Deep learning flood forecasting method based on hydrological state modulation confluence propagation operator

Abstract

The application discloses a deep learning flood prediction method based on hydrological state modulation confluence propagation operators, first, target area data is acquired and preprocessed to obtain multi-source hydrological data; according to the time scale difference of each variable in the multi-source hydrological data, time memory characteristics are extracted, and then flood triggering condition representation controlled by multi-source hydrological memory is obtained; then, the river network structure of the target area is extracted to obtain a static river network connection relationship, and the static river network connection relationship is time-varying corrected to obtain a hydrological state modulation confluence propagation operator; the flood triggering condition representation is input into a convolutional neural network, the convolution kernel weight of the convolutional neural network is structured modulated and dynamically reconstructed by using the hydrological state modulation confluence propagation operator, so that the response weight of the convolution kernel in the spatial convolution operation is constrained by the hydrological confluence path, and after spatial propagation calculation by the convolutional neural network, the flood occurrence probability prediction result of each grid unit is output.

Description

Technical Field

[0001] This invention relates to the fields of GNSS hydrological geography and intelligent disaster early warning technology, specifically a deep learning flood prediction method based on hydrological state modulation confluence propagation operators. Background Technology

[0002] Existing flood forecasting methods based on ground-based observation station data mainly rely on long-term continuous observation data from rain gauges, water level stations, and flow stations, combined with hydrological mechanism models, statistical analysis methods, and hydrodynamic numerical models to conduct flood simulation and prediction. These methods can directly characterize the correlation between rainfall processes and watershed response based on high-frequency continuous observation data from rain gauges, water level stations, and flow stations, offering the advantage of high time response accuracy in terms of data source. Furthermore, hydrological mechanism models such as the Xin'anjiang model and the Soil and Water Assessment Tool (SWAT) model can explicitly express the hydrological mechanism of rainfall-to-runoff conversion, exhibiting strong process interpretation capabilities and hydrological consistency. Statistical analysis methods such as the Autoregressive Integrated Moving Average (ARIMA) model, multiple regression models, and frequency analysis methods can mine the temporal correlation and statistical regularity of flood changes based on historical observation sequences, enabling rapid estimation of flood trends and probabilistic characteristics in the absence of complex physical parameters, and demonstrating high computational efficiency. Hydrodynamic numerical models, based on conservation equations, numerically simulate the movement of floods in river channels and floodplains, exhibiting high physical accuracy in flood inundation simulation and spatial propagation analysis. However, this flood modeling method based on ground observation data also has shortcomings: the structure of the hydrological mechanism model and the watershed topology are usually in a fixed, pre-set form, making it difficult to dynamically adjust the spatial hydrological connection strength according to the actual hydrological process; the statistical analysis method is mainly based on empirical fitting of historical observation sequences, lacking the ability to explicitly express the physical mechanism of the hydrological process; the calculation process of the hydrodynamic numerical model depends on high-resolution topographic data and fine boundary condition settings, which has high data requirements and is limited in large-scale watershed applications.

[0003] With the development of space geodesy technology, the Gravity Recovery and Climate Experiment (GRACE) satellite and its follow-up mission (GRACE Follow-On, GRACE-FO) have provided important data sources for regional and global-scale flood monitoring by inverting TWS changes through time-varying gravity fields. Terrestrial water storage, including surface water, soil water, groundwater, snowmelt, and vegetation water, can reflect the overall water storage status of a watershed. Based on this, the Flood Potential Index (FPI) and the Integrated Flood Potential Index (IFPI) are constructed to monitor flood events. GRACE overcomes the spatial distribution limitations of ground observation stations, possessing global-scale spatial coverage capabilities and enabling continuous monitoring of water storage in large-scale watersheds. However, GRACE products suffer from low temporal and spatial resolution, product release delays, and a one-year data gap between GRACE and GRACE-FO. Furthermore, the stability of flood monitoring relying solely on inverted TWS data is significantly affected by data quality, making it difficult to support high-precision, temporally continuous, and stable flood monitoring and forecasting.

[0004] The development of GNSS observation technology has provided a new means of acquiring hydrological information for flood monitoring. GNSS provides millimeter-scale daily station vertical displacement observation data that can accurately invert the dynamic changes of global or regional TWS (Transient Wave Delay) based on elastic load theory. Simultaneously, GNSS tropospheric delay can be inverted to obtain PWV (Potential Water Vapor Value), providing information on water vapor accumulation and potential triggering conditions before precipitation events. GNSS data has advantages such as all-weather operation, high temporal resolution, good continuity, and high accuracy, providing a high-quality data source for flood monitoring and forecasting. However, the PWV and TWS data obtained from GNSS inversion correspond to different physical roles in the flood formation and evolution mechanism, such as atmospheric water vapor supply and pre-flood water storage status, exhibiting significantly different temporal scale characteristics and memory effects. Therefore, it is necessary to construct multi-temporal scale memory representation mechanisms for different variables and further consider their synergistic relationship modeling in the flood triggering and evolution process.

[0005] In recent years, with the development of machine learning and deep learning technologies, data-driven deep learning-based flood prediction methods have gradually become a research hotspot. Related methods are mainly based on Artificial Neural Networks (ANNs) and Long Short-Term Memory (LSTM) networks, using precipitation, runoff, water level, and remote sensing hydrological data as inputs to construct features and automatically mine complex nonlinear relationships between multi-source hydrological data, showing good application potential in flood process simulation and prediction. However, existing deep learning methods typically treat multi-source hydrological variables as equally important input features for unified fusion modeling, lacking modeling of the differences in different hydrological variables across time scales. Secondly, these methods often rely on convolutional or recurrent structures to perform local or sequential modeling of raster data, primarily depending on data-driven learning of implicit spatial correlations. They lack physical constraints on the path changes and dynamic intensity adjustment mechanisms of hydrological connectivity within the watershed under different hydrological conditions, and further lack structured reconstruction of convolution kernel parameters driven by hydrological mechanisms. Therefore, they cannot achieve constraints on the real hydrological propagation mechanism that constrains the dynamic evolution of flood propagation paths and intensity with hydrological conditions.

[0006] Therefore, there is an urgent need to propose a deep learning method that integrates multi-source hydrological data as input, performs hierarchical representation and modeling based on the characteristic differences of different hydrological variables, introduces dynamic modulation of the spatial propagation process by hydrological state to construct a dynamic confluence propagation operator, and reconstructs the convolution kernel of CNN so that the propagation path of floods conforms to the dynamic watershed real hydrological connectivity structure and river network confluence law. This method can effectively improve the stability, accuracy and temporal continuity of flood prediction, enhance the physical expression of the hydrological laws of flood formation mechanism and triggering conditions, and strengthen the ability to characterize the changes in flood propagation path and intensity with hydrological state. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a deep learning flood prediction method based on a hydrological state modulation confluence propagation operator. This method predicts flood events by reconstructing the physical constraints of the convolution kernel using the hydrological state modulation confluence propagation operator. This invention integrates a flood triggering condition representation controlled by multi-source hydrological memory, and further uses the static river network connectivity as a structural basis. Based on the hydrological state modulation function obtained from four types of hydrological data, its propagation characteristics are dynamically reconstructed to form a hydrological state modulation confluence propagation operator. On this basis, the flood triggering condition representation is used as the model input signal. The hydrological state modulation confluence propagation operator is introduced to reconstruct and constrain the propagation weights of the spatial convolution kernel, achieving dynamic representation and prediction of the flood spatial propagation process under confluence path control.

[0008] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows:

[0009] A deep learning-based flood prediction method based on hydrological state modulation confluence propagation operators includes the following steps:

[0010] Step 1: Acquire data from the target area and preprocess it to obtain multi-source hydrological data, including atmospheric precipitable water, terrestrial water storage, precipitation data, and soil moisture data;

[0011] Step 2: Extract time memory features based on the time scale differences of each variable in the multi-source hydrological data. Construct the previous hydrological condition item and precipitation event trigger item based on the time memory features of different variables. Then, obtain the flood triggering condition characterization controlled by multi-source hydrological memory based on the previous hydrological condition item and precipitation event trigger item.

[0012] Step 3: Extract the river network structure of the target area to obtain the static river network connectivity. Construct a hydrological state modulation function based on multi-source hydrological data. Use the hydrological state modulation function to perform time-varying correction on the static river network connectivity to obtain the hydrological state modulation confluence propagation operator.

[0013] Step 4: Input the flood triggering condition representation into the convolutional neural network, and use the hydrological state modulation confluence propagation operator to perform structured modulation and dynamic reconstruction of the convolution kernel weights of the convolutional neural network, so that the response weights of the convolution kernel in the spatial convolution operation are constrained by the hydrological confluence path. After the convolutional neural network performs spatial propagation calculation, the flood occurrence probability prediction results of each grid unit are output.

[0014] Preferably, the preprocessing method includes performing interpolation, spatiotemporal alignment, and normalization on atmospheric precipitable water, terrestrial water storage, precipitation data, and soil moisture data to obtain multi-source hydrological data.

[0015] Preferably, in step 2, an early hydrological condition term is constructed based on the time memory characteristics of soil moisture and terrestrial water storage, and a precipitation event trigger term is constructed based on the time memory characteristics of atmospheric precipitable water and precipitation. The precipitation event trigger term is used as the basic flood trigger intensity, and the nonlinear enhancement modulation of the precipitation event trigger term by the early hydrological condition term is introduced to obtain a flood trigger condition characterization controlled by multi-source hydrological memory.

[0016] Preferably, the preliminary hydrological condition terms are constructed as follows:

[0017] A watershed wet state variable is constructed by weighted fusion of the time memory characteristics of soil moisture and the time memory characteristics of terrestrial water storage. A nonlinear saturation response mechanism is introduced into the watershed wet state variable. The nonlinear saturation response mechanism is as follows: when the watershed wet state variable exceeds a preset saturation critical threshold, a nonlinear enhancement response is triggered, and the watershed wet state variable is multiplied by the excess part after transformation by a nonlinear mapping function to obtain the previous hydrological condition term.

[0018] Preferably, the precipitation event triggering item is constructed as follows:

[0019] Based on the temporal memory characteristics of precipitation, absolute precipitation intensity factor, abrupt precipitation intensity factor, and anomaly precipitation intensity factor are extracted. Based on the temporal memory characteristics of atmospheric precipitable water, atmospheric precipitation potential factor is extracted. The absolute precipitation intensity factor, abrupt precipitation intensity factor, anomaly precipitation intensity factor, and atmospheric precipitation potential factor are weighted and fused, and after nonlinear mapping, a unified triggering intensity function is obtained. The temporal memory characteristics of precipitation are multiplied by the unified triggering intensity function to obtain the precipitation event triggering item.

[0020] Preferably, the nonlinear enhancement modulation of the precipitation event triggering term by introducing the prior hydrological condition term is as follows:

[0021] The precipitation event triggering term is used as the base value and multiplied by the product of the previous hydrological condition term and the preset fusion weight parameter to obtain the flood triggering condition characterization modulated by the previous hydrological conditions, so that the previous wet state of the watershed plays a nonlinear enhancement role in the flood triggering effect of precipitation events.

[0022] As a preferred method, the inversion method for atmospheric precipitable water is as follows:

[0023] Observational data from Global Navigation Satellite System (GNSS) stations are acquired, and data processing is performed using precise point positioning. The station coordinates and total tropospheric zenith delay are calculated by combining precise ephemeris and precise clock error products. Based on the air pressure, latitude, and elevation information of the station location, the zenith dry delay is calculated using the Saastamonin model. The zenith wet delay is obtained by subtracting the zenith dry delay from the total tropospheric zenith delay. The zenith wet delay is then converted into precipitable water content using the conversion relationship between the zenith wet delay and precipitable water content.

[0024] Preferably, the interpolation spatiotemporal alignment method is as follows:

[0025] Elevation data corresponding to receivers of each Global Navigation Satellite System (GNSS) within the study area are obtained. For each target interpolation grid point, the horizontal spatial distance and elevation difference between the target grid point and each GNSS receiver are calculated. A comprehensive weighting function is constructed based on the horizontal spatial distance and elevation difference. The comprehensive weighting function is inversely proportional to the power of the horizontal spatial distance and directly proportional to the negative exponential function of the elevation difference. The atmospheric precipitable water at each GNSS receiver is weighted and estimated based on the comprehensive weighting function to obtain the interpolated atmospheric precipitable water at the target grid point.

[0026] Preferably, in step 3, the static river network connectivity is constructed as follows:

[0027] Based on digital elevation model data, an eight-neighbor deterministic flow direction algorithm is used to determine the flow direction of each grid cell and establish a flow direction field. The upstream confluence contribution area of ​​each grid cell is calculated based on the flow direction field to obtain the cumulative flow value. The dominant confluence path is extracted based on a preset cumulative flow threshold to obtain a binary river network structure. The river network constraint connectivity relationship between grid cells is determined based on the upstream and downstream connection relationship determined by the eight-neighbor deterministic flow direction algorithm. A distance attenuation function is constructed based on the spatial distance between grid cells. A confluence capacity modulation factor is constructed based on the ratio of cumulative flow values ​​between grid cells. The static river network connectivity relationship is obtained by multiplying the river network constraint connectivity relationship, the distance attenuation function, and the confluence capacity modulation factor.

[0028] Preferably, the hydrological state modulation function is constructed as follows:

[0029] An upstream inflow driving term is constructed based on precipitation data, soil moisture data, and atmospheric precipitable water data of each grid cell. The upstream inflow driving term is used to characterize the runoff generation capacity of the grid cell. It is proportional to the precipitation data and is nonlinearly enhanced and modulated by the soil moisture data and atmospheric precipitable water data.

[0030] A hydrological state space similarity term is constructed based on the distance between the multi-source hydrological state vectors of each grid cell. The hydrological state space similarity term is used to characterize the possibility of adjacent grid cells coordinating responses in space.

[0031] A dynamic propagation attenuation term is constructed based on soil moisture data and terrestrial water storage data of downstream grid cells and the spatial distance between grid cells. The dynamic propagation attenuation term is used to characterize the energy loss caused by infiltration and storage of the underlying surface during the spatial propagation of floodwater.

[0032] The hydrological state modulation function is obtained by multiplying the upstream water inflow driving term, the hydrological state space similarity term, and the dynamic propagation attenuation term.

[0033] Preferably, the method of using the hydrological state modulation function to perform time-varying correction on the static river network connectivity is as follows:

[0034] Multiplying the static river network connectivity relationship with the hydrological state modulation function yields a spatiotemporal dynamic hydrological state modulation confluence propagation operator, which dynamically adjusts the confluence propagation intensity according to changes in atmospheric water vapor conditions, precipitation input intensity, soil moisture state, and terrestrial water storage state.

[0035] Preferably, the method for structurally modulating and dynamically reconstructing the convolutional kernel weights of the convolutional neural network is as follows:

[0036] The hydrological state modulation confluence propagation operator is multiplied element-wise with the weights of the basic convolutional kernel of the convolutional neural network to obtain a dynamic convolutional kernel, so that the spatial propagation weight distribution of the dynamic convolutional kernel is controlled in real time by the hydrological confluence path.

[0037] When the convolutional neural network performs spatial propagation calculations, the response of the convolutional kernel is determined by the weighted sum of the flood triggering condition characterization and the dynamic convolutional kernel, thereby enabling the information contribution between different grid units to be controlled by the hydrological state modulation confluence propagation operator.

[0038] Preferably, the training method for the convolutional neural network is as follows:

[0039] Obtain remote sensing flood inundation products and extract the real flood labels for each grid cell;

[0040] The binary cross-entropy loss function is used to calculate the data fitting error between the predicted flood probability and the actual flood label;

[0041] Based on the hydrological state modulation confluence propagation operator, a hydrological structure consistency constraint term is constructed. The hydrological structure consistency constraint term is used to constrain the difference in predicted flood probability between adjacent grid cells to be positively correlated with their hydrological connectivity strength.

[0042] The weighted sum of the data fitting error and the hydrological structure consistency constraint term is used as the overall optimization objective to iteratively optimize and solve the model parameters of the convolutional neural network.

[0043] Preferably, the following steps are also included:

[0044] The flood occurrence probability prediction results of each grid cell are organized into a spatial matrix according to spatial location and drawn into a flood probability distribution map to realize the visualization of the regional flood probability distribution;

[0045] High-risk flood areas are identified based on preset probability thresholds;

[0046] The system continuously outputs prediction results for multiple time steps to construct a flood evolution sequence, which is used to characterize the spatial expansion and propagation trend of floods, and realize flood event early warning and propagation prediction.

[0047] Preferably, the global navigation satellite system observation data includes tropospheric delay observation data and vertical displacement observation data from global navigation satellite system stations; the atmospheric precipitable water is obtained by inversion from the tropospheric delay observation data, and the terrestrial water storage is obtained by inversion from the vertical displacement observation data based on the elastic load theory.

[0048] This invention has the following characteristics and beneficial effects:

[0049] Multi-source data sources ensure stable prediction with small errors: This invention combines PWV, TWS, precipitation, and soil moisture data obtained from GNSS inversion, making full use of the advantages of different observation methods and solving the problems of low spatiotemporal resolution and limited prediction accuracy in traditional single-source prediction.

[0050] Enhanced ability to express flood mechanisms: Based on the hydrological evolution law in the flood formation process, this invention constructs a flood triggering condition representation based on the nonlinear modulation of the previous hydrological condition term and the precipitation event trigger term. Before the model input, it explicitly constructs a division of labor structure that conforms to the flood formation mechanism, thereby improving the physical consistency of the flood formation process.

[0051] Multi-scale temporal memory enhances early-stage hydrological characterization: By constructing temporal memory windows of different lengths to address the differences in the time scales of the effects of different hydrological variables on the flood formation process, a layered characterization of short-term atmospheric water vapor transport and precipitation, medium-term runoff generation processes, and long-term water storage changes can be achieved.

[0052] Enhanced consistency of actual confluence under river network path constraints: This ensures that the spatial propagation of floods strictly conforms to the unidirectional confluence pattern of natural river networks, enabling continuous advancement along the upstream-to-downstream path controlled by the actual terrain, thereby improving the consistency between the flood propagation path and the actual watershed confluence process.

[0053] To achieve dynamic adjustment of flood propagation path according to hydrological conditions: a dynamic propagation operator is introduced, which enables the convolution kernel weight to be dynamically adjusted according to changes in multiple hydrological conditions such as PWV, precipitation, SM (soil moisture), and TWS. This is more in line with the non-fixed evolution process of the path of real floods under different hydrological backgrounds, which is controlled by runoff generation and underlying surface hydrological conditions.

[0054] Wide applicability: This invention utilizes the complementary advantages of multi-source data, maintaining good prediction stability even in areas with limited observation data. It can be applied to real flood prediction situations, expanding the regional applicability and practical transferability of the method. Attached Figure Description

[0055] Figure 1 This is a flowchart of a deep learning flood prediction method based on a hydrological state modulation confluence propagation operator, according to an embodiment of the present invention.

[0056] Figure 2 This is a flowchart illustrating the construction of flood triggering condition characterization for multi-source hydrological memory control in this embodiment of the invention.

[0057] Figure 3 This is a flowchart illustrating the construction of the hydrological state modulation confluence propagation operator in an embodiment of the present invention.

[0058] Figure 4 This is a flowchart of a convolutional neural network flood prediction model based on hydrological state confluence propagation operators in an embodiment of the present invention.

[0059] Figure 5 This is a schematic diagram comparing the dynamic evolution of convolutional kernels based on hydrological state modulation in an embodiment of the present invention. Detailed Implementation

[0060] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0061] A deep learning flood prediction method based on hydrological state modulation confluence propagation operators, such as Figure 1 As shown, it includes the following steps:

[0062] Step 1: Acquire data from the target area and preprocess it to obtain multi-source hydrological data, including atmospheric precipitable water, terrestrial water storage, precipitation data, and soil moisture data.

[0063] It should be noted that the atmospheric precipitable water (PWV) in this embodiment is obtained by inversion using GNSS tropospheric delay data.

[0064] In this embodiment, GNSS station observation data for the target area is acquired, and data processing is performed using Precise Point Positioning (PPP). The station coordinates and Zenith Total Delay (ZTD) parameters are calculated by combining precise ephemeris and precise clock difference products from the International GNSS Service (IGS). Subsequently, based on the air pressure, latitude, and elevation information of the station location, the Zenith Hydrostatic Delay (ZHD) is calculated using the Saastamoinen model, and the Zenith Wet Delay (ZWD) is obtained by the difference between ZTD and ZHD. Furthermore, the conversion relationship between Zenith Wet Delay and atmospheric precipitable water is utilized to convert the Zenith Wet Delay into atmospheric precipitable water volume (PWV).

[0065] Specifically, the PPP calculation process uses IGS precise ephemeris and satellite clock products for positioning and estimates tropospheric delay parameters. ZTD consists of two parts: ZHD and ZWD, where ZHD is calculated using the Saastamoinen model.

[0066] ;

[0067] in, Indicates the latitude of the receiver; Indicates altitude; This indicates the pressure at sea level where the receiver is located.

[0068] ZWD can be calculated from ZTD and ZHD as follows:

[0069] ;

[0070] Next, using the relationship between tropospheric wet delay and atmospheric precipitable water, the ZWD is converted to PWV:

[0071] ;

[0072] in, The transformation factor is expressed as follows:

[0073] ;

[0074] in, Represents the density of liquid water vapor; The vapor-to-gas ratio constant; It is the atmospheric refractive constant; This represents the weighted average temperature of the atmosphere.

[0075] Furthermore, in this embodiment, the terrestrial water storage (TWS) is obtained by inversion using GNSS observation data.

[0076] In this embodiment, continuous coordinate time series data of GNSS stations are acquired and preprocessed, including outlier removal, step de-stepping, missing data interpolation, and trend term removal, to obtain a stable and continuous vertical displacement time series. Based on the theory of Earth's elastic load, a functional relationship model between mass load and station vertical displacement is constructed using the Green's function method, and an inversion equation is built with the grid equivalent water height as the unknown parameter. To improve the spatial continuity and stability of the inversion results, Laplace smoothing constraints are introduced during the solution process to regularize the solution of the regional equivalent water height, ultimately obtaining the TWS time series data of the study area.

[0077] Specifically, the Green's function matrix of the vertical displacement of the ground surface caused by the load can be expressed as:

[0078] ;

[0079] in, For the load Love number, ; , n > 0; is Newton's gravitational constant; It is the acceleration due to gravity; For Legendre polynomials; Angular distance between any two points on Earth; The radius of the Earth; This indicates the angular radius of the disk.

[0080] The inversion model based on Laplace matrix constraints (Argus et al. (2014)) is as follows:

[0081] ;

[0082] in, The standard deviation is the vector. The coefficient matrix of the Green's function; The equivalent water level of the grid points is high; For GNSS station vertical displacement observations; It is a smoothing factor; For the Laplace operator.

[0083] The equivalent water height for each grid within the region can be obtained from this model as follows:

[0084] .

[0085] Furthermore, the preprocessing method includes performing interpolation, spatiotemporal alignment, and normalization on atmospheric precipitable water, terrestrial water storage, precipitation data, and soil moisture data to obtain multi-source hydrological data.

[0086] In this embodiment, precipitation data and soil moisture (SM) from ground observation stations are acquired, and missing temporal values ​​are interpolated. A unified grid is used to divide the study area. The inverted TWS and SM are grid products and do not require interpolation. Precipitation data are interpolated to the corresponding grid using Kriging interpolation. Considering that PWV is significantly affected by topography and elevation, a constrained interpolation method based on horizontal spatial distance and elevation differences in DEM data is used. After spatiotemporal interpolation and alignment of the four types of data, normalization processing is performed to construct multi-source observation data at a unified scale.

[0087] Specifically, the PWV and TWS data obtained from inversion within the study area, along with the observed precipitation and SM data, were organized into corresponding time series. A unified grid was defined within the area. TWS and SM data are inherently gridded data and require no spatial interpolation. Precipitation data, derived from ground observation stations, was interpolated to the corresponding grid using Kriging interpolation. PWV data was inverted from GNSS station data; due to its significant influence from terrain elevation, a constrained interpolation method incorporating horizontal spatial distance and elevation differences was constructed. First, a comprehensive weighting function considering both spatial distance and elevation differences was constructed:

[0088] ;

[0089] in, , Target grid points and the first to participate in interpolation An GNSS receiver based on DEM elevation; This refers to the horizontal spatial distance between the target grid point and each GNSS receiver; This is the distance decay index, used to control the degree of influence of spatial distance on the weights; This is the elevation attenuation coefficient, used to adjust the degree of influence of terrain elevation difference in the weighting function; Represents target grid points With the The terrain elevation difference of each GNSS receiver.

[0090] The weighted estimate of PWV at the target grid point is performed using the weighting function described above, and its expression is as follows:

[0091] ;

[0092] in, For interpolation to the grid The PWV value on; For GNSS receiver PWV values ​​obtained from data inversion; This indicates the number of PWV data points corresponding to the GNSS receivers participating in the interpolation.

[0093] After interpolation and spatial-temporal alignment, the four types of data are normalized to map each variable to a uniform numerical scale.

[0094] Step 2: Extract time memory features based on the time scale differences of each variable in the multi-source hydrological data. Construct the previous hydrological condition item and precipitation event trigger item based on the time memory features of different variables. Then, obtain the flood triggering condition characterization controlled by multi-source hydrological memory based on the previous hydrological condition item and precipitation event trigger item.

[0095] In this embodiment, a preliminary hydrological condition term is constructed based on the temporal memory characteristics of soil moisture and terrestrial water storage, and a precipitation event trigger term is constructed based on the temporal memory characteristics of atmospheric precipitable water and precipitation. The precipitation event trigger term is used as the basic flood trigger intensity, and the nonlinear enhancement modulation of the precipitation event trigger term by the preliminary hydrological condition term is introduced to obtain a flood trigger condition characterization controlled by multi-source hydrological memory.

[0096] The construction method of the preliminary hydrological condition term is as follows: a watershed wet state variable is constructed based on the weighted fusion of the time memory characteristics of soil moisture and the time memory characteristics of terrestrial water storage; a nonlinear saturation response mechanism is introduced into the watershed wet state variable, which is as follows: when the watershed wet state variable exceeds a preset saturation critical threshold, a nonlinear enhancement response is triggered, and the watershed wet state variable is multiplied by the excess part after transformation by the nonlinear mapping function to obtain the preliminary hydrological condition term.

[0097] Specifically, such as Figure 2 As shown, firstly, a unified time memory representation process is applied to the multi-source hydrological time series. Different hydrological variables have different time scales of influence in the flood formation mechanism: precipitation usually exhibits a short-term triggering effect, PWV reflects short- to medium-term atmospheric transport processes, SM reflects the evolution of medium-term runoff conditions, while TWS characterizes the long-term water storage state of the basin. Therefore, this invention constructs time memory features by setting multi-scale time memory window lengths for different variables:

[0098] ;

[0099] in, ; This indicates the weight of contribution of different lag times to the formation of the current flood; Represents the observed values ​​of a variable at historical moments; Representing variables The time memory window length is used to limit the effective scope of historical information, and the length is set according to the response characteristics of different hydrological variables.

[0100] Then based on , An early hydrological condition term is constructed to characterize the overall wetting conditions and critical response state of the watershed before flooding. First, watershed wetting state variables are constructed using SM and TWS memory features:

[0101] ;

[0102] in, , These are learnable parameters.

[0103] The construction method of the preliminary hydrological condition term is as follows: a watershed wet state variable is constructed based on the weighted fusion of the time memory characteristics of soil moisture and the time memory characteristics of terrestrial water storage; a nonlinear saturation response mechanism is introduced into the watershed wet state variable, which is as follows: when the watershed wet state variable exceeds a preset saturation critical threshold, a nonlinear enhancement response is triggered, and the watershed wet state variable is multiplied by the excess part after transformation by the nonlinear mapping function to obtain the preliminary hydrological condition term, as shown in the following expression:

[0104] ;

[0105] in, This is the saturation critical threshold. This represents the Sigmoid nonlinear mapping function.

[0106] Next, build based on and The precipitation event triggering factors are considered. During flood triggering, precipitation not only exhibits an absolute intensity impact but also demonstrates short-term abrupt changes and anomalies relative to the climate background. Therefore, three complementary intensity characterization factors are introduced: absolute precipitation intensity... Sudden changes in precipitation intensity and abnormal intensity of precipitation And on this basis, introduce As an atmospheric potential factor, a unified triggering intensity function is constructed:

[0107] ;

[0108] in, ; ; , , , These are learnable weight parameters; This is the precipitation trigger threshold; This represents the Sigmoid nonlinear mapping function.

[0109] Next, the precipitation event triggering term is obtained from the memory-enhanced precipitation characteristics and the unified triggering intensity function. Specifically, the absolute precipitation intensity factor, abrupt precipitation intensity factor, and anomaly precipitation intensity factor are extracted based on the time memory characteristics of precipitation, and the atmospheric precipitation potential factor is extracted based on the time memory characteristics of atmospheric precipitable water. The absolute precipitation intensity factor, abrupt precipitation intensity factor, anomaly precipitation intensity factor, and atmospheric precipitation potential factor are weighted and fused, and then nonlinearly mapped to obtain the unified triggering intensity function. The precipitation event triggering term is obtained by multiplying the time memory characteristics of precipitation by the unified triggering intensity function, as shown in the following expression:

[0110] ;

[0111] Finally, the preceding hydrological conditions and precipitation event triggering factors are integrated. Specifically, the precipitation event triggering factor is used as a base value and multiplied by the product of the preceding hydrological conditions and a preset fusion weight parameter to obtain a flood triggering condition characterization modulated by the preceding hydrological conditions. This allows the preceding wet conditions of the watershed to exert a nonlinear enhancement effect on the flood triggering effect of precipitation events, resulting in a flood triggering condition characterization controlled by multi-source hydrological memory, expressed as follows:

[0112] ;

[0113] in, The weighting parameters are used to adjust the modulation intensity of the precipitation triggering effect of previous hydrological conditions.

[0114] Step 3: Extract the river network structure of the target area to obtain the static river network connectivity. Construct a hydrological state modulation function based on multi-source hydrological data. Use the hydrological state modulation function to perform time-varying correction on the static river network connectivity to obtain the hydrological state modulation confluence propagation operator.

[0115] In this embodiment, the static river network connectivity is constructed as follows: Based on digital elevation model data, an eight-neighbor deterministic flow direction algorithm is used to determine the flow direction of each grid unit, establishing a flow direction field; the upstream confluence contribution area of ​​each grid unit is calculated based on the flow direction field to obtain the cumulative flow value; the dominant confluence path is extracted based on a preset cumulative flow threshold to obtain a binary river network structure; the river network constraint connectivity relationship between grid units is determined based on the upstream and downstream connection relationship determined by the eight-neighbor deterministic flow direction algorithm; a distance attenuation function is constructed based on the spatial distance between grid units; a confluence capacity modulation factor is constructed based on the ratio of cumulative flow values ​​between grid units; the river network constraint connectivity relationship, the distance attenuation function, and the confluence capacity modulation factor are multiplied together to obtain the static river network connectivity relationship.

[0116] It should be noted that Digital Elevation Model (DEM) data is elevation data that represents the topographic relief of the ground. Each grid records the altitude of a location. It is a type of publicly available topographic elevation raster data that can be downloaded directly.

[0117] Furthermore, the hydrological state modulation function is constructed as follows:

[0118] An upstream inflow driving term is constructed based on precipitation data, soil moisture data, and atmospheric precipitable water data of each grid cell. The upstream inflow driving term is used to characterize the runoff generation capacity of the grid cell. It is proportional to the precipitation data and is nonlinearly enhanced and modulated by the soil moisture data and atmospheric precipitable water data.

[0119] A hydrological state space similarity term is constructed based on the distance between the multi-source hydrological state vectors of each grid cell. The hydrological state space similarity term is used to characterize the possibility of adjacent grid cells coordinating responses in space.

[0120] A dynamic propagation attenuation term is constructed based on soil moisture data and terrestrial water storage data of downstream grid cells and the spatial distance between grid cells. The dynamic propagation attenuation term is used to characterize the energy loss caused by infiltration and storage of the underlying surface during the spatial propagation of floodwater.

[0121] The hydrological state modulation function is obtained by multiplying the upstream water inflow driving term, the hydrological state space similarity term, and the dynamic propagation attenuation term.

[0122] Specifically, such as Figure 3 As shown. First, the regional river network structure is extracted based on DEM data. After filling depressions in the DEM data, the flow direction of each grid cell is calculated using the D8 method, and the direction with the largest slope in the eight neighboring areas is selected as the flow direction:

[0123] ;

[0124] in, For grid The direction of flow; For grid The set of eight neighborhoods; , They are grids and Elevation value; This represents the distance between the centers of the grid.

[0125] The upstream confluence contribution area of ​​each grid is calculated based on the flow field to obtain the cumulative flow value:

[0126] ;

[0127] in, , indicating grid The upstream contribution set.

[0128] Set thresholds based on accumulated traffic. Extracting the dominant confluence paths to form the river network structure:

[0129] ;

[0130] in, For the cumulative traffic threshold; when At that time, the grid represents Belonging to the river network unit, when When, it indicates a grid. It belongs to a non-river network unit.

[0131] Obtaining the binary river network structure Building upon this foundation, the river network structure is extended from a discrete representation to a continuous hydrological propagation intensity relationship. Firstly, based on the D8 flow direction results, a grid is defined. With grid The river network constraint connectivity between them is as follows:

[0132] ;

[0133] in, Represents the mesh determined based on the D8 flow direction algorithm. The downstream neighborhood set; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise.

[0134] Considering the energy loss of floodwater during its propagation through space, a distance attenuation function is introduced:

[0135] ;

[0136] in, This is the distance attenuation coefficient.

[0137] Next, cumulative flow values ​​are introduced to characterize the differences in the flow collection capacity of different river network units:

[0138] ;

[0139] in, It is a numerically stable term.

[0140] Considering the constraints of the river network structure, spatial attenuation effect, and confluence capacity modulation, the hydrological connectivity propagation intensity is defined as:

[0141] ;

[0142] in, For grid To grid The propagation intensity of river network hydrological connectivity reflects the static propagation relationship determined by river network structure, spatial distance, and confluence capacity. However, flood propagation is not only controlled by topographical river network structure but also dynamically influenced by the current multi-source hydrological state. Therefore, three types of physical processes are introduced: upstream inflow driving term, hydrological state spatial similarity term, and propagation attenuation term, to construct a multi-source hydrological state modulation function. Used for static Dynamic modulation is performed to achieve time-varying physical constraint modeling of the flood propagation process.

[0143] First, hydrological state vectors for grid cells are constructed using PWV, precipitation, soil moisture, and terrestrial water storage.

[0144] ;

[0145] During flood propagation, adjacent areas with similar hydrological conditions are more likely to form a coordinated response. Therefore, a hydrological state space similarity term can be constructed using state vector distance:

[0146] ;

[0147] In the process of precipitation runoff formation, precipitation is the direct hydrological input, soil moisture reflects the underlying surface's runoff generation and transformation capacity, and PWV reflects the enhancing effect of atmospheric water vapor background on precipitation events. Therefore, the upstream inflow driving term is constructed as follows:

[0148] ;

[0149] in, Represents a grid Runoff generation capacity; Soil moisture modulation coefficient; This is the water vapor enhancement coefficient.

[0150] During the spatial propagation of floodwaters, propagation attenuation occurs due to surface infiltration capacity, watershed storage capacity, and spatial distance effects. This invention constructs a dynamic propagation attenuation term to characterize the changing characteristics of propagation efficiency under different hydrological conditions:

[0151] ;

[0152] in, Represents a grid To grid The dynamic propagation attenuation modulation factor; This is the soil moisture loss modulation coefficient; The modulation coefficient for land water storage loss; This is the distance attenuation coefficient; Represents a grid With grid The spatial distance between them.

[0153] Combining the three components constructed above, we obtain the dynamic hydrological state modulation function:

[0154] ;

[0155] Finally, the static hydrological connectivity propagation intensity is corrected using a multi-source hydrological state modulation function to obtain the hydrological state modulation confluence propagation operator:

[0156] ;

[0157] Thus, the dynamic spatial propagation path of floods under the constraints of real river networks and changes in real-time hydrological conditions can be characterized by the hydrological state modulation confluence propagation operator.

[0158] Step 4: Input the flood triggering condition representation into the convolutional neural network, and use the hydrological state modulation confluence propagation operator to perform structured modulation and dynamic reconstruction of the convolution kernel weights of the convolutional neural network, so that the response weights of the convolution kernel in the spatial convolution operation are constrained by the hydrological confluence path. After the convolutional neural network performs spatial propagation calculation, the flood occurrence probability prediction results of each grid unit are output.

[0159] In this embodiment, a model of the spatial propagation process of flood is built based on a constructed hydrologically modulated confluence propagation operator. This operator is then introduced into the spatial convolution operation and combined with the basic convolution weights. This causes the convolution kernel to evolve into a propagation weight distribution kernel driven by hydrological conditions and constrained by the river network confluence path. Its response undergoes spatial redistribution and path adjustment under the influence of changes in hydrological conditions. This allows the information contribution between different grids in the CNN during the convolution process to be controlled by the hydrologically modulated dynamic confluence propagation operator. This enables the calculation of the neighborhood propagation of the previous flood trigger intensity along the real confluence path, obtaining a spatial propagation feature representation. Finally, the probability of flood occurrence for each grid at the next moment is output through nonlinear mapping. During model training, remote sensing flood inundation products are used as supervision signals. Error constraints based on the actual spatial distribution of floods are introduced, and a structural consistency constraint term based on the dynamic hydrologically modulated propagation operator is constructed to ensure that the prediction results are consistent with the actual flood evolution process in terms of spatial distribution and propagation path.

[0160] Specifically, such as Figure 4 As shown, the method for structured modulation and dynamic reconstruction of the convolutional neural network kernel weights is as follows: the hydrological state modulation confluence propagation operator is multiplied element-wise with the basic convolutional kernel weights of the convolutional neural network to obtain a dynamic convolutional kernel, so that the spatial propagation weight distribution of the dynamic convolutional kernel is controlled in real time by the hydrological confluence path; when the convolutional neural network performs spatial propagation calculations, the response of the convolutional kernel is determined by the weighted sum of the flood triggering condition characterization and the dynamic convolutional kernel, thereby allowing the information contribution between different grid cells to be controlled by the hydrological state modulation confluence propagation operator.

[0161] use Propagation weights of the basic convolution kernel during spatial convolution. Physical modulation and reconstruction are performed to make the spatial convolution process based on the static river network connectivity relationship, and its propagation characteristics are dynamically reconstructed through hydrological state modulation terms. This embeds the confluence and transport process under hydrological conditions into the spatial weighted propagation calculation, thereby realizing the physically consistent propagation expression of flood driving forces in space.

[0162] Hydrological state modulation confluence propagation operator Physical constraints are applied to modulate the effective propagation weights of the basic spatial convolutional kernel to construct a dynamic convolutional kernel:

[0163] ;

[0164] in, These are the weights of the convolutional kernel in the basic space.

[0165] The dynamic influence of hydrological state modulation on convolution kernel parameters and a comparison diagram between dynamic convolution kernels and basic convolution kernels are shown below. Figure 5 As shown.

[0166] The dynamic convolution kernel is used to characterize the flood formation triggering intensity. Spatial propagation calculations are performed to obtain the convolutional response features:

[0167] ;

[0168] in, For grid The spatial neighborhood set.

[0169] Based on the spatial propagation response, a nonlinear mapping function is introduced to map the spatial propagation characteristics to the probability of flood occurrence:

[0170] ;

[0171] in, Indicates the prediction time step; It is a nonlinear mapping function; This is the Sigmoid function.

[0172] Furthermore, in this embodiment, regarding the training of the model:

[0173] Obtain remote sensing flood inundation products and extract the real flood labels for each grid cell;

[0174] The binary cross-entropy loss function is used to calculate the data fitting error between the predicted flood probability and the actual flood label;

[0175] Based on the hydrological state modulation confluence propagation operator, a hydrological structure consistency constraint term is constructed. The hydrological structure consistency constraint term is used to constrain the difference in predicted flood probability between adjacent grid cells to be positively correlated with their hydrological connectivity strength.

[0176] The weighted sum of the data fitting error and the hydrological structure consistency constraint is used as the overall optimization objective to iteratively optimize the model parameters of the convolutional neural network. The expression is as follows:

[0177] Constructing real labels based on Sentinel-1 SAR flood extraction results And a binary classification cross-entropy loss function is used:

[0178] ;

[0179] in, Labeled as a real flood; To predict the probability of flooding using the model.

[0180] Meanwhile, to ensure that the spatial propagation results of floods conform to the consistency of hydrological connectivity under dynamic regulation, a structural constraint term based on the hydrological state modulation confluence propagation operator is introduced:

[0181] ;

[0182] Therefore, the overall optimization objective of the model consists of both the data fitting error and the hydrological structure constraints:

[0183] ;

[0184] in, The weights are the constraints for the hydrological structure.

[0185] Finally, the model parameters are iteratively optimized based on the overall optimization objective to obtain the converged model parameters. The final output is the spatiotemporal distribution prediction of the flood occurrence probability for each grid cell.

[0186] Understandably, the spatial distribution of floods within the target area is expressed and analyzed based on the above prediction results. The predicted probabilities of each grid are displayed according to spatial location, forming a flood probability distribution map to visualize the likelihood of flooding in different areas at the predicted time. Based on this, the flood-prone areas are divided according to the set probability thresholds and the average flood probability within the area, obtaining the spatial distribution results of the flood's impact range and achieving the identification of flood-prone areas. Simultaneously, the prediction results at different time steps are continuously output to form a flood evolution process sequence, used to characterize the spatial expansion and propagation trends of floods for flood monitoring, risk assessment, and early warning applications.

[0187] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning flood prediction method based on hydrological state modulation confluence propagation operators, characterized in that, Includes the following steps: Step 1: Acquire data from the target area and preprocess it to obtain multi-source hydrological data, including atmospheric precipitable water, terrestrial water storage, precipitation data, and soil moisture data; Step 2: Extract time memory features based on the time scale differences of each variable in the multi-source hydrological data. Construct the previous hydrological condition item and precipitation event trigger item based on the time memory features of different variables. Then, obtain the flood triggering condition characterization controlled by multi-source hydrological memory based on the previous hydrological condition item and precipitation event trigger item. Step 3: Extract the river network structure of the target area to obtain the static river network connectivity. Construct a hydrological state modulation function based on multi-source hydrological data. Use the hydrological state modulation function to perform time-varying correction on the static river network connectivity to obtain the hydrological state modulation confluence propagation operator. Step 4: Input the flood triggering condition representation into the convolutional neural network, and use the hydrological state modulation confluence propagation operator to perform structured modulation and dynamic reconstruction of the convolution kernel weights of the convolutional neural network, so that the response weights of the convolution kernel in the spatial convolution operation are constrained by the hydrological confluence path. After the convolutional neural network performs spatial propagation calculation, the flood occurrence probability prediction results of each grid unit are output.

2. The method according to claim 1, characterized in that, The preprocessing method includes performing interpolation, spatiotemporal alignment, and normalization on atmospheric precipitable water, terrestrial water storage, precipitation data, and soil moisture data to obtain multi-source hydrological data.

3. The method according to claim 1, characterized in that, In step 2, a preliminary hydrological condition term is constructed based on the time memory characteristics of soil moisture and terrestrial water storage, and a precipitation event trigger term is constructed based on the time memory characteristics of atmospheric precipitable water and precipitation. The precipitation event trigger term is used as the basic flood trigger intensity, and the nonlinear enhancement modulation of the precipitation event trigger term by the preliminary hydrological condition term is introduced to obtain a flood trigger condition characterization controlled by multi-source hydrological memory.

4. The method according to claim 3, characterized in that, The method for constructing the aforementioned prior hydrological condition items is as follows: A watershed wet state variable is constructed by weighted fusion of the time memory characteristics of soil moisture and the time memory characteristics of terrestrial water storage. A nonlinear saturation response mechanism is introduced into the watershed wet state variable. The nonlinear saturation response mechanism is as follows: when the watershed wet state variable exceeds a preset saturation critical threshold, a nonlinear enhancement response is triggered, and the watershed wet state variable is multiplied by the excess part after transformation by a nonlinear mapping function to obtain the previous hydrological condition term.

5. The method according to claim 4, characterized in that, The precipitation event trigger is constructed as follows: Based on the temporal memory characteristics of precipitation, absolute precipitation intensity factor, abrupt precipitation intensity factor, and anomaly precipitation intensity factor are extracted. Based on the temporal memory characteristics of atmospheric precipitable water, atmospheric precipitation potential factor is extracted. The absolute precipitation intensity factor, abrupt precipitation intensity factor, anomaly precipitation intensity factor, and atmospheric precipitation potential factor are weighted and fused, and after nonlinear mapping, a unified triggering intensity function is obtained. The temporal memory characteristics of precipitation are multiplied by the unified triggering intensity function to obtain the precipitation event triggering item.

6. The method according to claim 5, characterized in that, The method of nonlinearly enhancing the modulation of the precipitation event triggering term by introducing the previous hydrological condition term is as follows: The precipitation event triggering term is used as the base value and multiplied by the product of the previous hydrological condition term and the preset fusion weight parameter to obtain the flood triggering condition characterization modulated by the previous hydrological conditions, so that the previous wet state of the watershed plays a nonlinear enhancement role in the flood triggering effect of precipitation events.

7. The method according to claim 1, characterized in that, The method for retrieving atmospheric precipitable water is as follows: Observational data from Global Navigation Satellite System (GNSS) stations are acquired, and data processing is performed using precise point positioning. The station coordinates and total tropospheric zenith delay are calculated by combining precise ephemeris and precise clock error products. Based on the air pressure, latitude, and elevation information of the station location, the zenith dry delay is calculated using the Saastamonin model. The zenith wet delay is obtained by subtracting the zenith dry delay from the total tropospheric zenith delay. The zenith wet delay is then converted into precipitable water content using the conversion relationship between the zenith wet delay and precipitable water content.

8. The method according to claim 2, characterized in that, The interpolation spatiotemporal alignment method is as follows: Elevation data corresponding to receivers of each Global Navigation Satellite System (GNSS) within the study area are obtained. For each target interpolation grid point, the horizontal spatial distance and elevation difference between the target grid point and each GNSS receiver are calculated. A comprehensive weighting function is constructed based on the horizontal spatial distance and elevation difference. The comprehensive weighting function is inversely proportional to the power of the horizontal spatial distance and directly proportional to the negative exponential function of the elevation difference. The atmospheric precipitable water at each GNSS receiver is weighted and estimated based on the comprehensive weighting function to obtain the interpolated atmospheric precipitable water at the target grid point.

9. The method according to claim 1, characterized in that, In step 3, the static river network connectivity is constructed as follows: Based on digital elevation model data, an eight-neighbor deterministic flow direction algorithm is used to determine the flow direction of each grid cell and establish a flow direction field. The upstream confluence contribution area of ​​each grid cell is calculated based on the flow direction field to obtain the cumulative flow value. The dominant confluence path is extracted based on a preset cumulative flow threshold to obtain a binary river network structure. The river network constraint connectivity relationship between grid cells is determined based on the upstream and downstream connection relationship determined by the eight-neighbor deterministic flow direction algorithm. A distance attenuation function is constructed based on the spatial distance between grid cells. A confluence capacity modulation factor is constructed based on the ratio of cumulative flow values ​​between grid cells. The static river network connectivity relationship is obtained by multiplying the river network constraint connectivity relationship, the distance attenuation function, and the confluence capacity modulation factor.

10. The method according to claim 1, characterized in that, The hydrological state modulation function is constructed as follows: An upstream inflow driving term is constructed based on precipitation data, soil moisture data, and atmospheric precipitable water data of each grid cell. The upstream inflow driving term is used to characterize the runoff generation capacity of the grid cell. It is proportional to the precipitation data and is nonlinearly enhanced and modulated by the soil moisture data and atmospheric precipitable water data. A hydrological state space similarity term is constructed based on the distance between the multi-source hydrological state vectors of each grid cell. The hydrological state space similarity term is used to characterize the possibility of adjacent grid cells coordinating responses in space. A dynamic propagation attenuation term is constructed based on soil moisture data and terrestrial water storage data of downstream grid cells and the spatial distance between grid cells. The dynamic propagation attenuation term is used to characterize the energy loss caused by infiltration and storage of the underlying surface during the spatial propagation of floodwater. The hydrological state modulation function is obtained by multiplying the upstream water inflow driving term, the hydrological state space similarity term, and the dynamic propagation attenuation term.

11. The method according to claim 1, characterized in that, The method of using hydrological state modulation functions to perform time-varying correction on static river network connectivity is as follows: Multiplying the static river network connectivity relationship with the hydrological state modulation function yields a spatiotemporal dynamic hydrological state modulation confluence propagation operator, which dynamically adjusts the confluence propagation intensity according to changes in atmospheric water vapor conditions, precipitation input intensity, soil moisture state, and terrestrial water storage state.

12. The method according to claim 1, characterized in that, The method for structurally modulating and dynamically reconstructing the convolutional kernel weights of the convolutional neural network is as follows: The hydrological state modulation confluence propagation operator is multiplied element-wise with the weights of the basic convolutional kernel of the convolutional neural network to obtain a dynamic convolutional kernel, so that the spatial propagation weight distribution of the dynamic convolutional kernel is controlled in real time by the hydrological confluence path. When the convolutional neural network performs spatial propagation calculations, the response of the convolutional kernel is determined by the weighted sum of the flood triggering condition characterization and the dynamic convolutional kernel, thereby enabling the information contribution between different grid units to be controlled by the hydrological state modulation confluence propagation operator.

13. The method according to claim 1, characterized in that, The training method for the convolutional neural network is as follows: Obtain remote sensing flood inundation products and extract the real flood labels for each grid cell; The binary cross-entropy loss function is used to calculate the data fitting error between the predicted flood probability and the actual flood label; Based on the hydrological state modulation confluence propagation operator, a hydrological structure consistency constraint term is constructed. The hydrological structure consistency constraint term is used to constrain the difference in predicted flood probability between adjacent grid cells to be positively correlated with their hydrological connectivity strength. The weighted sum of the data fitting error and the hydrological structure consistency constraint term is used as the overall optimization objective to iteratively optimize and solve the model parameters of the convolutional neural network.

14. The method according to claim 1, characterized in that, It also includes the following steps: The flood occurrence probability prediction results of each grid cell are organized into a spatial matrix according to spatial location and drawn into a flood probability distribution map to realize the visualization of the regional flood probability distribution. High-risk flood areas are identified based on preset probability thresholds; The system continuously outputs prediction results for multiple time steps to construct a flood evolution sequence, which is used to characterize the spatial expansion and propagation trend of floods, and realize flood event early warning and propagation prediction.

15. The method according to claim 1, characterized in that: The target area data includes tropospheric delay observation data and vertical displacement observation data from global navigation satellite system stations; the atmospheric precipitable water is obtained by inversion from the tropospheric delay observation data, and the terrestrial water storage is obtained by inversion from the vertical displacement observation data based on the elastic load theory.

Images