rainfall data or optical remote sensing data), and do not fully integrate multimodal remote sensing information such as "optical + radar + infrared" (e.g., optical data reflects vegetation cover, radar data penetrates the cloud layer to obtain terrain, and infrared data monitors soil moisture); moreover, traditional deep learning models (such as CNN and LSTM) are continuous numerical calculations, which have high energy consumption and large computational latency when deployed at the edge nodes of UAVs, making it difficult to adapt to the limited computing power and endurance resources of UAVs.

[0005] 3. Weak ability to model dynamic spatiotemporal features: Flash flood evolution is a typical spatiotemporal dynamic process (such as the temporal correlation of rainfall-confluence-inundation, and the spatial differences in flow velocity / water depth in different regions). Traditional models mostly focus on static features (such as historical inundation range) and lack accurate modeling of "spatiotemporal coupling features". Moreover, it does not take advantage of the biological realism of brain-like spiking networks - SNNs encode spatiotemporal information with discrete pulse sequences, and the energy consumption is only 1 / 100 to 1 / 10 of that of traditional neural networks. Furthermore, it can efficiently learn the dynamic process laws through the pulse time-dependent plasticity (STDP) rule, but it has not been applied to flash flood assessment scenarios.

[0006] To address the aforementioned issues, there is an urgent need for a flash flood disaster assessment method that integrates the wide-coverage, multimodal data acquisition capabilities of multi-UAV collaborative remote sensing with the low-energy, high-real-time spatiotemporal modeling capabilities of brain-like pulse networks, thereby achieving an end-to-end process of "real-time data acquisition - efficient feature extraction - accurate risk assessment". Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a flash flood disaster assessment method based on multi-UAV remote sensing and brain-like pulse networks, so as to achieve wide coverage, real-time, low energy consumption and high precision assessment of flash flood risk.

[0008] The technical solution adopted in this invention is as follows: A flash flood disaster assessment method based on unmanned aerial vehicles and neuromorphic spur networks includes the following steps: S1. An unmanned aerial vehicle (UAV) cluster equipped with optical remote sensing sensors, synthetic aperture radar (SAR) remote sensing sensors, infrared sensors, and rainfall sensors was used to collect multi-source remote sensing data, infrared soil moisture data, and rainfall data for the study area. Data timestamp synchronization was achieved through the UAV communication protocol, and the geographic coordinates, acquisition time, and sensor type of each frame of data were labeled to form the raw dataset. , i is the UAV number, j is the sensor type, and t is the data acquisition time; S2. Multi-source data preprocessing: The data is corrected to eliminate noise and systematic errors, and then pixel-level and feature-level fusion is performed to form a standardized dataset. The multimodal feature vector is defined by time t and spatial location p. S3. Construct a brain-like spiking network (SNN) based on leakage integral-ignition neurons. Using historical UAV remote sensing data and on-site disaster investigation data, with the spatiotemporal characteristics of the preprocessed historical data as input and the flash flood disaster risk level as output, train the SNN model and optimize the model. S4. Process the pre-processed real-time multi-UAV data Input the trained SNN model, and the model outputs the risk level of each spatial grid at time t. For the risk levels of adjacent grids, spatial autocorrelation analysis was used to eliminate isolated outliers, resulting in a spatial distribution map of flash flood risk in the study area.

[0009] The aforementioned flash flood disaster assessment method also includes a step of dynamic model updating, which involves periodically collecting new flash flood event data and UAV remote sensing data, and using an incremental learning strategy to update the synaptic weights of the SNN.

[0010] The aforementioned flash flood disaster assessment method also includes steps for outputting risk level maps, key risk indicators, and early warning recommendations.

[0011] In the above method, the pixel-level fusion in step S2 involves fusing optical remote sensing data and SAR remote sensing data of the same sub-region through wavelet transform, preserving the vegetation texture of the optical data and the topographic relief features of the SAR data; the feature-level fusion involves spatiotemporally matching infrared soil moisture and rainfall data with the pixel-level fused topographic data to form a standardized dataset. It is a multimodal feature vector at time t and spatial location p, containing slope, water system density, vegetation cover, soil moisture, and rainfall intensity.

[0012] The neuromorphic spiking network (SNN) mentioned in step S3 specifically includes: Pulse code module: will The continuous features are converted into discrete pulse sequences. Poisson coding is applied to each feature dimension to obtain a set of pulse sequences. k is the feature dimension. A pulse sequence with time t, position p, and dimension k; Input layer: The number of neurons is set to the number of feature dimensions × the number of spatial grids. Each neuron corresponds to a pulse sequence of one feature dimension, realizing a one-to-one mapping between a single feature dimension and a neuron. Hidden layer: LIF neurons are used, with the number being 1 / 10 of that in the input layer. The synaptic weights are adjusted using the STDP learning rule (a Heblin-type learning rule that adjusts synaptic weights based on the relative timing of impulses) to extract spatiotemporal coupling features. Output layer: Three neurons are set up, corresponding to "low risk", "medium risk" and "high risk" respectively, and the level is determined by the pulse firing frequency.

[0013] Step S3 employs a pulse time error minimization optimization model, which minimizes the time difference between the model's output pulse sequence and the standard pulse sequence corresponding to the labeled risk level.

[0014] Another objective of this invention is to provide a flash flood disaster assessment system based on unmanned aerial vehicles and neuromorphic spur networks, comprising: The multi-source data acquisition module employs a drone array equipped with optical remote sensing sensors, synthetic aperture radar (SAR) remote sensing sensors, infrared sensors, and rainfall sensors to collect multi-source remote sensing data, infrared soil moisture data, and rainfall data for the study area. Data timestamp synchronization is achieved through the drone communication protocol, labeling each frame with its geographic coordinates, acquisition time, and sensor type to form the raw dataset. , i is the UAV number, j is the sensor type, and t is the data acquisition time; The multi-source data preprocessing module corrects the data to eliminate noise and systematic errors, and then performs pixel-level fusion and feature-level fusion to form a standardized dataset. The multimodal feature vector is defined by time t and spatial location p. The brain-like spiking network (SNN) model construction module constructs a brain-like spiking network (SNN) based on leakage integral-ignition neurons. It uses historical UAV remote sensing data and field disaster investigation data, with the spatiotemporal characteristics of the preprocessed historical data as input and the flash flood disaster risk level as output, to train and optimize the SNN model. The flash flood disaster prediction module will process pre-processed real-time multi-UAV data. Input the trained SNN model, and the model outputs the risk level of each spatial grid at time t. For the risk levels of adjacent grids, spatial autocorrelation analysis was used to eliminate isolated outliers, resulting in a spatial distribution map of flash flood risk in the study area.

[0015] The aforementioned flash flood disaster assessment system based on UAVs and brain-like spiking networks also includes an optimized deployment module for the SNN model, which deploys the trained SNN model at the edge nodes of the UAV.

[0016] The aforementioned flash flood disaster assessment system based on UAVs and brain-like spiking networks also includes a dynamic update module, which periodically collects new flash flood event data and UAV remote sensing data, and updates the synaptic weights of the SNN using an incremental learning strategy.

[0017] The beneficial effects of this invention are: 1. Wide data coverage and high real-time performance: Multi-UAV collaborative data collection overcomes the limitations of fixed stations and satellite remote sensing, achieving full coverage from "macro" to "micro" perspectives, with a data sampling frequency of 5-10 minutes per time, meeting the needs of flash flood emergency assessment; 2. Low energy consumption and strong edge adaptability: The neuromorphic spiking network encodes information with discrete pulses, and its energy consumption is only 1 / 50 to 1 / 10 of that of traditional CNNs. It can be directly deployed on the edge nodes of drones, avoiding the delay of data transmission to the cloud. 3. High assessment accuracy: Through multimodal data fusion and spatiotemporal feature modeling of SNN, the accuracy of risk level determination is ≥92%, and the positioning error of high-risk areas is ≤50m, which is better than traditional methods (accuracy ≤85%, error ≥100m). 4. Highly practical: The output risk maps, key indicators and early warning suggestions can directly provide decision support for emergency management departments, and are especially suitable for remote mountainous areas without fixed monitoring stations. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This diagram illustrates the collaborative deployment and data collection of multiple drones. Detailed Implementation

[0019] The present invention will be further described below with reference to specific embodiments.

[0020] Example 1: A flash flood disaster assessment method based on unmanned aerial vehicles and neuromorphic spur networks, comprising the following steps: S1. A drone team equipped with optical remote sensing sensors, synthetic aperture radar (SAR) remote sensing sensors, infrared sensors, and rainfall sensors was used to collect multi-source remote sensing data, infrared soil moisture data, and rainfall data for the study area. Data timestamp synchronization was achieved through the drone communication protocol, and the geographic coordinates, acquisition time, and sensor type of each frame of data were labeled to form the original dataset. UAV swarm deployment: Based on the terrain of the study area (e.g., mountain valleys, watershed boundaries), divide the area into M monitoring sub-areas and deploy N UAVs (N≥3, including multi-rotor UAVs and fixed-wing UAVs). Fixed-wing UAV (endurance ≥ 2 hours): Equipped with optical sensors (such as a 5-megapixel RGB camera with a resolution of 0.5m) and synthetic aperture radar (SAR, resolution 1m, penetrating clouds), responsible for collecting terrain, water system, and vegetation cover data over a large area (such as a sub-region area ≥ 10km²), with a sampling frequency of 1 time / 10 minutes. Multi-rotor UAV (flexible and maneuverable): Equipped with an infrared sensor (monitoring soil moisture, accuracy ±2%) and a rain sensor (collecting local rainfall intensity, accuracy ±0.1mm / h), it is responsible for micro-scale data collection in small areas (such as steep slopes or the area around villages within a sub-region), with a sampling frequency of 1 time / 5 minutes; Data synchronization and annotation: Data timestamp synchronization is achieved through drone swarm communication protocols (such as MQTT), and the geographic coordinates (latitude, longitude, and elevation), acquisition time, and sensor type of each frame of data are annotated to form the original dataset. (i is the drone number, j is the sensor type, and t is the data acquisition time).

[0021] S2. Multi-source data preprocessing: right Preprocessing is performed to eliminate noise and system errors, and to unify the data format and spatial resolution (set to 0.5m).

[0022] Sensor calibration: Optical remote sensing data were subjected to radiometric correction (to eliminate the effects of atmospheric scattering) and geometric correction (based on ground control points GCP, with a positioning error ≤1m) using ENVI software. SAR remote sensing data were phase-corrected using a range-Doppler model to remove speckle noise (using Lee filtering). Infrared data, based on temperature calibration using a blackbody radiation source, is converted into soil volumetric water content (using a soil dielectric constant model).

[0023] Data fusion: A two-level fusion strategy of "pixel-level fusion + feature-level fusion" is adopted. Pixel-level fusion: Optical remote sensing and SAR remote sensing data of the same sub-region are fused through wavelet transform to preserve the vegetation texture of the optical remote sensing data and the topographic relief features of the SAR remote sensing data. Feature-level fusion: Infrared soil moisture and rainfall data are spatiotemporally matched with fused topographic data to form a standardized dataset. It is a multimodal feature vector at time t and spatial location p, containing slope, water system density, vegetation cover, soil moisture, and rainfall intensity.

[0024] S3. Construct a brain-like spiking network (SNN) based on leakage integral-ignition neurons. Utilize historical UAV remote sensing data and on-site disaster survey data, using the preprocessed spatiotemporal features of historical data as input and the flash flood disaster risk level as output, train the SNN model and optimize the model: Construct an SNN model based on Leakage Integral-Ignition (LIF) neurons, inputting... Based on the spatiotemporal characteristics, the risk level of flash flood disaster (low / medium / high) is output, specifically including: Pulse code module: will The continuous features are converted into discrete pulse sequences. For each feature dimension (such as slope α, rainfall intensity R), Poisson coding is used: the feature value is set to be positively correlated with the pulse emission probability, that is, the larger the feature value, the more pulses are emitted per unit time (e.g., rainfall intensity). At that time, the pulse frequency is ≥10Hz; (When the pulse frequency is ≤2Hz); after encoding, a pulse sequence set is obtained. (k is the feature dimension, (A pulse sequence with time t, position p, and dimension k).

[0025] SNN network structure: It adopts a three-layer structure of "input layer - hidden layer - output layer". Input layer: Number of neurons = Number of feature dimensions × Number of spatial grids (e.g., if the study area is divided into 1000×1000 grids, with a feature dimension of 5, then the number of neurons in the input layer = 5×10). 6 Each neuron corresponds to a pulse sequence of one feature dimension; Hidden layer: Employs LIF neurons (membrane time constant) Threshold voltage The number of layers is 1 / 10 of the input layer. The synaptic weights are adjusted by the STDP learning rule (the weights increase when the presynaptic pulse is fired before the postsynaptic pulse, and decrease otherwise), and spatiotemporal coupling features (such as the synergistic effect of "steep slope + high rainfall") are extracted. Output layer: 3 neurons, corresponding to "low risk", "medium risk" and "high risk" respectively. The level is determined by the pulse firing frequency (e.g., high risk neuron pulse frequency ≥ 8Hz, medium risk 4~8Hz, low risk < 4Hz).

[0026] Model training: Training dataset: Historical flash flood disaster case data (such as multi-source data collected by drones from a flash flood event in a mountainous area from 2018 to 2023 + on-site disaster investigation data), labeled with risk levels; Loss function: Pulse time error (PTE) is used to minimize the time difference between the model output pulse sequence and the standard pulse sequence corresponding to the labeled risk level; Deployment optimization: Deploy SNN at the edge node of the drone (such as a drone equipped with NVIDIA Jetson Nano) to reduce power consumption through quantization compression (8-bit weight precision) and achieve inference latency ≤100ms.

[0027] S4. Process the pre-processed real-time multi-UAV data Input the trained SNN model, and the model outputs the risk level of each spatial grid at time t. Spatial autocorrelation analysis was used to eliminate isolated outliers for the risk levels of adjacent grids, resulting in a spatial distribution map of flash flood risk in the study area. Real-time inference: This involves processing pre-processed real-time multi-UAV data. Input the trained SNN model, and the model outputs the risk level of each spatial grid p at time t. ; Risk fusion: Spatial autocorrelation analysis (Moran index) is used to eliminate isolated outliers for the risk levels of adjacent grids, resulting in a spatial distribution map of flash flood risk in the study area; Output results: Three types of information will be output. Risk level map: Displays "low / medium / high" risk areas in the form of a heat map, and marks the core area of ​​high-risk areas (such as areas with a flooding probability ≥80%). Key indicators: area of ​​high-risk zone, number of affected villages, and estimated maximum water depth (based on SNN output and historical water depth-risk correlation model). Warning Recommendations: For high-risk areas, evacuation routes are recommended (based on terrain slope and water system distribution).

[0028] To address the need for calculating the "estimated maximum water depth" indicator, a historical water depth-risk correlation model is constructed. By using the risk level and multimodal features output by the SNN, the estimated maximum water depth of the corresponding area is back-mapped, solving the quantitative connection problem between "risk level" and "specific water depth" and improving the practicality of the indicator.

[0029] The model outputs a risk level (low / medium / high) through an SNN, and measures multimodal baseline features (slope, soil moisture, rainfall intensity, and water system density, from...). The system outputs the estimated maximum water depth (unit: m, accuracy ±0.1 m) of the target area, along with historical water depth labels obtained from the survey (maximum water depth of flash flood events in the field survey). Model Structure and Training Process: The Gradient Boosting Regression Tree (GBRT) model is employed. This model can adapt to the non-linear mapping relationship between "features" and "water depth," and can output the importance of features (such as the contribution weight of rainfall intensity to water depth), facilitating result interpretation. Multiple regression trees are trained iteratively, with each tree fitting the residuals of the preceding model. Finally, a weighted sum is calculated to obtain the predicted water depth value. The specific process is as follows: Historical data: Samples of flash flood events from 2018 to 2023, based on “risk level + multimodal characteristics + actual water depth”, were compiled and divided into a training set (70%) and a test set (30%) in a 7:3 ratio. Data augmentation: For scarce samples (such as high-risk, deep water samples), the "feature perturbation method" is used to augment them (e.g., fine-tuning within the range of rainfall intensity ±5mm / h and slope ±2° while keeping the water depth label unchanged). Key parameters: number of trees (100-200), learning rate (0.05-0.1), tree depth (5-8 layers). The optimal combination is selected through 5-fold cross-validation (with the goal of minimizing the root mean square error (RMSE) of the test set). Performance metrics: The test set RMSE should be ≤0.15m (for example, when the actual water depth is 1.2m, the prediction error should be ≤0.15m), and the coefficient of determination R^2 should be ≥0.85 (indicating that the model can explain more than 85% of the water depth variation).

[0030] To address the recommendation requirements for "evacuation routes in high-risk areas," a shortest safe path model constrained by terrain and water systems is constructed. This model calculates the optimal route from affected villages to safe high ground while avoiding dangerous areas (such as steep slopes and areas near water systems), ensuring both route safety and timeliness. The required parameter types for the model include raster data of the terrain, masked raster data of dangerous areas, and coordinates of the start and end points.

[0031] By acquiring basic topographic data of the study area and establishing a digital elevation model (DEM, resolution 0.5m, from SAR data processing results), the raster data of the terrain (including elevation and slope values ​​for each raster) can be obtained.

[0032] By masking dangerous areas, such as high-risk flooding areas (SNN output L_t(p) = high-risk grid) and dangerous terrain areas (areas with a slope ≥ 25°, areas ≤ 50m from water systems), a binary grid (1 = dangerous, 0 = safe) can be obtained. The coordinates of the proposed evacuation start and end points in the danger zone (start point: latitude and longitude coordinates of the residential gathering point of the affected village, end point: safe high ground (areas with a slope of <10° and an elevation more than 1m above the historical maximum flood level, pre-marked in GIS) can be obtained through the Geographic Information System (GIS).

[0033] The model has the following core constraints: Safety constraints apply: routes must not cross hazardous areas (grids with hazardous area mask = 1). Time constraints apply: the total route length is ≤3km (ensuring residents can reach the destination on foot within 30 minutes, meeting the time requirements for emergency evacuation during flash floods). Slope constraints: The average slope of the route is ≤15° (avoiding steep slopes that are difficult for the elderly and children to pass through).

[0034] The model can adopt an improved A* algorithm, which introduces a "safety cost" weight on the basis of the traditional A* algorithm's "shortest path" objective, and prioritizes the selection of safe areas (such as areas far away from water systems and low slope areas), taking into account both path length and safety.

[0035] Design a multi-factor cost function F(n) = G(n) + H(n) + C(n), where: G(n): The actual path length from the starting point to the current node n (unit: m); H(n): The estimated straight-line distance from the current node n to the endpoint (heuristic function to ensure algorithm efficiency); C(n): The safety cost of the current node n (C(n)=5 when the slope is ≥15°, C(n)=3 when the distance from the water system is 20-50m, C(n)=0 in the safe zone, and the higher the cost, the higher the priority to avoid it).

[0036] The route generation steps are as follows: Dangerous area masking: Dangerous areas are masked in the DEM raster map (dangerous grates are set to "impassable"), leaving only safe areas as path search space; Node initialization: Set the gathering point of the affected villages as the starting point S, the safe high ground as the ending point T, and initialize the cost function of the starting point F(S)=0; Path search: Iteratively expand the adjacent safe grids of the current node, calculate the F(n) value of each grid, and select the grid with the smallest F(n) as the next search node; When the destination T is found, backtrack the path nodes to generate a complete route from S to T. Route optimization: The generated route is "smoothed" (redundant inflection points are removed, and detours are reduced); Mark key information about the route: total length (e.g., 1.5km), estimated walking time (e.g., 15 minutes), and key points along the way (e.g., "passing XX Bridge, 30m away from waterways").

[0037] The final output includes both text description and GIS map annotations for early warning suggestions.

[0038] The text description reads: "Residents of XX village are advised to evacuate to the northwest. The total route is 1.5km long, passing through XX plot (slope 8°), and finally reaching XX highland (elevation 520m, high safety level). The estimated walking time is 15 minutes, avoiding dangerous areas within 20m of waterways." GIS Map: Evacuation routes (solid red lines), starting points (red dots), ending points (green dots), and danger zones (red shaded areas) are overlaid on the risk level heat map to facilitate intuitive deployment by emergency departments.

[0039] S5. Dynamic Model Update: Regularly (e.g., quarterly) collect new flash flood event data and UAV remote sensing data, and use an incremental learning strategy to update the synaptic weights of the SNN (based on STDP rule fine-tuning, without retraining the entire model) to improve the model's adaptability to new terrain and climate conditions.

[0040] To further illustrate this, let's take the flash flood disaster assessment data of a certain mountainous watershed as an example: (1) Experimental area and UAV deployment, Regional Overview: A watershed in a mountainous area in Southwest my country was selected (area 500 km², including 3 villages, terrain slope 0°~60°, average number of flash floods 2~3 times per year). Drone swarm: Deploy 5 drones, including 2 fixed-wing drones (equipped with RGB cameras + SAR) and 3 multi-rotor drones (equipped with infrared sensors + rain sensors), divided into 5 monitoring sub-areas (each area is 100km²).

[0041] (2) Data acquisition and preprocessing, Data collection period: July 2024 (rainy season), continuous collection for 72 hours, with fixed-wing UAVs collecting data once every 10 minutes and multi-rotor UAVs collecting data once every 5 minutes; Preprocessing results: After geometric correction, the positioning error of optical data is 0.8m, the SAR data speckle noise removal rate is ≥85%, the soil moisture inversion error is ±1.5%, the resolution of fused data is 0.5m, and the spatiotemporal matching accuracy is ≤1 minute / 1m.

[0042] (3) SNN model training and evaluation, Training data: UAV data (a total of 100,000 frames) from three flash flood events in the basin from 2018 to 2023 were used and labeled as "low risk" (inundation area < 1km²), "medium risk" (1~5km²), and "high risk" (>5km²). Model parameters: LIF neuron membrane time constant 20ms, threshold voltage 15mV, STDP learning rate 0.01, number of input layer neurons 5×(1000×1000)=5×10 6 Hidden layer 5×10 5 Output layer 3; Evaluation results: Risk level determination accuracy was 93.2%, high-risk area positioning error was 42m, edge node inference latency was 85ms, and energy consumption was 1.2W (traditional CNN energy consumption is 60W).

[0043] (4) Real-time evaluation results, At 14:00 on July 15, 2024, the rainfall intensity in sub-region 3 reached 65 mm / h, the soil moisture was 45%, and the area with a slope of ≥30° accounted for 35%; SNN output: Sub-region 3 is "high risk", with a core area of ​​2.8 km², involving 1 village, and an estimated maximum water depth of 1.2 m; Warning Recommendation: It is recommended that residents of the village evacuate to higher ground with a slope of less than 10° in the northwest direction, with an evacuation route of 1.5km in length.

[0044] Example 2: A flash flood disaster assessment system based on unmanned aerial vehicles and brain-like spur networks, comprising: The multi-source data acquisition module employs a drone array equipped with optical remote sensing sensors, synthetic aperture radar (SAR) remote sensing sensors, infrared sensors, and rainfall sensors to collect multi-source remote sensing data, infrared soil moisture data, and rainfall data for the study area. Data timestamp synchronization is achieved through the drone communication protocol, labeling each frame with its geographic coordinates, acquisition time, and sensor type to form the raw dataset. , i is the UAV number, j is the sensor type, and t is the data acquisition time; The multi-source data preprocessing module corrects the data to eliminate noise and systematic errors, and then performs pixel-level fusion and feature-level fusion to form a standardized dataset. The multimodal feature vector is defined by time t and spatial location p. The brain-like spiking network (SNN) model construction module constructs a brain-like spiking network (SNN) based on leakage integral-ignition neurons. It uses historical UAV remote sensing data and field disaster investigation data, with the spatiotemporal characteristics of the preprocessed historical data as input and the flash flood disaster risk level as output, to train and optimize the SNN model. The flash flood disaster prediction module will process pre-processed real-time multi-UAV data. Input the trained SNN model, and the model outputs the risk level of each spatial grid at time t. Spatial autocorrelation analysis was used to eliminate isolated outliers for the risk levels of adjacent grids, resulting in a spatial distribution map of flash flood risk in the study area. The dynamic update module periodically collects new flash flood event data and UAV remote sensing data, and uses an incremental learning strategy to update the synaptic weights of the SNN.

[0045] The aforementioned flash flood disaster assessment system based on UAVs and brain-like spiking networks also includes an optimized deployment module for the SNN model, which deploys the trained SNN model at the edge nodes of the UAV.

[0046] The above is a further description of the present invention in conjunction with the embodiments, and the scope of protection of the present invention is not limited thereto.

Claims

1. A flash flood disaster assessment method based on unmanned aerial vehicles and neuromorphic spur networks, characterized by: The steps include the following: S1. An unmanned aerial vehicle (UAV) cluster equipped with optical remote sensing sensors, synthetic aperture radar (SAR) remote sensing sensors, infrared sensors, and rainfall sensors was used to collect multi-source remote sensing data, infrared soil moisture data, and rainfall data for the study area. Data timestamp synchronization was achieved through the UAV communication protocol, and the geographic coordinates, acquisition time, and sensor type of each frame of data were labeled to form the raw dataset. , i is the UAV number, j is the sensor type, and t is the data acquisition time; S2. Multi-source data preprocessing: The data is corrected to eliminate noise and systematic errors, and then pixel-level and feature-level fusion is performed to form a standardized dataset. The multimodal feature vector is defined by time t and spatial location p. S3. Construct a brain-like spiking network (SNN) based on leakage integral-ignition neurons. Using historical UAV remote sensing data and on-site disaster investigation data, with the spatiotemporal characteristics of the preprocessed historical data as input and the flash flood disaster risk level as output, train the SNN model and optimize the model. S4. Process the pre-processed real-time multi-UAV data Input the trained SNN model, and the model outputs the risk level of each spatial grid at time t. For the risk levels of adjacent grids, spatial autocorrelation analysis was used to eliminate isolated outliers, resulting in a spatial distribution map of flash flood risk in the study area.

2. The flash flood disaster assessment method based on unmanned aerial vehicles and neuromorphic impulse networks according to claim 1, characterized in that, The aforementioned flash flood disaster assessment method also includes a step of dynamic model updating, which involves periodically collecting new flash flood event data and UAV remote sensing data, and using an incremental learning strategy to update the synaptic weights of the SNN.

3. The flash flood disaster assessment method based on unmanned aerial vehicles and neuromorphic impulse networks according to claim 1, characterized in that, The pixel-level fusion described in step S2 involves fusing optical remote sensing data and SAR remote sensing data of the same sub-region using wavelet transform, preserving the vegetation texture of the optical data and the topographic relief features of the SAR data. The feature-level fusion involves spatiotemporally matching infrared soil moisture and rainfall data with the pixel-level fused topographic data to form a standardized dataset. It is a multimodal feature vector at time t and spatial location p, containing slope, water system density, vegetation cover, soil moisture, and rainfall intensity.

4. The flash flood disaster assessment method based on unmanned aerial vehicles and neuromorphic impulse networks according to claim 1, characterized in that, The neuromorphic spiking network (SNN) mentioned in step S3 specifically includes: Pulse code module: will The continuous features are converted into discrete pulse sequences. Poisson coding is applied to each feature dimension to obtain a set of pulse sequences. k is the feature dimension. A pulse sequence with time t, position p, and dimension k; Input layer: The number of neurons is set to the number of feature dimensions × the number of spatial grids, and each neuron corresponds to a pulse sequence of one feature dimension; Hidden layer: LIF neurons are used, with the number being 1 / 10 of that in the input layer. Synaptic weights are adjusted using the STDP learning rule to extract spatiotemporal coupling features. Output layer: Set up 3 neurons, corresponding to "low risk", "medium risk" and "high risk" respectively, and determine the level by the pulse firing frequency.

5. The flash flood disaster assessment method based on unmanned aerial vehicles and neuromorphic impulse networks according to claim 1, characterized in that, Step S3 employs a pulse time error minimization optimization model, which minimizes the time difference between the model's output pulse sequence and the standard pulse sequence corresponding to the labeled risk level.

6. The flash flood disaster assessment method based on unmanned aerial vehicles and neuromorphic impulse networks according to claim 1, characterized in that, The aforementioned flash flood disaster assessment method also includes steps for outputting risk level maps, key risk indicators, and early warning recommendations.

7. A flash flood disaster assessment system based on unmanned aerial vehicles and neuromorphic spur networks, characterized in that, include: The multi-source data acquisition module employs a drone array equipped with optical remote sensing sensors, synthetic aperture radar (SAR) remote sensing sensors, infrared sensors, and rainfall sensors to collect multi-source remote sensing data, infrared soil moisture data, and rainfall data for the study area. Data timestamp synchronization is achieved through the drone communication protocol, labeling each frame with its geographic coordinates, acquisition time, and sensor type to form the raw dataset. , i is the UAV number, j is the sensor type, and t is the data acquisition time; The multi-source data preprocessing module corrects the data to eliminate noise and systematic errors, and then performs pixel-level fusion and feature-level fusion to form a standardized dataset. The multimodal feature vector is defined by time t and spatial location p. The brain-like spiking network (SNN) model construction module constructs a brain-like spiking network (SNN) based on leakage integral-ignition neurons. It uses historical UAV remote sensing data and field disaster investigation data, with the spatiotemporal characteristics of the preprocessed historical data as input and the flash flood disaster risk level as output, to train and optimize the SNN model. The flash flood disaster prediction module will process pre-processed real-time multi-UAV data. Input the trained SNN model, and the model outputs the risk level of each spatial grid at time t. For the risk levels of adjacent grids, spatial autocorrelation analysis was used to eliminate isolated outliers, resulting in a spatial distribution map of flash flood risk in the study area.

8. A flash flood disaster assessment system based on unmanned aerial vehicles and neuromorphic impulse networks according to claim 7, characterized in that, It also includes an optimization deployment module for SNN models, which deploys the trained SNN models at the edge nodes of the drone.

9. A flash flood disaster assessment system based on unmanned aerial vehicles and neuromorphic impulse networks according to claim 7, characterized in that, It also includes a dynamic update module that regularly collects new flash flood event data and UAV remote sensing data, and uses an incremental learning strategy to update the synaptic weights of the SNN.

Images