Unmanned aerial vehicle-mounted hydrological state measurement, parameter assimilation and edge prediction device, method and system for mountain torrent small watershed

Abstract

The application discloses a kind of unmanned aerial vehicle-mounted hydrological state measurement, parameter assimilation and edge forecasting equipment, method and system for mountain torrent small watershed, and relates to mountain torrent disaster early warning technical field. Edge computing unit reads known boundary data and public or measured digital elevation model of target mountain torrent small watershed, pre-generates HRU hydrological response unit grid or regular grid space index, and uses the space index for route search, sensor footprint matching and model calculation. Unified to state quantity data cube, the mirror surface reflection point of GNSS-R, local incidence angle and surface roughness are corrected using laser radar DEM and unmanned aerial vehicle attitude, vegetation transmission loss is compensated using multispectral vegetation parameters, and surface soil moisture state quantity with confidence and observation error covariance is generated. Extract the characteristics of underlying surface, river section and water-blocking obstacle, and output flow, water level, inundation, water depth, risk level and uncertainty.

Description

Technical Field

[0001] This invention relates to the field of flash flood disaster early warning technology, and in particular to a UAV-equipped device, method and system for measuring hydrological state quantities, parameter assimilation and edge forecasting for flash flood small watersheds. Background Technology

[0002] Mountain torrent basins are characterized by short runoff times, strong spatial heterogeneity, sparse monitoring stations, sudden localized rainstorms, and rapid changes in underlying surface conditions. Traditional hydrological forecasts often rely on single-station rainfall, water levels, or empirical parameters from adjacent basins to estimate the current basin status, making it difficult to reflect in a timely manner the physical conditions that have a crucial impact on runoff generation and flood evolution, such as slope soil moisture, surface roughness, vegetation status, river cross-section, temporary water-blocking barriers, and local micro-topography.

[0003] In hilly areas with flash flood watersheds, the following technical problems still exist: First, although the target watershed area can be known in advance, existing UAV missions often lack HRU spatial indexes for hydrological models, leading to fragmentation of flight path planning, sensor footprint matching, and model computation units; Second, GNSS-R inversion in hilly areas is easily affected by terrain, roughness, vegetation, and attitude disturbances, making it difficult to obtain reliable soil moisture state quantities that can be used for hydrological models; Third, multi-source UAV observations are difficult to unify into the same hydrological response unit and convert into model parameters such as CN, initial loss rate, Manning roughness rate, and water-blocking cross-section ratio; Fourth, traditional two-dimensional hydrodynamic models are difficult to run quickly under limited flight time and edge computing power conditions; Fifth, the results of previous detections of the same watershed are not structured and accumulated as priors for subsequent forecasts, causing each mission to start approximately from scratch.

[0004] Therefore, there is a need for a UAV equipped with hydrological state quantity measurement, parameter assimilation and edge forecasting equipment, methods and systems for flash flood small watersheds, in order to reduce the uncertainty caused by asynchronous, non-basin and low spatial representative observations in flash flood short-term forecasting. Summary of the Invention

[0005] To address the aforementioned issues, embodiments of this invention propose an unmanned aerial vehicle (UAV) equipped with hydrological state quantity measurement, parameter assimilation, and edge forecasting equipment, methods, and systems for flash flood small watersheds. This system integrates GNSS-R, lidar, and multispectral three-source sensing on a single UAV platform, and completes HRU spatial index construction, collaborative correction, parameter assimilation, edge forecasting, and watershed event memory learning on the airborne or near-aircraft end. This reduces the uncertainty in short-term flash flood forecasting caused by asynchronous, non-watershed, and low spatial representative observations.

[0006] This invention relates to an unmanned aerial vehicle (UAV) for flash flood small watersheds, equipped with hydrological state quantity measurement, parameter assimilation, and edge forecasting equipment, comprising: a flight platform; a modular payload compartment installed at one end of the lower part of the flight platform; a multi-source sensing unit installed inside the modular payload compartment, the multi-source sensing unit including a GNSS-R direct / reflected signal receiving module, a lidar module, and a multispectral imaging module; a spatiotemporal synchronization unit installed inside the modular payload compartment and connected to the multi-source sensing unit; and an edge computing unit located inside the modular payload compartment and connected to the multi-source sensing unit. The computational unit utilizes a lidar module and a multispectral imaging module to generate surface soil moisture state quantities with confidence levels and observation error covariance. The edge computing unit, based on computation time budget, remaining power, computational load, grid size, and state quantity confidence levels, selects an event-driven confluence model, a local hydrodynamic model, or a physically constrained inundation proxy model, outputting cross-sectional flow processes, characteristic cross-sectional water levels, inundation range, water depth, risk level, and corresponding uncertainties. A communication unit, located inside the modular payload compartment, is connected to the multi-source sensing unit. A power thermal management unit, also located inside the modular payload compartment, is connected to the multi-source sensing unit.

[0007] The flight platform is an industrial-grade multi-rotor UAV with RTK or PPK positioning capabilities and multiple payload installation capabilities.

[0008] The modular payload compartment includes a vibration damping mounting plate, an electromagnetic shielding layer, a quick-release interface, an independent power supply interface, a heat dissipation channel, a solid-state storage, and a wiring harness fixing structure. The modular payload compartment is configured to accommodate the GNSS-R radio frequency front end, baseband processing board, edge computing board, power management board, and heat dissipation structure, and is connected to the rigid structure of the UAV body without being stacked on the LiDAR gimbal vibration damper.

[0009] The GNSS-R direct / reflected signal receiving module includes a right-hand circularly polarized GNSS antenna mounted on top of the UAV, a left-hand circularly polarized or dual-circularly polarized GNSS antenna mounted below the UAV, a low-noise amplifier, an RF filter, a dual-channel coherent software radio receiver, and a baseband processing board. The right-hand circularly polarized GNSS antenna is used to receive direct GNSS signals, and the left-hand circularly polarized or dual-circularly polarized GNSS antenna is used to receive GNSS signals reflected from the ground surface.

[0010] The lidar module is an airborne lidar scanning module integrating an inertial measurement unit and positioning calculation capabilities, installed on the main lower gimbal of the UAV or the main lower mounting interface; the multispectral imaging module is a lightweight five-band or five-band plus panchromatic multispectral camera, and is installed on the UAV through a second lower interface, SkyPort interface, QuickMount interface or independent quick-release bracket.

[0011] The present invention relates to an unmanned aerial vehicle (UAV) system for measuring hydrological state quantities, assimilation of parameters, and edge forecasting for flash flood watersheds, comprising the equipment, ground station, and early warning platform described in this invention. The ground station is used to receive state quantities, parameters, risk products, and quality control information output by the equipment, and to perform post-processing verification, baseline HRU spatial index maintenance, historical sample library maintenance, model retraining, or watershed event memory backup. The early warning platform is used to generate short-term flash flood early warning information based on the risk level and uncertainty.

[0012] The method for measuring hydrological state parameters, assimilating parameters, and predicting the edge of flash floods in small watersheds using the device described in this invention includes the following steps: pre-generating an HRU spatial index based on the known target watershed boundary and a publicly available or measured digital elevation model; planning flight routes covering key runoff-generating areas, confluence channels, river cross-sections, and potential hazard points; simultaneously acquiring GNSS-R direct / reflected signals, lidar point clouds, multispectral imagery, and UAV attitude positioning data; constructing a state parameter data cube; performing GNSS-R terrain, roughness, and vegetation compensation in mountainous areas to obtain surface soil moisture state parameters with confidence level and observation error covariance; extracting underlying surface structural features, river cross-section features, and water-blocking obstacle features; and mapping and assimilating the state parameters and structural features to update them into dynamic runoff curves. Initial loss rate Manning roughness Compared with water-blocking cross section At the edge, execute the event generation and confluence model, local hydrodynamic model, or physically constrained inundation proxy model; output the cross-sectional flow process, water level, inundation range, water depth, risk level, and uncertainty; and write the posterior parameters, error covariance, model residuals, and low-confidence regions of this mission into the watershed event memory as priors for subsequent exploration missions.

[0013] The planning of the route does not require full coverage of a fixed area as a necessary condition. Instead, it determines one or more of the following routes based on the baseline HRU spatial index, risk level, historical residuals, high uncertainty areas, rainfall forecast, remaining flight time, and safety constraints of the target flash flood watershed: baseline route for the entire watershed, route for key blocks, fast route before the event, re-measurement route after the event, or supplementary route.

[0014] The number of dynamic runoff curves Initial loss rate Manning roughness The calculation formula is:

[0015]

[0016]

[0017]

[0018] Where VI is the vegetation index, slope is the gradient, impervious is the impermeability, roughness is the surface roughness, M is the historical correction term or posterior parameter term provided by the watershed event memory, clip indicates truncation by physical boundary, and θ is the estimated soil moisture value. This is the initial value for the dynamic runoff curve number. Let the initial value be the Manning roughness. This is the initial value of the initial loss rate. For production flow parameters The lower limit of the value, For production flow parameters The upper limit of the value, Loss rate coefficient The lower limit value, Loss rate coefficient The upper limit, This represents the lower limit of the Manning surface roughness. This represents the upper limit of the surface roughness of the Manning landmass. This is the dynamic correction amount for the CN value. Loss rate coefficient The dynamic correction amount, This is the dynamic correction amount for surface roughness.

[0019] The number of dynamic runoff curves and initial loss rate Used for event runoff calculation, when rainfall P is greater than initial loss. At that time, the runoff depth Q satisfies Q=(PI) a ) ² / (PI a +S), when P is not greater than I a When Q=0;

[0020] In the formula, I a =λ t S, S=25400 / CN t -254, where P is rainfall, Q is runoff depth, and S is the initial water storage capacity of the watershed. This is the initial loss.

[0021] The water-blocking cross-sectional ratio Defined as the projected area of ​​an obstacle in the effective flood discharge section. With effective flood discharge cross-sectional area The ratio, the formula is:

[0022]

[0023] In the formula, The water-blocking cross-section ratio, For the projected area, The effective flood discharge cross-sectional area.

[0024] like Figure 1 As shown, the workflow of the edge computing unit is as follows:

[0025] A. Read the known watershed boundary data, publicly available or measured digital elevation model data, and optional land use and hydrological soil group data of the target flash flood watershed, and pre-generate the hydrological response unit HRU grid or regular raster spatial index.

[0026] The boundary data of the target flash flood watershed is the watershed boundary after the administrative boundary is trimmed, or the boundary of the flash flood prevention target is artificially designated; the digital elevation model data is a 12.5 m precision DEM, a 30 m precision DEM, an airborne lidar DEM or a combination thereof that are publicly available and freely obtained.

[0027] The HRU grid is generated through digital elevation model hydrological correction, flow direction calculation, runoff area calculation, channel network extraction, slope classification, land use merging, hydro-soil group merging, and neighborhood merging. Each cell in the HRU grid has a unique number, geometric boundary, area, average slope, dominant land cover, dominant hydro-soil group, runoff path, downstream connectivity, and risk priority.

[0028] B. Use the HRU grid or regular raster spatial index as a spatial reference for UAV route retrieval, multi-source observation footprint matching, and model calculation.

[0029] The route retrieval process includes: overlaying the GNSS-R reflection footprint, lidar scanning bandwidth, and multispectral imaging swath onto the HRU grid, calculating the coverage ratio, observation time, and observation quality of each HRU covered by each sensor, and then generating key HRU routes, supplementary HRU routes, or post-event re-measurement routes.

[0030] C. Unify GNSS-R direct / reflection observations, lidar point clouds, multispectral imagery, and UAV attitude positioning data to the same time and spatial coordinate reference, and construct a state variable data cube for the target flash flood watershed.

[0031] The state data cube includes the elevation, slope, aspect, flow direction, catchment area, topographic humidity index, land cover category, vegetation index, canopy height, topsoil moisture, topsoil moisture confidence level, observation error covariance, surface roughness, river distance, building or bridge obstacle identification, cross-sectional geometric parameters, flood control and water obstruction characteristics, and observation timestamp for each HRU or raster.

[0032] D. Using the digital elevation model generated by the lidar module and the attitude positioning data of the UAV, the specular reflection point, local incident angle and surface roughness attenuation of the GNSS-R reflection signal are corrected. The vegetation parameters obtained by the multispectral imaging module are used to compensate for the vegetation transmission loss of the GNSS-R reflection signal, and the surface soil moisture state quantity with confidence and observation error covariance is generated.

[0033] Correcting the specular reflection point and local incident angle includes: solving for the intersection of the GNSS reflection path and the local terrain surface based on the GNSS satellite ephemeris, UAV positioning attitude, phase center of the up-and-down antenna, and digital elevation model to obtain the corrected specular reflection point; extracting the slope and aspect of the neighborhood of the corrected specular reflection point, correcting the nominal incident angle under the assumption of a horizontal surface to the local true incident angle, and correcting the Fresnel reflection coefficient, normalized reflectivity, or delayed Doppler energy distribution accordingly.

[0034] Correcting surface roughness attenuation includes: calculating at least one roughness index from the root mean square elevation of lidar point cloud, local slope fitting residual, micro-topographic undulation, point cloud intensity, or point cloud height quantile within the neighborhood of the corrected specular reflection point, and applying attenuation compensation to the GNSS-R coherent reflection component using the roughness index, or converting the roughness index into a weighting factor for the confidence level of surface soil moisture observation.

[0035] Compensation for vegetation transmission loss includes: calculating at least one vegetation parameter from multispectral images, such as NDVI, EVI, LAI, canopy coverage, or red edge index, and estimating vegetation optical thickness or vegetation transmittance; correcting the amplitude of GNSS-R reflected signals based on vegetation transmittance; and reducing the confidence level of surface soil moisture state quantity or increasing its observation error covariance when the vegetation index, canopy height, or canopy coverage exceeds a preset threshold.

[0036] Topsoil moisture state parameters include soil moisture estimate θ and confidence level. and observation error covariance Confidence level The covariance of observation error is determined by at least three of the following: GNSS-R signal-to-noise ratio, number of visible satellites, local slope, surface roughness, vegetation parameters, reflection footprint mixing degree, UAV attitude stability, and soil moisture inversion residuals; With confidence level Negative correlation.

[0037] E. Based on digital elevation models, multispectral images, and surface soil moisture state data, extract underlying surface structural features, river cross-sectional features, and flood control and water obstruction features.

[0038] F. Map and assimilate the state variables and structural features to update them into hydrological and hydrodynamic parameters, which include at least the number of dynamic runoff curves. Initial loss rate Manning roughness Compared with water-blocking cross section .

[0039] The assimilation update employs ensemble Kalman filtering, extended Kalman filtering, unscented Kalman filtering, particle filtering, Bayesian recursion, or a combination thereof; the assimilation state vector includes at least the soil moisture estimate θ. , , , Cross-sectional flow or water level The four types are: GNSS-R soil moisture estimation, multispectral water or vegetation status, surface rainfall, radar precipitation, water level, flow velocity, or cross-sectional flow.

[0040] The state vector reflects the hydrological and hydrodynamic parameters and physical state quantities of the physical processes within the target flash flood watershed. Specifically, this includes traditional hydrological state quantities describing the current physical characteristics of the watershed (such as surface soil moisture θ, cross-sectional discharge). and water level ), and dynamic model parameters derived from state variables and structural features (such as , , and ).

[0041] The observation vector refers to the actual measurement data acquired through UAV multi-source sensing devices or external monitoring methods. During the assimilation process, the observation vector serves as a true external reference value, used to constrain and correct the assimilation state vector. This includes GNSS-R soil moisture estimation, multispectral water or vegetation status, surface rainfall, radar precipitation, water level, flow velocity, or cross-sectional flow rate. It belongs to externally directly measured or sensed state quantities. These are the actual observation outputs captured by sensors at specific moments, serving as the input basis for the assimilation system to update the state and correct errors.

[0042] The mapping of hydrological and hydrodynamic parameters includes: obtaining initial values ​​of dynamic runoff curve numbers from tables based on land use type and hydrological soil group. and initial value of Manning roughness Based on the estimated soil moisture θ, vegetation index, slope, impermeability, surface roughness, canopy height, and water-blocking cross-section ratio. Calculate the parameter correction amount; combine the parameter correction amount with the initial value from the lookup table to obtain... , , and And apply physical value boundaries to the parameters.

[0043] Water-blocking cross-sectional ratio The projected area of ​​water obstruction within the flood discharge section includes bridge piers, culverts, accumulated materials, floating objects, shrubs, fallen trees, buildings, or temporary water-blocking structures. and corresponding effective flood discharge cross-sectional area The ratio is determined, that is Water-blocking projected area It is determined by at least one of the following: lidar point cloud, multispectral image, visible light image, or their fusion recognition results.

[0044] Number of dynamic runoff curves Initial loss rate and Manning roughness Satisfying one or more of the following mapping relationships:

[0045]

[0046]

[0047]

[0048] Where VI is the vegetation index, slope is the gradient, impervious is the impermeability, roughness is the surface roughness, M is the historical correction term or posterior parameter term provided by the watershed event memory, clip indicates truncation by physical boundary, and θ is the estimated soil moisture value. This is the initial value for the dynamic runoff curve number. Let the initial value be the Manning roughness. This is the initial value of the initial loss rate. For production flow parameters The lower limit of the value, For production flow parameters The upper limit of the value, Loss rate coefficient The lower limit value, Loss rate coefficient The upper limit, This represents the lower limit of the Manning surface roughness. This represents the upper limit of the surface roughness of the Manning landmass. This is the dynamic correction amount for the CN value. Loss rate coefficient The dynamic correction amount, This is the dynamic correction amount for surface roughness.

[0049] Number of dynamic runoff curves and initial loss rate Used for event runoff calculation, when rainfall P is greater than initial loss. At that time, the runoff depth Q satisfies Q=(PI) a ) ² / (PI a +S), when P is not greater than I a When Q=0; I a =λ t S, S=25400 / CN t -254, P, Q, and S are in millimeters.

[0050] G. Within the edge computing unit, based on the computing time budget, remaining power, computing load, grid size, and state quantity confidence, select the event generation and confluence model, local hydrodynamic model, or physically constrained inundation proxy model, and output the cross-sectional flow process, characteristic cross-sectional water level, inundation range, water depth, risk level, and corresponding uncertainties.

[0051] The local hydrodynamic model is based on a digital elevation model, net rainfall input, boundary flow, water level boundary, and Manning roughness. The diffuse wave model, local inertial shallow water model, or simplified Saint-Venant model are used to calculate water depth, flow direction, flow velocity, and overflow range in the neighborhood of the target hazard point.

[0052] The physically constrained inundation surrogate model is a pruned, quantized, or distilled random forest, gradient boosting tree, lightweight convolutional neural network, graph neural network, physical information neural network, or a combination thereof; the inputs to the physically constrained inundation surrogate model include digital elevation model, slope, catchment area, land cover, etc. , , The output of the physical constraint inundation proxy model includes inundation probability, water depth grid, inundation range, and risk level. The model can be configured to consider rainfall characteristics, cross-sectional flow processes, or water level boundaries.

[0053] When training or inferring the physical constraint flooding proxy model, at least two of the following constraints are applied: non-negativity of water depth, consistency of flooding extent with watershed or levee boundary, monotonic constraint that water depth does not decrease as flow or water level increases, low-lying connectivity constraint based on digital elevation model, local mass conservation residual constraint, and error penalty constraint consistent with hydrodynamic model sample library.

[0054] Model selection is generated by a scoring function or threshold rule, which utilizes at least a computational time budget. Remaining battery power Calculate the overall confidence level of the load L and state variables. Risk level of hidden danger points and historical model residuals The three items in the text.

[0055] H. Write the state variable data cube, correction coefficients, parameter posterior values, error covariance, model residuals, low confidence regions, and observation quality identifiers of this mission into the watershed event memory. In subsequent exploration missions, use the historical parameter posteriors and error covariances in the watershed event memory as prior inputs for parameter assimilation and model selection.

[0056] The watershed event memory includes baseline digital elevation models, baseline land cover, baseline river cross sections, baseline HRU spatial indexes, state data cubes from each mission, GNSS-R terrain roughness and vegetation correction coefficients, soil moisture inversion residuals, posterior values ​​of hydrological and hydrodynamic parameters, error covariance, physical constraint inundation surrogate model residuals, high-risk areas, high-uncertainty areas, quality threshold identifiers, and mission timestamps.

[0057] The watershed event memory is used in subsequent exploration missions for: using historical parameter posteriors as parameter assimilation priors; generating key HRU routes or supplementary survey areas based on historical low-confidence areas and model residual hotspots; updating GNSS-R terrain roughness and vegetation correction coefficients; and updating... , , and The parameters are posteriorized; and when the observation quality is below the quality threshold, the uncertainty is only increased without being written into the long-term parameter library.

[0058] A computer-readable storage medium having a computer program stored thereon, which, when executed by an edge computing unit of the device of the present invention, implements the method of the present invention.

[0059] The beneficial effects of this invention are as follows: By pre-generating HRU spatial indexes using known watershed boundaries and publicly available or measured DEMs, consistency is maintained between flight path planning, sensor footprint attribution, state variable database construction, and model calculation units, avoiding mismatches between simple image gratings and hydrological units. By unifying temporal and spatial references, GNSS-R, lidar, multispectral, and UAV attitude positioning data are mapped to the same HRU or grating unit, reducing state estimation errors caused by asynchronous multi-source data. Local mirror point, local incidence angle, and roughness compensation driven by lidar DEM improve the physical consistency of GNSS-R soil moisture inversion in hilly areas, and avoids excessive influence of low-reliability observations on parameter assimilation through confidence level and observation error covariance. Multispectral vegetation parameters simultaneously serve GNSS-R transmittance compensation, land cover identification, and hydrological and hydrodynamic parameter correction, enabling verifiable collaborative correction relationships between sensors, rather than simple parallel measurements. Through state variables to… , , and The mapping and assimilation process transforms UAV remote sensing products into a dynamic parameter set that can directly drive event generation and confluence, local hydrodynamics, and inundation inference models. Through edge-end hierarchical models and physically constrained inundation proxy inference, it outputs flow, water level, inundation, and risk products suitable for short-term flash flood warnings under limited payload, computing power, and flight time. By utilizing a watershed event memory database, the posterior parameters, error covariance, and residual hotspots from previous surveys are used for subsequent tasks, enabling rolling learning capabilities for the same watershed's surveys and forecasts, thus improving the stability and speed of subsequent forecasts. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the overall system architecture and HRU spatial index closed loop of the present invention.

[0061] Figure 2 The flowchart shows the HRU pre-partitioning and route retrieval process driven by the known watershed extent and public DEM.

[0062] Figure 3 This is a schematic diagram of a single UAV platform with three payloads and a GNSS-R dual-antenna structure.

[0063] Figure 4 A flowchart for generating confidence scores for GNSS-R collaborative correction and soil moisture in mountainous areas.

[0064] Figure 5 This is a closed-loop diagram for parameter mapping, assimilation, edge prediction, and watershed event memory learning.

[0065] Figure 6 This is a schematic diagram of the structure of the UAV equipped with hydrological state quantity measurement, parameter assimilation and edge forecasting equipment for small mountain torrent watersheds according to the present invention.

[0066] Figure 7 This is a schematic diagram of the structure of the lidar module of the present invention.

[0067] Figure 8 This is a schematic diagram of the GNSS-R direct / reflected signal receiving module of the present invention.

[0068] Figure 9 This is a schematic diagram of the structure of the multispectral imaging module of the present invention.

[0069] Figure label:

[0070] LiDAR module 1; modular payload compartment 2; upward-looking RHCP antenna 201; downward-looking LHCP antenna 202; multispectral imaging module 3; main lower-mounted gimbal 4. Detailed Implementation

[0071] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0072] like Figures 1-9 As shown, the UAV for flash flood watersheds of the present invention is equipped with hydrological state quantity measurement, parameter assimilation and edge forecasting equipment, including: flight platform, modular payload cabin, multi-source sensing unit, spatiotemporal synchronization unit, edge computing unit, communication unit and power thermal management unit.

[0073] The flight platform is an industrial-grade multi-rotor drone with RTK or PPK positioning capabilities and multi-payload mounting capabilities. The flight platform is either a DJI Matrice 300 RTK or an industrial-grade multi-rotor drone with equivalent RTK / PPK positioning capabilities, payload capacity, gimbal interface, and extended power supply capabilities. The drone is an M 300 RTK with multi-payload mounting capabilities, employing a multi-gimbal system that supports up to three payloads simultaneously and is compatible with various industry payloads.

[0074] Modular payload compartment 2 is installed at one end of the lower part of the flight platform; Modular payload compartment 2 includes a vibration damping mounting plate, an electromagnetic shielding layer, a quick-release interface, an independent power supply interface, a heat dissipation channel, a solid-state memory and a wiring harness fixing structure. The modular payload compartment 2 is configured to accommodate the GNSS-R radio frequency front end, baseband processing board, edge computing board, power management board and heat dissipation structure, and is connected to the rigid structure of the UAV body without being stacked on the laser radar gimbal vibration damper.

[0075] The multi-source sensing unit is installed inside the modular payload bay 2. The multi-source sensing unit includes a GNSS-R direct / reflected signal receiving module, a lidar module 1, and a multispectral imaging module 3. The radio frequency and baseband processing components of the GNSS-R direct / reflected signal receiving module are installed inside the modular payload bay, while the lidar module 1 and the multispectral imaging module 3 are independently installed outside the flight platform via mounting interfaces.

[0076] The GNSS-R direct / reflected signal receiving module includes a right-hand circularly polarized GNSS antenna (upward-looking RHCP antenna 201) mounted on top of the UAV, a left-hand circularly polarized or dual-circularly polarized GNSS antenna (downward-looking LHCP antenna 202 or dual-circularly polarized antenna) mounted below the UAV, a low-noise amplifier, an RF filter, a dual-channel coherent software radio receiver, and a baseband processing board. The right-hand circularly polarized GNSS antenna is used to receive direct GNSS signals, while the left-hand circularly polarized or dual-circularly polarized GNSS antenna is used to receive GNSS signals reflected from the ground surface.

[0077] The GNSS-R upward-looking antenna 201 is mounted on the top of the fuselage or on a small mast, avoiding obstruction from the propeller sweep surface, the UAV's built-in RTK antenna, and the multispectral downlink optical sensor; the GNSS-R downward-looking antenna 202 is mounted on the belly of the fuselage or below the payload bay, avoiding entry into the scanning field of view of the lidar and the imaging field of view of the multispectral camera. The dual-channel software-defined radio receiver, low-noise amplifier, filter, edge computing board, solid-state storage, and DC / DC conversion module are housed in a lightweight shielded payload bay.

[0078] To reduce electromagnetic interference, shielded coaxial cables are used for the GNSS-R radio frequency line and are kept as short as possible. The radio frequency link is separated from the UAV's high-voltage power cable bundle. The payload compartment maintains a predetermined isolation distance from the motors, ESCs, and high-current cables, and a grounded shielding layer is installed. Before flight, ground static spectrum testing, rotor idle speed and full throttle electromagnetic interference testing, low-altitude constant altitude flight testing, and payload trim checks are performed.

[0079] The GNSS-R direct / reflected signal receiving module is connected to the spatiotemporal synchronization unit via an external second pulse, a 10 MHz reference clock, or a hardware trigger signal, and outputs at least one GNSS-R reflection characteristic among the following: relative power of the direct / reflected signal, coherent or incoherent integrated power, delayed Doppler graph, signal-to-noise ratio sequence, carrier-to-noise ratio sequence, or normalized reflectivity.

[0080] The lidar module 1 is an airborne lidar scanning module that integrates an inertial measurement unit and positioning calculation capabilities. It is installed on the main lower gimbal 4 of the UAV or the main lower mounting interface. The multispectral imaging module 3 is a lightweight five-band or five-band plus panchromatic multispectral camera, which is installed on the UAV through the second lower interface, SkyPort interface, QuickMount interface or independent quick-release bracket.

[0081] The lidar module 1 generates DEM, DSM, point cloud roughness, and river cross-section; the multispectral imaging module 3 acquires spectral states such as vegetation index, land cover, impermeable surface, and water body.

[0082] The multispectral imaging module 3 acquires at least five narrow band images from the blue, green, red, red-edge, and near-infrared bands, and can optionally acquire panchromatic or visible light images; the downlink optical sensor that accompanies the multispectral imaging module 3 is installed in an unobstructed position above the UAV body for irradiance correction.

[0083] The spatiotemporal synchronization unit, edge computing unit, communication unit, and power thermal management unit are installed inside the modular payload bay and are all connected to the multi-source sensing unit. The spatiotemporal synchronization unit includes a PPS second pulse, hardware trigger line, RTK / PPK track, inertial measurement data, unified timestamp management module, and extrinsic parameter management module. The extrinsic parameter management module stores the lidar coordinate system, multispectral camera center, GNSS-R up-looking antenna phase center, GNSS-R down-looking antenna phase center, levers, rotation matrix, and calibration errors between the downlink optical sensor and the UAV system.

[0084] The communication unit is used to transmit surface soil moisture grids and confidence levels, dynamic data. The parameter grid, flow process line of key sections, highest water level of characteristic sections, inundation range, water depth grid, risk level, uncertainty level, low confidence area and supplementary measurement suggestions are sent to ground stations, mobile terminals or flash flood warning platforms.

[0085] The edge computing unit or ground station preprocessing unit first reads the known boundary data, publicly available or measured DEM, land use, and hydrological and soil group data of the target flash flood watershed. Through DEM hydrological correction, flow direction calculation, runoff area calculation, channel network extraction, slope classification, land use merging, hydrological and soil group merging, and neighborhood merging, it pre-generates an HRU hydrological response unit grid or a regular raster spatial index. The spatial index is used for flight path retrieval, GNSS-R reflection footprint matching, lidar point cloud assignment, multispectral image assignment, and model calculation unit definition during flight missions.

[0086] During flight mission execution, the edge computing unit constructs a target watershed state quantity data cube based on a unified time reference, a unified coordinate reference, and sensor extrinsic parameters. Then, it uses lidar DEM, UAV attitude, and GNSS satellite ephemeris to solve for GNSS-R corrected specular reflection points and local true incident angles. It uses lidar point cloud roughness index to compensate for coherent reflection attenuation caused by surface roughness, and uses multispectral vegetation index or LAI to estimate vegetation transmittance and compensate for GNSS-R signal penetration loss, thereby outputting surface soil moisture state quantities with confidence level and observation error covariance.

[0087] Subsequently, the edge computing unit extracts underlying surface structure, river cross-section, and water-blocking barrier features from the digital elevation model, multispectral imagery, and surface soil moisture state variables, and maps these state variables and structural features into dynamic runoff curves. Initial loss rate Manning roughness Compared with water-blocking cross section The posterior parameters and error covariance are updated using assimilation methods such as ensemble Kalman filtering, unscented Kalman filtering, particle filtering, or Bayesian recursion.

[0088] Furthermore, the edge computing unit selects an event-driven confluence model, a local hydrodynamic model, or a physically constrained inundation surrogate model based on the computation time budget, remaining power, computational load, state quantity confidence, risk level of potential hazards, and historical model residuals. It then outputs the flow process at key cross-sections, water levels at characteristic cross-sections, inundation range, water depth, risk level, and uncertainty. The system also establishes a watershed event memory to store the HRU spatial index, state quantity data cubes, correction coefficients, parameter posteriors, error covariance, model residuals, and low-confidence regions from each task. These are used as priors for parameter assimilation, model selection, and key route generation in subsequent tasks.

[0089] The present invention relates to an unmanned aerial vehicle (UAV) system for measuring hydrological state quantities, assimilation of parameters, and edge forecasting for flash flood watersheds, comprising the equipment, ground station, and early warning platform described in the present invention. The ground station is used to receive state quantities, parameters, risk products, and quality control information output by the equipment described in the present invention, and to perform post-processing verification, baseline HRU spatial index maintenance, historical sample library maintenance, model retraining, or watershed event memory backup. The early warning platform is used to generate short-term flash flood early warning information based on the risk level and uncertainty.

[0090] The method for measuring hydrological state parameters, assimilating parameters, and predicting the edge of flash floods in small watersheds using the device described in this invention includes the following steps: pre-generating an HRU spatial index based on the known target watershed boundary and a publicly available or measured digital elevation model; planning flight routes covering key runoff-generating areas, confluence channels, river cross-sections, and potential hazard points; simultaneously acquiring GNSS-R direct / reflected signals, lidar point clouds, multispectral imagery, and UAV attitude positioning data; constructing a state parameter data cube; performing GNSS-R terrain, roughness, and vegetation compensation in mountainous areas to obtain surface soil moisture state parameters with confidence level and observation error covariance; extracting underlying surface structural features, river cross-section features, and water-blocking obstacle features; and mapping and assimilating the state parameters and structural features to update them into dynamic runoff curves. Initial loss rate Manning roughness Compared with water-blocking cross section At the edge, execute the event generation and confluence model, local hydrodynamic model, or physically constrained inundation proxy model; output the cross-sectional flow process, water level, inundation range, water depth, risk level, and uncertainty; and write the posterior parameters, error covariance, model residuals, and low-confidence regions of this mission into the watershed event memory as priors for subsequent exploration missions.

[0091] Example 1: Target Watershed Preprocessing and HRU Spatial Indexing

[0092] like Figure 2 As shown, for a target flash flood watershed with a known range, the system reads the watershed boundary SHP data, publicly available 12.5 m precision DEM data or higher precision DEM data before the formal flight, as well as optional land use, soil type, hydro-soil group, historical disaster points, bridge and culvert points, and settlement data. The edge computing unit or ground station preprocessing module performs coordinate unification, cropping, depression filling, outlier correction, and hydrological connectivity correction on the DEM.

[0093] On the hydrologically corrected DEM, the system calculates flow direction, catchment area, slope, aspect, topographic humidity index, and channel network. Combining land use type, hydro-soil group, slope classification, catchment path consistency, and artificially designated hazard points, the target watershed is divided into regular grids or HRUs. The principle for HRU division is that topographic control, land cover, hydro-soil group, and catchment path within the same unit have relative consistency, and the area is not less than the lower limit of the effective sensor observation footprint or the smallest unit scale calculated by the subsequent model.

[0094] Each HRU stores a unique ID, geometric boundaries, area, average elevation, average slope, dominant land cover, dominant hydrosomatic group, downstream HRU ID, river distance, whether it includes channels or hazard points, historical residual level, and priority score. When generating a flight mission, the system generates an executable or supplementary flight path based on HRU priority, GNSS-R reflection footprint, multispectral swath width, lidar scanning bandwidth, terrain safety altitude, and remaining flight time.

[0095] Example 2: Spatiotemporal Synchronization and Extrinsic Parameter Calibration

[0096] The spatiotemporal synchronization unit includes a PPS second pulse, a 10 MHz reference clock, a hardware trigger line, RTK / PPK tracks, inertial measurement data, and a timestamp management module. GNSS-R raw IF / IQ samples or reflection features, LiDAR point clouds, multispectral image frames, and UAV attitude and positioning data are all appended with a unified timestamp. The edge computing unit transforms the extrinsic parameters of each sensor to the UAV's body coordinate system, and then to the geographic coordinate system, forming a dataset within the same coordinate framework.

[0097] The extrinsic parameters that need to be calibrated include the rotation and translation of the lidar coordinate system to the system, the rotation and translation of the multispectral camera center to the system, the struts of the GNSS-R up-view and down-view antenna phase centers to the system, the position of the downlink optical sensor to the system, and the offset between the payload bay's center of mass and the UAV's overall center of gravity. Extrinsic parameter errors are then addressed through solutions for specular reflection points, local incidence angle correction, point cloud image registration, and soil moisture confidence level calculation.

[0098] Example 3: Flight Mission and State Quantity Data Cube

[0099] Before flight, a flight path is planned to cover key runoff-generating slopes, confluence channels, river cross-sections, and potential hazard points based on the HRU spatial index, gullies, settlements, bridges and culverts, historical disaster sites, ground rain gauges, weather radar, or rainfall forecasts of the target flash flood watershed. The flight path altitude and heading are determined comprehensively based on the GNSS-R footprint, lidar point density, multispectral ground sampling distance, terrain safety distance, and remaining flight time. This invention does not require single-pass full coverage of a fixed area as a necessary condition.

[0100] Edge computing units divide the target small watershed into regular grids or HRUs. Each unit stores elevation, slope, aspect, flow direction, catchment area, topographic moisture index, land cover type, vegetation index, canopy height, topsoil moisture, soil moisture confidence level, observation error covariance, surface roughness, channel distance, obstacle type, cross-sectional geometric parameters, and observation timestamps, forming a state variable data cube. .

[0101]

[0102] In the formula, For state variable data cube, Indicates slope, Indicates vegetation index, This represents an estimated soil moisture content. Indicates the confidence level. denoted by _z_, representing the observation error covariance; _roughness_, representing surface roughness; and _z_, representing elevation. Indicates slope direction. Indicates the direction of water flow, and A represents the catchment area. Indicates the topographic humidity index. Indicates land cover category, Indicates the distance of the river channel. Indicates the water-blocking cross-sectional area ratio. The observation time is indicated by 'i', which represents the spatial unit index, specifically referring to the i-th hydrological response unit (HRU) or regular grid.

[0103] Example 4: GNSS-R Soil Moisture Inversion and Confidence Level in Mountainous Areas

[0104] like Figure 4 As shown, the edge computing unit first determines the candidate reflection geometry based on GNSS satellite ephemeris, UAV position and attitude, and the phase centers of the top and bottom look-ahead antennas. Then, it solves for the intersection of the reflection path and the local terrain surface on the lidar DEM to obtain the corrected specular reflection point. The slope and aspect of the neighborhood of this point are used to calculate the local true angle of incidence, replacing the nominal angle of incidence under the assumption of a horizontal surface.

[0105] Within the neighborhood of the corrected specular reflection point, the edge computing unit calculates the roughness index using the lidar point cloud elevation residual, local slope fitting residual, root mean square elevation, or point cloud height quantile difference. This index is used to correct for coherent reflectance energy or reduce observation weights. Multispectral imagery is used to calculate NDVI, EVI, LAI, or canopy cover, and to estimate vegetation transmittance. The GNSS-R characteristics after terrain, roughness, and vegetation correction can be expressed as:

[0106]

[0107] in, These are the corrected GNSS-R characteristic quantities. To observe the reflection characteristics, For local incident angle, This is the antenna attitude compensation term. For roughness index, The soil moisture content is the vegetation transmittance. Soil moisture estimation can be achieved using dielectric model lookup tables, semi-empirical regression, random forests, gradient boosting, or machine learning residual models. The edge computing unit simultaneously outputs the estimated soil moisture content θ and its confidence level. and observation error covariance When the slope, vegetation obstruction, or roughness exceeds a threshold, the system does not force a high-confidence soil moisture reading. Instead, it marks the cell as low-confidence and adds this information to the parameter assimilation process. .

[0108] Example 5: Feature extraction of underlying surface, river cross-section, and water-blocking obstacles

[0109] After ground point filtering, lidar point clouds are used to generate a Digital Surface Model (DEM) and a Digital Surface Model (DSM). The DEM is used to extract slope, aspect, channel network, runoff area, cross-sectional geometry, and depression connectivity. The difference between the DSM and the DEM is used to estimate canopy height, building height, or obstacle height. Multispectral imagery, after radiometric and orthorectification correction, is used to extract vegetation indices, bare land, water bodies, impervious surfaces, farmland, woodland, grassland, and construction land categories.

[0110] For potential hazards in river channels or ditches, the system generates cross-sections at preset locations or narrow locations automatically identified by the DEM. Piers, culverts, landslide debris, floating objects, dense shrubs, fallen trees, or temporary structures can be identified through a combination of point cloud height abrupt changes, local geometric occlusion, texture features, multispectral categories, and image target detection. The water-blocking cross-sectional area ratio is defined as the projected area of ​​the obstacle within the effective flood discharge cross-section. With effective flood discharge cross-sectional area The ratio.

[0111]

[0112] In the formula, The water-blocking cross-section ratio, For the projected area, The effective flood discharge cross-sectional area.

[0113] Example 6: Mapping and Assimilation of Hydrological and Hydrodynamic Parameters

[0114] For each HRU or grid, the system obtains the data from a table based on land use and hydro-soil groups. and Soil moisture estimate θ, vegetation index VI, slope, impervious, roughness, canopy height, and water-blocking section ratio. The historical correction term M provided by the watershed event memory is used to generate parameter correction values. Dynamic parameters can be expressed as:

[0115]

[0116]

[0117]

[0118] Where VI is the vegetation index, slope is the gradient, impervious is the impermeability, roughness is the surface roughness, M is the historical correction term or posterior parameter term provided by the watershed event memory, clip indicates truncation by physical boundary, and θ is the estimated soil moisture value. This is the initial value for the dynamic runoff curve number. Let the initial value be the Manning roughness. This is the initial value of the initial loss rate. For production flow parameters The lower limit of the value, For production flow parameters The upper limit of the value, Loss rate coefficient The lower limit value, Loss rate coefficient The upper limit, This represents the lower limit of the Manning surface roughness. This represents the upper limit of the surface roughness of the Manning landmass. This is the dynamic correction amount for the CN value. Loss rate coefficient The dynamic correction amount, This is the dynamic correction amount for surface roughness.

[0119] Number of dynamic runoff curves and initial loss rate For event runoff calculation, the modified SCS-CN method can be used when the rainfall P is greater than the initial loss. At that time, the runoff depth Q satisfies:

[0120] Q=(PI a ) ² / (PI a +S),

[0121] I a =λ t S,

[0122] S=25400 / CN t -254,

[0123] When P is not greater than When Q=0; where P is the rainfall, Q is the runoff depth, and S is the initial water storage capacity of the watershed. This is the initial loss.

[0124] The assimilated state vector can be represented as:

[0125]

[0126] in, This represents the estimated soil moisture value for time period t. For the number of dynamic runoff curves, Initial loss rate, For Manning roughness, The water-blocking cross-section ratio, This refers to the cross-sectional flow rate. This refers to the water level or water depth.

[0127] Observation vectors can include GNSS-R soil moisture, multispectral water body or vegetation status, surface rainfall, radar precipitation, water level, flow velocity, or cross-sectional discharge. Edge computing units use ensemble Kalman filtering, unscented Kalman filtering, particle filtering, or Bayesian recursion to update the state vector and error covariance. For low-confidence soil moisture units, the observation error variance is increased, thereby reducing the impact of that observation on parameter updates.

[0128] Example 7: Hydrological and hydrodynamic calculations at the edge of the boundary

[0129] The edge computing unit incorporates a hierarchical model library. Level 1 models are event-driven runoff models, used for quickly estimating runoff depth, runoff timing, and critical section flow processes for each sub-basin or HRU. Level 2 models are localized diffuse wave models, localized inertial shallow water models, or simplified Saint-Venant models, used to calculate water depth, flow direction, velocity, and overflow extent in the neighborhood of potential hazard points. Level 3 models are physically constrained flooding proxy models, used to replace full two-dimensional hydrodynamic iterations when edge computing power is insufficient, the grid size is too large, or rapid rolling updates are required.

[0130] Model selection based on time budget Remaining battery power Calculate the overall confidence level of the load L and state variables. Risk level of hidden danger points and historical model residuals A joint decision. When Sufficient and When the temperature is high, execute the second-level hydrodynamic model; when If the threshold is insufficient or L exceeds the threshold, the three-level proxy model is executed; when When the risk level is low and key inputs are lacking, only the risk level and uncertainty warnings are output, and a supplementary test route is generated or a request for verification is submitted to the ground station.

[0131] The physically constrained inundation surrogate model can be trained using historical hydrodynamic simulation samples, measured flood inundation samples, and UAV-measured topographic imagery samples. Input channels include DEM, slope, catchment area, land cover, and... , , The system includes rainfall duration and intensity, cross-sectional flow process, and water level boundaries; the output channels include a water depth grid, inundation probability, inundation boundary, and risk level. During training or inference, constraints such as non-negative water depth, watershed or levee boundary consistency, flow monotonicity, topographic connectivity, and quality conservation residuals are applied to improve the physical consistency of the extrapolated scenario.

[0132] Example 8: Watershed Event Memory Learning Closed Loop

[0133] The watershed event memory is used to store the baseline HRU spatial index, state data cube, GNSS-R terrain roughness vegetation correction coefficient, soil moisture inversion residual, hydrological and hydrodynamic parameter posterior values, error covariance, physical constraint inundation surrogate model residual, high-risk area, high uncertainty area, quality threshold identifier, and task timestamp for the same flash flood watershed under different seasons, different rainfall events, and different detection tasks.

[0134] Before a mission, the edge computing unit or ground station reads the watershed event memory to obtain the posterior parameters, error covariance, low-confidence HRUs, and model residual hot zones from the previous mission. Route generation does not rely on multi-aircraft collaborative scheduling, but rather generates key routes based on high-risk and high-uncertainty HRUs under single-UAV conditions; the entire watershed baseline can be updated infrequently, with priority given to rolling updates of key areas during rainfall events.

[0135] During the mission, GNSS-R, lidar, and multispectral data were acquired simultaneously; lidar constrained GNSS-R for terrain and roughness, and multispectral data constrained for vegetation transmission, outputting θ, and Low-confidence observations are automatically downweighted. Parameter updates use the memory parameters as priors and the current observations as likelihoods, employing EnKF, UKF, PF, or Bayesian recursion for updates. , , , And error covariance.

[0136] After the task, the forecast results are verified using ground rainfall, water level, flow velocity, floodmarks, manual verification, or high-precision post-processing models, and residuals are calculated. Only data that passes the quality threshold is written into long-term memory; low-quality data is only used as temporary markers. To avoid becoming increasingly biased in learning, all learning is more evenly distributed. , , and Physical boundary constraints; when the observation confidence is low, only the uncertainty is increased, and long-term parameters are not forcibly updated; seasonal mutations, engineering disturbances, landslide accumulation or large-scale changes in vegetation trigger local reconstruction or reset.

[0137] Example 9: Output, Early Warning, and Quality Control

[0138] The equipment outputs the following results to the ground station or early warning platform: surface soil moisture grid and confidence level, dynamic... The output includes a parameter raster, flow process curves at key sections, maximum water level at characteristic sections, inundation range of potential hazard points, water depth raster, risk level, low-confidence areas, and uncertainty descriptions. Output can be in GeoTIFF, GeoJSON, NetCDF, CSV, or lightweight binary message format.

[0139] Quality control includes checking sensor timestamp integrity, GNSS-R signal-to-noise ratio, number of visible satellites, lidar point cloud density, multispectral exposure and cloud shadow, attitude anomaly, DEM depression consistency, extrinsic parameter consistency, and model residuals. If key quality indicators fail to meet thresholds, the edge computing unit marks the area as low confidence and displays the uncertainty level in the early warning product.

[0140] Optionally, the multi-source sensing unit may also include millimeter-wave water level radar, surface flow velocity radar, visible light camera, or thermal infrared camera to acquire water level, surface flow velocity, flood marks, or water body boundary observations above bridges, flooded road surfaces, or ditch cross-sections, and use them as assimilation observation vectors. The edge computing unit can also work in conjunction with ground stations or cloud platforms, with the edge completing rapid risk output and the cloud performing high-precision two-dimensional model verification, model retraining, and historical sample database updates.

[0141] Although the above embodiments have been shown and described, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Any changes, modifications, substitutions and variations made to the above embodiments by those skilled in the art are within the protection scope of the present invention.

Claims

1. A drone equipped with hydrological state quantity measurement, parameter assimilation, and edge prediction equipment for small flash flood watersheds, characterized in that, include: Flight platform; A modular payload bay, which is installed at one end of the lower part of the flight platform; A multi-source sensing unit is installed inside the modular payload compartment. The multi-source sensing unit includes a GNSS-R direct / reflected signal receiving module, a lidar module, and a multispectral imaging module. A spatiotemporal synchronization unit is installed inside the modular payload compartment and is connected to a multi-source sensing unit. An edge computing unit, located inside a modular payload compartment and connected to a multi-source sensing unit, utilizes a lidar module and a multispectral imaging module to generate surface soil moisture state parameters with confidence level and observation error covariance. Based on computation time budget, remaining power, computational load, grid size, and state parameter confidence level, the edge computing unit selects an event-driven flow model, a local hydrodynamic model, or a physically constrained inundation proxy model, and outputs the cross-sectional flow process, characteristic cross-sectional water level, inundation range, water depth, risk level, and corresponding uncertainties. A communication unit is located inside the modular payload compartment and is connected to a multi-source sensing unit. A power thermal management unit is located inside the modular payload compartment and is connected to a multi-source sensing unit.

2. The UAV-equipped hydrological state quantity measurement, parameter assimilation, and edge prediction equipment for flash flood small watersheds according to claim 1, characterized in that, The flight platform is an industrial-grade multi-rotor UAV with RTK or PPK positioning capabilities and multiple payload installation capabilities.

3. The UAV-equipped hydrological state quantity measurement, parameter assimilation, and edge prediction equipment for flash flood watersheds as described in claim 1, characterized in that, The workflow of the edge computing unit is as follows: A. Read the known watershed boundary data, publicly available or measured digital elevation model data, and optional land use and hydrological soil group data of the target flash flood watershed, and pre-generate the hydrological response unit HRU grid or regular raster spatial index. B. Use the HRU grid or regular raster spatial index as a spatial reference for UAV route retrieval, multi-source observation footprint matching, and model calculation; C. Unify GNSS-R direct / reflection observations, lidar point clouds, multispectral imagery, and UAV attitude positioning data to the same time and spatial coordinate reference, and construct a state quantity data cube for the target flash flood watershed. D. Using the digital elevation model generated by the lidar module and the attitude positioning data of the UAV, the specular reflection point, local incident angle and surface roughness attenuation of the GNSS-R reflection signal are corrected, and the vegetation transmission loss of the GNSS-R reflection signal is compensated by the vegetation parameters obtained by the multispectral imaging module, and the surface soil moisture state quantity with confidence and observation error covariance is generated. E. Based on the digital elevation model, multispectral images, and surface soil moisture state data, extract underlying surface structural features, river cross-sectional features, and flood control and water-blocking obstacle features; F. Map and assimilate the state variables and structural features to update them into hydrological and hydrodynamic parameters, which include at least the number of dynamic runoff curves. Initial loss rate Manning roughness Compared with water-blocking cross section ; G. Within the edge computing unit, based on the computing time budget, remaining power, computing load, grid size, and state quantity confidence, select the event generation and confluence model, local hydrodynamic model, or physically constrained inundation proxy model, and output the cross-sectional flow process, characteristic cross-sectional water level, inundation range, water depth, risk level, and corresponding uncertainties. H. Write the state variable data cube, correction coefficients, parameter posterior values, error covariance, model residuals, low confidence regions, and observation quality identifiers of this mission into the watershed event memory. In subsequent exploration missions, use the historical parameter posteriors and error covariances in the watershed event memory as prior inputs for parameter assimilation and model selection.

4. The UAV-equipped hydrological state quantity measurement, parameter assimilation, and edge prediction equipment for flash flood watersheds according to claim 1, characterized in that, The GNSS-R direct / reflected signal receiving module includes a right-hand circularly polarized GNSS antenna mounted on the top of the UAV body, a left-hand circularly polarized or dual-circularly polarized GNSS antenna mounted on the bottom of the UAV body, a low-noise amplifier, an RF filter, a dual-channel coherent software radio receiver, and a baseband processing board. The right-hand circularly polarized GNSS antenna is used to receive direct GNSS signals, and the left-hand circularly polarized or dual-circularly polarized GNSS antenna is used to receive GNSS signals reflected from the ground surface.

5. The UAV-equipped hydrological state quantity measurement, parameter assimilation, and edge prediction equipment for flash flood small watersheds according to claim 1, characterized in that, The lidar module is an airborne lidar scanning module integrating an inertial measurement unit and positioning calculation capabilities, installed on the main lower gimbal of the UAV or the main lower mounting interface; the multispectral imaging module is a lightweight five-band or five-band plus panchromatic multispectral camera, and is installed on the UAV through a second lower interface, SkyPort interface, QuickMount interface or independent quick-release bracket.

6. A UAV-equipped system for hydrological state measurement, parameter assimilation, and edge prediction in a small flash flood watershed, characterized in that: The device, ground station, and early warning platform described in any one of claims 1-5 are included; the ground station is used to receive status variables, parameters, risk products, and quality control information output by the device, and to perform post-processing verification, baseline HRU spatial index maintenance, historical sample library maintenance, model retraining, or watershed event memory backup; the early warning platform is used to generate flash flood short-term early warning information based on the risk level and uncertainty.

7. A method for measuring hydrological state quantities, assimilating parameters, and predicting edge conditions using the equipment described in any one of claims 1 to 5, characterized in that, Includes the following steps: Based on known target watershed boundaries and publicly available or measured digital elevation models, a pre-generated HRU spatial index is generated; flight routes covering key runoff areas, confluence channels, river cross-sections, and potential hazard points are planned; GNSS-R direct / reflected signals, lidar point clouds, multispectral imagery, and UAV attitude positioning data are collected simultaneously; a state quantity data cube is constructed; GNSS-R topography, roughness, and vegetation compensation are performed in mountainous areas to obtain surface soil moisture state quantities with confidence level and observation error covariance; underlying surface structure features, river cross-section features, and water-blocking obstacle features are extracted; The state variables and structural features are mapped and assimilated to update the dynamic runoff curve number. Initial loss rate Manning roughness Compared with water-blocking cross section ; Execute event generation and confluence models, local hydrodynamic models, or physically constrained inundation proxy models at the edge; output cross-sectional flow processes, water levels, inundation ranges, water depths, risk levels, and uncertainties; and write the posterior parameters, error covariance, model residuals, and low-confidence regions of this mission into the watershed event memory as priors for subsequent exploration missions.

8. The method for measuring hydrological state quantities, assimilating parameters, and predicting edges using the equipment described in claim 7, characterized in that, The number of dynamic runoff curves Initial loss rate Manning roughness The calculation formula is: ； ； ； Where VI is the vegetation index, slope is the gradient, impervious is the impermeability, roughness is the surface roughness, M is the historical correction term or posterior parameter term provided by the watershed event memory, clip indicates truncation by physical boundary, and θ is the estimated soil moisture value. This is the initial value for the dynamic runoff curve number. Let the initial value be the Manning roughness. This is the initial value of the initial loss rate. For production flow parameters The lower limit of the value, For production flow parameters The upper limit of the value, Loss rate coefficient The lower limit value, Loss rate coefficient The upper limit, This represents the lower limit of the Manning surface roughness. This represents the upper limit of the surface roughness of the Manning landmass. This is the dynamic correction amount for the CN value. Loss rate coefficient The dynamic correction amount, This is the dynamic correction amount for surface roughness.

9. The method for measuring hydrological state quantities, assimilating parameters, and predicting edges using the equipment according to claim 8, characterized in that, The number of dynamic runoff curves and initial loss rate Used for event runoff calculation, when rainfall P is greater than initial loss. At that time, the runoff depth Q satisfies Q=(PI) a ) ² / (PI a +S), when P is not greater than I a When Q=0; In the formula, I a =λ t S, S=25400 / CN t -254, P is rainfall, Q is runoff depth, S is the initial water storage capacity of the watershed, I a This is the initial loss.

10. The method for measuring hydrological state quantities, assimilating parameters, and predicting edges using the equipment according to claim 7, characterized in that, The water-blocking cross-sectional ratio Defined as the projected area of ​​an obstacle in the effective flood discharge section. With effective flood discharge cross-sectional area The ratio, the formula is: ； In the formula, The water-blocking cross-section ratio, For the projected area, The effective flood discharge cross-sectional area.

Images