Deep learning-based historical flood inundation dynamic reconstruction method and system

Abstract

The application provides a deep learning-based historical flood inundation dynamic reconstruction method and system, and relates to the technical field of flood simulation. The method comprises the following steps: obtaining a standardized historical flood data set; locating P topographic cavity defects; matching and reusing P adaptive filling models in a topographic filling model library according to P missing feature vectors; performing multi-thread parallel processing of the P topographic cavity defects; performing pixel-level replacement of the P topographic cavity defects using P corrected topographic structures to obtain corrected DEM data; after constructing a two-dimensional water dynamic model based on the corrected DEM data, driving the two-dimensional water dynamic model to perform flood simulation by setting a water dynamic boundary control set, and outputting a historical flood dynamic evolution process. The application solves the technical problem that the flood simulation accuracy is limited due to incomplete topographic data and insufficient simulation generalization ability in different regions in the prior art, thereby affecting the accuracy of flood dynamic reconstruction, and improves the flood simulation accuracy.

Description

Technical Field

[0001] This application relates to the field of flood simulation technology, and in particular to a method and system for dynamic reconstruction of historical flood inundation based on deep learning. Background Technology

[0002] Currently, reconstructing historical flood dynamics often requires integrating multivariate flood data and multivariate topographic data. However, due to technical limitations during raw data acquisition, data corruption, or improper post-processing, the data often contains missing or incomplete information. Topographic data imputation is a crucial step in reconstructing historical flood inundation dynamics. Existing methods for imputing missing data generally lack sufficient generalization ability, making it difficult to adapt to the diverse needs of different geographical regions and complex terrain features. The incompleteness of topographic data and the insufficient generalization ability of imputation methods directly limit the accuracy of flood simulations, resulting in significant differences from actual conditions and thus affecting the accuracy of flood dynamic reconstruction.

[0003] In summary, existing technologies suffer from limitations in flood simulation accuracy due to incomplete terrain data and insufficient generalization ability across different regions, which in turn affects the accuracy of flood dynamic reconstruction. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for dynamic reconstruction of historical flood inundation based on deep learning, in order to solve the technical problem in the prior art that the accuracy of flood simulation is limited due to incomplete terrain data and insufficient simulation generalization ability in different regions, thereby affecting the accuracy of dynamic flood reconstruction.

[0005] In view of the above problems, this application provides a method and system for dynamic reconstruction of historical flood inundation based on deep learning.

[0006] Firstly, this application provides a deep learning-based method for dynamic reconstruction of historical flood inundation. This method is implemented through a deep learning-based system for dynamic reconstruction of historical flood inundation. The method includes: unifying and fusing historical multivariate flood data and topographic multivariate impact data to obtain a standardized historical flood dataset; locating P missing topographic cavities in the standardized historical flood dataset, each identified by a missing feature vector; matching and reusing P suitable infill models in a topographic infill model library based on the P missing feature vectors; driving the P suitable infill models to perform multi-threaded parallel processing of the P missing topographic cavities, outputting P corrected topographic structures; using the P corrected topographic structures in the standardized historical flood dataset to perform pixel-level replacement of the P missing topographic cavities, obtaining corrected DEM data; constructing a two-dimensional hydrodynamic model based on the corrected DEM data, and driving the two-dimensional hydrodynamic model to perform flood simulation by setting a hydrodynamic boundary control set, outputting the dynamic evolution process of historical floods.

[0007] Optionally, the confidence review engine is driven to traverse and perform three-dimensional confidence reviews on the historical multivariate flood data and topographic multivariate impact data; if the review passes, the historical multivariate flood data and topographic multivariate impact data are spatiotemporally aligned and then fused to obtain the standardized historical flood dataset.

[0008] Optionally, if the review fails, a multivariate defect dataset is separated from the historical multivariate flood data and the topographic multivariate impact data; the multivariate defect dataset is bidirectionally corrected to obtain a multivariate corrected dataset; after the multivariate corrected dataset is used to perform defect data fusion and replacement on the historical multivariate flood data and the topographic multivariate impact data, coordinate unification data fusion is performed to output the standardized historical flood dataset.

[0009] Optionally, the multivariate defect dataset is aggregated based on defect type to output an observation defect set, a mutation defect set, and an event defect set; a multi-channel defect correction model is pre-constructed, wherein the multi-channel defect correction model includes a parallel architecture of an observation correction channel, a mutation correction channel, and an event supplementary channel; after loading the observation defect set, the mutation defect set, and the event defect set into the multi-channel defect correction model, the observation correction channel, the mutation correction channel, and the event supplementary channel are driven to perform bidirectional defect correction in parallel, outputting an observation correction set, a mutation correction set, and an event supplementary set; by restoring the data structure of the observation correction set, the mutation correction set, and the event supplementary set, the multivariate correction dataset is obtained.

[0010] Optionally, step a: drive the observation correction channel to call the same source data of neighboring stations to perform spatiotemporal interpolation of the observation defect set to obtain the observation correction set; step b: drive the mutation correction channel to apply the double cumulative curve method to perform systematic bias correction on the mutation defect set to obtain the mutation correction set; step c: drive the event supplementation channel to extrapolate and supplement the event defect set based on the rainfall-runoff model to obtain the event supplementation set.

[0011] Optionally, connected component analysis is performed on the standardized historical flood dataset to locate multiple missing topographic cavities; P missing topographic cavities with ≥100 raster cells are selected from the multiple missing topographic cavities; a first range proportion, a first continuity indicator, a first edge gradient, and a first cavity type are extracted from the first missing topographic cavity; the first range proportion, the first continuity indicator, the first edge gradient, and the first cavity type are vectorized and concatenated to output a first missing feature vector; after constructing the P missing feature vectors by analogy, the P missing topographic cavities and the P missing feature vectors are associated through key-value pair mapping to generate a cavity location index matrix.

[0012] Optionally, a standard imputation model is invoked based on a predefined multi-scale range ratio threshold for the multi-scale missing vectors to obtain multiple sample imputation models. After locally invoking multiple complete DEM data samples, invalid regions of the multiple complete DEM data samples are manually constructed based on the multi-scale range ratio threshold for the multi-scale missing vectors to obtain multiple multi-scale sample terrain cavity sets. The multiple multi-scale sample terrain cavity sets and the multiple complete DEM data samples are used as training data, and a weighted loss function is used for training. After mapping and training the multiple sample imputation models, the terrain imputation model library is obtained by associating and storing the multi-scale missing vectors and the multiple sample imputation models. Vector matching is performed by traversing the terrain imputation model library based on the P missing feature vectors to reuse the P adapted imputation models.

[0013] Optionally, the discriminators of the multiple sample imputation models are all configured with the PatchGAN architecture, and the standard imputation models include the U-Net model and the PFRNet model.

[0014] Optionally, the topographic multivariate impact data is processed to unify coordinates to generate a river channel associated region; based on the topographic complexity, the river channel associated region is divided into a main channel sub-region and a floodplain sub-region; using the river channel associated region as a spatial alignment reference, the main channel sub-region is partitioned with a 5-10m high-resolution unstructured grid, and the floodplain sub-region is partitioned with a 10-50m medium-resolution unstructured grid to obtain an initial river channel model; after loading the river cross-section data into the initial river channel model, the corrected DEM data is mapped to the grid nodes of the initial river channel model using bilinear interpolation to generate the two-dimensional hydrodynamic model.

[0015] Secondly, this application also provides a deep learning-based historical flood inundation dynamic reconstruction system for executing the deep learning-based historical flood inundation dynamic reconstruction method described in the first aspect. The deep learning-based historical flood inundation dynamic reconstruction system includes: a coordinate unification and fusion module for performing coordinate unification and fusion on collected historical multivariate flood data and topographic multivariate impact data to obtain a standardized historical flood dataset; a missing location module for locating P missing topographic cavities in the standardized historical flood dataset, wherein the P missing topographic cavities are identified by P missing feature vectors; and a model matching module for matching data based on... The system uses P missing feature vectors to match and reuse P adapted infill models in a terrain filling model library; a multi-threaded parallel module drives the P adapted infill models to perform multi-threaded parallel processing of the P missing terrain cavities, outputting P corrected terrain structures; a pixel-level replacement module uses the P corrected terrain structures in the standardized historical flood dataset to perform pixel-level replacement of the P missing terrain cavities, obtaining corrected DEM data; and a flood simulation module constructs a two-dimensional hydrodynamic model based on the corrected DEM data, drives the two-dimensional hydrodynamic model to perform flood simulation by setting a hydrodynamic boundary control set, and outputs the dynamic evolution process of historical floods.

[0016] One or more technical solutions provided in this application have at least the following beneficial effects: by locating topographic cavities on a standardized historical flood dataset, identifying missing areas, matching suitable topographic filling models for multi-threaded parallel repair, the accuracy and adaptability of cavity repair are improved; the corrected topographic structure is accurately applied to the location of topographic cavities for pixel-level replacement; and a two-dimensional hydrodynamic model is constructed for flood simulation, improving the accuracy and generalization ability of flood simulation, thereby enhancing the accuracy of flood dynamic reconstruction. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the deep learning-based dynamic reconstruction method for historical flood inundation proposed in this application.

[0018] Figure 2This is a schematic diagram of the structure of the historical flood inundation dynamic reconstruction system based on deep learning in this application.

[0019] Figure labeling: Coordinate unification and fusion module 11, Missing location module 12, Model matching module 13, Multi-threaded parallel module 14, Pixel-level replacement module 15, Flood simulation module 16. Detailed Implementation

[0020] This application provides a deep learning-based method and system for dynamic reconstruction of historical flood inundation, addressing the technical problem in existing technologies where incomplete terrain data and insufficient generalization ability across different regions limit the accuracy of flood simulation, thus affecting the accuracy of dynamic flood reconstruction. By locating terrain voids on a standardized historical flood dataset, identifying missing areas, and matching suitable terrain filling models for multi-threaded parallel repair, the accuracy and adaptability of void repair are improved. The corrected terrain structure is precisely applied to the locations of terrain voids for pixel-level replacement, and a two-dimensional hydrodynamic model is constructed for flood simulation, improving the accuracy and generalization ability of flood simulation and thus enhancing the accuracy of dynamic flood reconstruction.

[0021] Example 1, please refer to the appendix. Figure 1 This application provides a deep learning-based method for dynamic reconstruction of historical flood inundation, which specifically includes the following steps:

[0022] By unifying and fusing the collected historical multivariate flood data and topographic multivariate impact data with coordinates, a standardized historical flood dataset is obtained.

[0023] Furthermore, this application also includes the following steps: driving the confidence review engine to traverse and perform three-dimensional confidence reviews on the historical multivariate flood data and topographic multivariate impact data; if the review passes, then after spatiotemporally aligning the historical multivariate flood data and topographic multivariate impact data, performing coordinate unification fusion to obtain the standardized historical flood dataset.

[0024] Specifically, historical multivariate flood data collection includes hydrological monitoring records, flood surveys, and flood event-related data. Hydrological monitoring records include water level and flow data, recording the flood's occurrence time, peak water level, duration, peak flow, and hydrological monitoring records for the affected area. Flood surveys and flood event-related data typically estimate peak water levels, flow rates, and inundation extent by identifying water level marks left by floodwaters on structures, rocks, bridges, etc. Flood inundation extent is obtained through satellite or aerial imagery. The intensity of rainfall, cumulative rainfall, and spatiotemporal distribution of rainfall corresponding to each flood event also require appropriate hydrological analysis. During data collection, it is crucial to ensure data completeness and accuracy; data from different sources and in different formats must undergo standardized preprocessing.

[0025] Multi-dimensional topographic impact data were collected, including DEM data, DOM data, oblique photogrammetry data of key river sections, and surface topographic data of key river sections acquired using airborne lidar, while underwater topographic data of key river sections was acquired using RTK cross-sectional data. DEM data, as the fundamental data reflecting the topographic elevation of the study area, is the core of constructing the topographic framework for flood simulation. DOM data provides clear surface texture information, which can help identify surface feature types and provide a reference for analyzing the impact of the underlying surface on flood evolution. Oblique photogrammetry data and surface topographic data of key river sections can further improve the accuracy of DEM data. Cross-sectional measurement data of the river channel accurately reflects the morphological characteristics of the river channel, including riverbed elevation, channel width, and slope gradient, which is crucial for accurately simulating flood flow and floodplain processes within the river channel.

[0026] Oblique photogrammetry data acquisition for key river sections is primarily used to obtain 3D texture information and spatial morphology of features along the riverbanks, accurately presenting detailed characteristics of above-ground elements such as riverbank buildings, vegetation cover, and levee engineering. The acquisition process is typically carried out by drones or manned aircraft equipped with multi-lens cameras. The target area is photographed from multiple angles (usually including vertical and four oblique directions), and then post-processed to generate high-resolution 3D models that combine geometric accuracy and texture realism. Water topographic data for key river sections is acquired using airborne LiDAR. An airborne laser scanner emits laser pulses to the ground surface, and the distance is calculated using the pulse echo time difference, thus quickly obtaining the 3D coordinate information of the water topography. Its advantages include being unaffected by weather and lighting conditions, efficiently covering large areas, and providing high data accuracy (horizontal and elevation accuracy up to decimeter level). It is particularly suitable for topographic mapping of complex river morphologies, such as shoals, sandbars, and floodplains, accurately reflecting the undulating characteristics of the water topography above the riverbed, providing crucial data support for river evolution analysis and flood control capacity calculation.

[0027] Underwater topographic data for key river sections was acquired using RTK cross-sectional data collection. This involved mounting an RTK-GPS receiver and a depth sounder on a survey vessel or platform, and then moving along a pre-defined cross-section for measurement. RTK technology provides real-time, centimeter-level accuracy in both horizontal and vertical positioning. Combined with the simultaneous measurement of underwater depth by the depth sounder, the three-dimensional coordinates of each point on the cross-section line can be directly obtained. After data stitching, the underwater topographic cross-sectional data is formed, suitable for detailed measurement of river underwater topography.

[0028] The confidence review engine traverses and performs three-fold confidence reviews on the acquired historical multivariate flood data and topographic multivariate impact data. The confidence review engine is a software system used to check and verify data quality. Through certain rules and standards, it analyzes the reliability, consistency, and applicability of data, ensuring that the data meets requirements and can be used for further analysis and processing. The three-fold confidence review is a core part of hydrological data compilation and quality control, aiming to ensure the reliability, consistency, and applicability of data, including systematic verification and analysis of the reliability, consistency, and representativeness of hydrological data. Reliability review is the core part of data quality control, aiming to verify the authenticity and accuracy of data sources and investigate potential errors in the observation, recording, and transmission processes. It mainly includes key aspects such as observation method review, human error review, and data source credibility review. The observation method review focuses on verifying whether the data acquisition methods (such as manual observation, automatic sensor monitoring, etc.) comply with industry standards, and confirming whether the observation instruments are calibrated regularly as required, to ensure the scientific and standardized nature of data acquisition. Record integrity and outlier review involves a comprehensive examination of the data sequence to check for missing data and outliers (such as negative flow rates or water levels exceeding historical extremes) to ensure data integrity and reasonableness. Human error review focuses on identifying errors during data transcription (such as misspelled numbers or misplaced decimal points) and unit confusion (e.g., miswriting flow units as volume units), minimizing the impact of human factors on data quality. Data source credibility review requires comparing the data with historical flood survey results to verify consistency with flood records, water conservancy chronicles, and other relevant data, thereby assessing the reliability of the data source.

[0029] Logical checks, such as verifying whether the flow-water-level relationship at a certain station conforms to the station's established flow-water-level curve; comparative verification, such as performing correlation analysis between the data and concurrent data from neighboring stations or meteorological data; error tracing, which involves tracing and investigating the sources of data errors by reviewing original record books, instrument operation logs, and other materials. The core objective of consistency review is to ensure that the data is comparable in time or space, thereby eliminating data abrupt changes caused by non-natural factors (such as station relocation, instrument replacement, etc.). This includes key aspects such as review of the impact of station changes, review of changes in observation methods, and review of climate and underlying surface consistency. The review of the impact of station changes focuses on changes in station location and elevation, as well as whether river regulation projects (such as dam construction, river straightening, etc.) have disrupted the continuity of the data sequence, leading to unreasonable breaks in the data. Review of changes in observation methods: When the observation method changes (e.g., from manual observation to automatic monitoring), it assesses whether this change will introduce systematic bias and affect data consistency. Climate and underlying surface consistency review: Analyze whether changes in watershed land use (such as urbanization, deforestation, etc.) have altered the watershed's runoff generation and distribution relationships, thereby affecting data consistency.

[0030] In the practice of consistency review, the relationship between cumulative precipitation and cumulative runoff is compared using the double cumulative curve method. If the slope of the curve changes abruptly, it means that the underlying surface has changed or there is inconsistency in the data. Time series analysis is used to check whether there are abrupt changes in the mean and variance of the data. Historical comparison at the same station is used to analyze whether the water level and flow relationship curves of the same station over many years have shifted, thereby judging the consistency of the data.

[0031] The core purpose of representativeness review is to assess the extent to which data reflects the hydrological characteristics of the study area and to determine whether it can effectively represent the long-term hydrological patterns and extreme hydrological conditions of the region. Temporal coverage review needs to focus on whether the data series is sufficiently long, such as whether it includes at least different hydrological years, including high-water years, normal-water years, and low-water years. If the data series is too short, extreme hydrological events may be missed, thus affecting the understanding of long-term hydrological patterns. Spatial representativeness review examines whether the distribution of observation stations can cover key areas of the basin, such as the upstream headwaters and the confluence areas of major tributaries, ensuring that the data does not overgeneralize due to incomplete spatial coverage and accurately reflects the hydrological conditions of the entire basin. The extreme event inclusion review focuses on whether the data series includes historical major floods, exceptionally dry years, and other extreme hydrological events. These extreme events are crucial for hydrological frequency analysis and directly affect the reliability of the analysis results.

[0032] Frequency analysis verifies whether the data series meets the hydrological statistical assumptions and assesses statistical representativeness; comparison with climate conditions verifies whether the hydrological data matches the changing trends of climate elements such as precipitation and temperature during the same period, enhancing the rationality of the data; interpolation extension supplements the data when there are missing data by using observation data from nearby stations or tools such as rainfall and runoff models to improve the completeness and representativeness of the data.

[0033] Once the three-dimensional confidence review is passed, spatiotemporal alignment is performed on historical multivariate flood data and topographic multivariate impact data. Spatiotemporal alignment aims to resolve inconsistencies in the time and spatial dimensions of the data and ensure comparability. As a fundamental step in data fusion, spatiotemporal alignment mainly consists of time alignment and spatial alignment. Time alignment includes two core dimensions: unified time resolution and synchronization of time references. Unified time resolution aims to address the problem of inconsistent data time scales. For data with different time interval characteristics, such as hourly observation data and observation data added at irregular intervals, linear interpolation or cubic spline interpolation and aggregation operations are required to standardize them to a unified time granularity, ensuring data comparability in the time dimension. Synchronization of time references focuses on eliminating differences in time measurement systems. Specifically, it is necessary to complete the unified processing of time zones, standardize the definition of time formats, and fill in the gaps in the data series by supplementing with source data from neighboring stations or by professional model extrapolation, ensuring the integrity and continuity of the time series.

[0034] The core purpose of spatial alignment is to achieve accurate matching and integrated analysis of multi-source data by unifying the coordinate systems, geometric benchmarks, or spatial resolutions of different data sources. As another important dimension of spatiotemporal alignment, spatial alignment mainly encompasses two key aspects: spatial resolution matching and spatial coverage adjustment. Spatial resolution matching aims to solve the problem of inconsistent spatial scales in data. For raster data (satellite or aerial imagery), resampling techniques, including commonly used algorithms such as bilinear interpolation and nearest neighbor methods, are needed to transform it to a uniform spatial resolution, ensuring that the data has a consistent level of detail in the spatial dimension. Spatial coverage adjustment focuses on achieving consistency in the spatial extent of the data. Through cropping or masking, the effective spatial extent of different data sources is strictly limited to the target study area; for discretely distributed station observation data, Kriging interpolation is used to convert it into continuous spatial field data, thereby eliminating spatial matching barriers caused by differences in data presentation formats.

[0035] After aligning historical multivariate flood data and topographic multivariate impact data in time and space, coordinate unification and fusion are performed. Coordinate unification aims to eliminate spatial reference differences and achieve consistency in geographic coordinates. It is a crucial step in spatial data integration, primarily encompassing three aspects: projection transformation, datum correction, and elevation unification. Projection transformation aims to eliminate differences caused by different coordinate systems. For data using different coordinate systems, specialized tools such as ArcGIS are used to convert them to a unified projection through specific transformation methods, thus avoiding distortions in distance and area calculations. Datum correction mainly addresses the offset problem between different geodetic datums. When data is based on different geodetic datums, seven-parameter or three-parameter transformation methods are used for correction to ensure consistency across geodetic datums. Elevation unification ensures that elevation data is on the same datum. For data such as water level elevations, they need to be unified to the same elevation datum, correcting any potential vertical deviations to guarantee the accuracy and comparability of the elevation data.

[0036] After undergoing confidence verification, spatiotemporal alignment, and coordinate unification fusion, all data (historical multivariate flood data and topographic multivariate impact data) will be unified within the same temporal and spatial framework. All data has been verified and processed to ensure its reliability, consistency, and applicability, resulting in a standardized historical flood dataset. This standardized historical flood dataset is a consistent and reliable historical flood dataset that has undergone review and processing. The data in the dataset has been coordinate unified and spatiotemporally aligned, and meets the requirements of flood simulation.

[0037] Through rigorous three-dimensional review, a large amount of low-quality and unreliable data was filtered out, preventing erroneous data from interfering with subsequent simulation results and improving the reliability of the simulation. Spatiotemporal alignment and coordinate unification fusion solved the problem of directly using multi-source heterogeneous data, integrating the data into a unified and standardized framework, enabling different types of data to be matched and analyzed collaboratively. The resulting standardized historical flood dataset exhibits significantly improved completeness, accuracy, consistency, and spatiotemporal matching, directly reducing uncertainties in the modeling process. This allows the hydrodynamic model constructed from the standardized historical flood dataset to accurately reflect the real physical processes of historical floods, thereby improving the accuracy and reliability of the dynamic reconstruction of historical floods.

[0038] Furthermore, this application also includes the following steps: if the review fails, a multivariate defect dataset is separated from the historical multivariate flood data and the topographic multivariate impact data; the multivariate defect dataset is bidirectionally corrected to obtain a multivariate corrected dataset; after the multivariate corrected dataset is used to perform defect data fusion and replacement on the historical multivariate flood data and the topographic multivariate impact data, coordinate unification data fusion is performed to output the standardized historical flood dataset.

[0039] Furthermore, this application also includes the following steps: aggregating the multivariate defect dataset based on defect type, and outputting an observation defect set, a mutation defect set, and an event defect set; pre-constructing a multi-channel defect correction model, wherein the multi-channel defect correction model includes a parallel architecture of an observation correction channel, a mutation correction channel, and an event supplementary channel; loading the observation defect set, the mutation defect set, and the event defect set into the multi-channel defect correction model, and driving the observation correction channel, the mutation correction channel, and the event supplementary channel to perform bidirectional defect correction in parallel, outputting an observation correction set, a mutation correction set, and an event supplementary set; obtaining the multivariate correction dataset by restoring the data structure of the observation correction set, the mutation correction set, and the event supplementary set.

[0040] Furthermore, this application also includes the following steps: Step a: Driving the observation correction channel to call the same source data from neighboring stations to perform spatiotemporal interpolation of the observation defect set, thereby obtaining the observation correction set; Step b: Driving the mutation correction channel to apply the double cumulative curve method to perform systematic bias correction on the mutation defect set, thereby obtaining the mutation correction set; Step c: Driving the event supplementation channel to extrapolate and supplement the event defect set based on the rainfall-runoff model, thereby obtaining the event supplementation set.

[0041] Specifically, when the confidence review engine determines that the input historical multivariate flood data and topographic multivariate impact data fail the three-dimensional confidence review, it does not directly discard this data. Instead, it initiates a data repair process. This involves separating the multivariate defect dataset from the historical multivariate flood data and topographic multivariate impact data; that is, identifying and separating all data points or data blocks that do not meet the quality standards and classifying them into the multivariate defect dataset. The multivariate defect dataset undergoes bidirectional correction, meaning the correction process is carried out in both directions.

[0042] A multivariate defect dataset refers to the defective portion of historical flood and topographic data. Based on defect type, the multivariate defect dataset is aggregated by defect type, i.e., the defective data in the dataset is classified and organized. According to different defect types, the multivariate defect dataset is divided into observation defect set, mutation defect set, and event defect set. The observation defect set contains defective data caused by observation equipment failure, transmission errors, or improper data processing, such as abnormal water level or flow readings. The mutation defect set contains data mutations or deviations caused by sudden events (such as equipment replacement or changes in data processing methods). The event defect set contains data missing or abnormal data caused by specific events (such as extreme weather or human intervention).

[0043] A multi-channel defect correction model is constructed, which is a deep learning model containing multiple parallel correction channels to correct different types of defective data. Each channel specializes in handling a specific type of defective data, including observation defect correction channels, abrupt change defect correction channels, and event defect supplementation channels. The observation correction channel corrects data problems caused by observational defect sets by performing spatiotemporal interpolation using data from neighboring stations to fill in missing or inconsistent observational data. The abrupt change correction channel handles abrupt change defective data, applying the double cumulative curve method to correct systematic biases in the data and ensure the smoothness of data trends. The event supplementation channel handles defective data caused by sudden events. It uses a rainfall-runoff model to extrapolate and supplement the event defect set, filling in data gaps caused by precipitation or flood events.

[0044] In a multi-channel correction model, the observation correction channel, mutation correction channel, and event supplement channel work simultaneously in parallel, correcting different types of defective data respectively. Each channel corrects data within its specific scope and outputs a corrected dataset. After each channel completes its correction task, the three corrected datasets (observation correction set, mutation correction set, and event supplement set) are combined. Parallel architecture refers to multiple modules or channels working in parallel, independently processing different data streams. Each channel is responsible for processing different features or defects in the data and works independently, ultimately merging the corrected data. The observation correction channel, mutation correction channel, and event supplement channel work in parallel, processing different types of defective data respectively.

[0045] After loading the observation defect set, mutation defect set, and event defect set into the multi-channel defect correction model, spatiotemporal interpolation of the observation defect set is performed by calling nearby stations with similar source data through the observation correction channel. For example, precipitation data for a certain observation station is missing for a specific period. To fill in the missing data, data from spatially proximate stations with the same source are selected. The data from these nearby stations should be similar to the missing data and can serve as a good reference. Using spatiotemporal interpolation methods, data from adjacent time periods (if it is time series data) or adjacent spatial locations (if it is geographic data) are selected. For example, assuming that precipitation data for station A in March 2005 is missing, while precipitation data from stations B and C are available, the precipitation for station A is estimated using interpolation methods based on the precipitation data from stations B and C. Through spatiotemporal interpolation methods, the missing data in the observation defect set is repaired, resulting in the observation correction set, which has filled in the originally missing observation data. For example, if the rainfall at station B is 50 mm and the rainfall at station C is 45 mm, the rainfall at station A can be obtained by weighted averaging. The rainfall at station A is (50 mm + 45 mm) / 2 = 47.5 mm.

[0046] In the mutation correction channel, mutation defects caused by equipment failure, data fluctuations, etc., are addressed. Mutation defects typically manifest as abnormal data changes, such as a sudden increase or decrease in water level data within a certain period, deviating from the normal trend. The double cumulative curve method detects data mutations by constructing two curves: the first cumulative curve calculates the cumulative sum of the original data, reflecting the overall trend; the second cumulative curve smooths the original data, reducing short-term fluctuations and yielding a smoothed cumulative sum. By comparing the two cumulative curves, the deviation between the original and smoothed data is checked, identifying the mutation. Once a mutation is detected, the double cumulative curve method corrects the mutated data according to the overall trend. After correction using the double cumulative curve method, a mutation correction set is obtained, where the abnormal fluctuations have been smoothed. For example, suppose the water level at station B suddenly increases from the normal 10 meters to 50 meters. The smoothed trend should be a gradual increase from 10 meters to 12 meters, but the sudden increase of 50 meters clearly deviates from this trend. Using the double cumulative curve method, the 50-meter increase is corrected to 12.5 meters, conforming to the overall trend.

[0047] In the event supplementation channel, missing data caused by event deficiencies (such as rainstorms or floods) is supplemented based on rainfall-runoff models. For example, in a rainstorm event, some flow data may be missing due to monitoring equipment failure or transmission interruption. The missing data is deduced through the event supplementation channel. Rainfall-runoff models typically calculate runoff volume based on known precipitation data, simulating how precipitation affects surface runoff and, consequently, flood events. Missing flow data is predicted and supplemented using rainfall-runoff models based on historical precipitation, flow, and other known data. Through this deduction and supplementation based on rainfall-runoff models, an event supplement set is obtained, filling in the missing flow data during rainstorm or flood events. For example, assuming flow data is missing at station C, and a rainstorm occurred in the area with 100 mm of precipitation, the corresponding flow calculated using the rainfall-runoff model should be 500 cubic meters per second.

[0048] After completing the above three correction processes, we obtain the observation correction set, the mutation correction set, and the event supplement set, respectively. These sets are then restored and combined into a complete multivariate corrected dataset. The structure of the corrected dataset is restored to ensure that each type of defective data is appropriately filled and that data types remain consistent. The observation correction set, mutation correction set, and event supplement set are then integrated to form a complete multivariate corrected dataset. The defects in this dataset have been corrected, resulting in high quality and applicability. Through techniques such as bidirectional correction, spatiotemporal interpolation, the double cumulative curve method, and rainfall-runoff models, defects in historical flood data, especially those related to missing data, anomalies, and sudden events, can be effectively repaired.

[0049] This method involves fusing and replacing defective data in historical multivariate flood data and topographic multivariate impact data using a multivariate corrected dataset. The corrected data replaces inaccurate or missing data in the original dataset. The corrected data is then combined with the original data, replacing incomplete or inaccurate portions. With the support of data fusion algorithms, the corrected data is appropriately embedded into the original data structure to ensure seamless integration. For example, if precipitation data for a particular station is missing, the original missing value will be replaced with precipitation data obtained through forward correction.

[0050] After data fusion and replacement, all historical flood and topographic data have been repaired and merged. Next, coordinate unification data fusion is required: unifying the historical flood and topographic data into a standard coordinate system to ensure accurate overlay of data from different sources and periods. After coordinate unification, data from different sources (such as historical precipitation data, flow data, digital elevation models, etc.) are merged in the same coordinate system to ensure correct data alignment. After bidirectional correction, defective data fusion and replacement, and coordinate unification data fusion, a standardized historical flood dataset is obtained. Through bidirectional correction and defective data replacement, missing, anomaly, and inconsistency issues in the original data are resolved, ensuring data reliability and consistency.

[0051] Locate P missing topographic cavities in the standardized historical flood dataset, wherein each of the P missing topographic cavities is identified by a missing feature vector.

[0052] Furthermore, this application also includes the following steps: performing connected component analysis on the standardized historical flood dataset to locate multiple missing topographic cavities; selecting P missing topographic cavities with ≥100 raster cells from the multiple missing topographic cavities; extracting the first range proportion, first continuity indicator, first edge gradient, and first cavity type of the first missing topographic cavity; vectorizing and concatenating the first range proportion, first continuity indicator, first edge gradient, and first cavity type to output a first missing feature vector; and then, after constructing the P missing feature vectors by analogy, associating the P missing topographic cavities and the P missing feature vectors through key-value pair mapping to generate a cavity location index matrix.

[0053] Specifically, connectivity analysis is performed on standardized historical flood datasets. Connectivity analysis is an image processing method typically used to identify and analyze continuous regions in data, specifically identifying which areas are continuous and may contain holes or missing data. By analyzing the connectivity between different values ​​in raster data, holes or missing regions in topographic data can be located. All possible hole regions are identified by analyzing the connectivity between adjacent cells in raster data. Connectivity analysis can be performed based on the adjacency relationships between pixel values. For example, in a Digital Elevation Model (DEM), if some raster cells have missing values ​​(such as NaN or invalid values), these adjacent invalid raster cells will be marked as a hole region. Suppose the DEM data for a certain area is a 10x10 grid, and some areas have missing data due to sensor malfunction; these missing raster cells will be identified as topographic holes.

[0054] Raster cells are the fundamental building blocks of grid data, with each cell representing a spatial region. In terrain data, a raster cell typically represents the elevation value of a ground location (e.g., a 5x5 meter or 10x10 meter area). During connected component analysis, multiple terrain voids were detected. From these voids, voids with an area of ​​100 or more raster cells were selected. These larger voids are generally more significant and may have a greater impact on flood simulations, thus requiring priority processing. The area of ​​each void region was calculated, which can be obtained by counting the number of raster cells within the void. If a void has an area of ​​100 or more raster cells, it is marked as a void requiring further processing. For example, if void A contains 150 raster cells and void B contains 80 raster cells, void A will be selected as the void requiring priority processing, while void B will be ignored.

[0055] Through filtering, P terrain holes are identified, where P is a positive integer greater than or equal to 1. For the selected P terrain holes (i.e., holes with an area ≥ 100 raster cells), the extent ratio, continuity marker, edge gradient, and hole type are extracted. The first extent ratio is the proportion of the area of ​​the first terrain hole to the total area of ​​all P terrain holes. For example, if hole A has an area of ​​150 raster cells, and the total area of ​​all P terrain holes is 1000 raster cells, then the extent ratio of hole A is 15%.

[0056] The first continuity criterion determines whether the first topographic cavity is isolated or connected to other cavities. For example, if cavity A and cavity B are connected by a small area, they might be considered a larger cavity. If the cavity is continuous, the criterion is 1; otherwise, it's 0. The first edge gradient assesses the transition between the first topographic cavity and the surrounding valid area. If the boundary of the first topographic cavity is too abrupt (e.g., completely discontinuous), the edge gradient is large; if the transition is smooth, the gradient is small. The first cavity type is classified based on the shape, size, and distribution of the first topographic cavity, such as circular, rectangular, or irregular shapes. For example, cavity A has an area of ​​150 grid cells, a range ratio of 15%, a small edge gradient, and is classified as an irregular shape; its continuity criterion is isolation.

[0057] The first range proportion, first continuity indicator, first edge gradient, and first cavity type extracted from the first terrain cavity are numerically processed. The first range proportion is directly represented by a floating value, such as 0.15; the first continuity indicator is represented by a binary number, where 1 indicates an isolated cavity and 0 indicates a connected cavity; the first edge gradient is a numerical representation of the rate of change of the edge, which can be a floating value, usually obtained through gradient calculation or variance, such as an edge gradient of 0.2; the first cavity type can be encoded by numbers, such as 1 for a circle, 2 for a rectangle, and 3 for an irregular shape. Assume the cavity type is 3. The first range proportion, first continuity indicator, first edge gradient, and first cavity type are vectorized and concatenated to obtain the missing feature vector. For example, with a range proportion of 0.15, a continuity indicator of 1, an edge gradient of 0.2, and a cavity type of 3, the concatenated missing feature vector is: 0.15, 1, 0.2, 3.

[0058] The same steps are performed for each of the P missing terrain holes to obtain P missing feature vectors. For example, consider the following holes: Hole A: area of ​​150 grid cells, range ratio of 0.15, continuity flag of 1 (isolated), edge gradient of 0.2, hole type of 3 (irregular shape); Hole B: area of ​​120 grid cells, range ratio of 0.12, continuity flag of 0 (connected), edge gradient of 0.3, hole type of 2 (rectangular); Hole C: area of ​​180 grid cells, range ratio of 0.18, continuity flag of 1 (isolated), edge gradient of 0.1, hole type of 1 (circular). The missing feature vector for each hole will be: Hole A: [0.15, 1, 0.2, 3], Hole B: [0.12, 0, 0.3, 2], Hole C: [0.18, 1, 0.1, 1].

[0059] By using key-value pair mapping, P missing terrain holes are associated with P missing feature vectors to obtain a hole location index matrix. The hole location (e.g., raster coordinates) serves as the key. For example, hole A is located at (10, 20), hole B at (15, 25), and hole C at (30, 35). The value corresponds to the missing feature vector of the hole. For example, the feature vector of hole A is [0.15, 1, 0.2, 3]. Through key-value pair mapping, the spatial location of each hole is associated with its corresponding feature vector. The hole location index matrix not only records the location of the hole but also the associated feature information.

[0060] By using connected component analysis, void areas in terrain data can be quickly identified, especially larger voids that require focused repair. Key features of voids (such as extent ratio and continuity indicators) are extracted and vectorized, allowing the spatial characteristics of voids to be expressed in numerical form. The generated void location index matrix can provide accurate positioning and feature support for the subsequent void filling process, ensuring the effectiveness and accuracy of the filling model.

[0061] Based on the P missing feature vectors, P suitable filling models are matched and reused in the terrain filling model library.

[0062] Furthermore, this application also includes the following steps: calling a standard imputation model based on a predefined multi-scale range ratio threshold for multi-scale missing vectors to obtain multiple sample imputation models; after locally calling multiple complete DEM data samples, manually constructing invalid regions of the multiple complete DEM data samples based on the multi-scale range ratio threshold for the multi-scale missing vectors to obtain multiple multi-scale sample terrain cavity sets; using the multiple multi-scale sample terrain cavity sets and the multiple complete DEM data samples as training data, training with a weighted loss function, mapping and training the multiple sample imputation models, and obtaining the terrain imputation model library by associating and storing the multi-scale missing vectors and the multiple sample imputation models; traversing the terrain imputation model library based on the P missing feature vectors to perform vector matching, so as to reuse the P adapted imputation models.

[0063] Furthermore, this application also includes the following steps: the discriminators of the plurality of sample imputation models are all configured with the PatchGAN architecture, and the standard imputation models include the U-Net model and the PFRNet model.

[0064] Specifically, the core of building a terrain-filling model based on adversarial networks (ANNs) is to have two networks compete against each other and improve together. The generator is responsible for generating fake data that is as realistic as possible, attempting to deceive the discriminator. The discriminator is responsible for distinguishing between real data and the fake data generated by the generator, aiming to accurately identify the difference. During training, the two networks continuously compete, eventually reaching a Nash equilibrium, where the data generated by the generator is sufficiently realistic that the discriminator cannot distinguish between real and fake data; that is, the discriminator guesses randomly with a 50% probability. Ultimately, this enables the generator to generate realistic data. The ANN consists of a generator and a discriminator: The generator takes a random noise vector and DEM data as input, extracts terrain features through an encoder-decoder convolutional layer, and outputs fake sample DEM data; the discriminator takes the fake sample DEM data and the real DEM data from the training sample set as input, and uses convolutional layers to determine the authenticity of local terrain. The generator and discriminator are optimized through adversarial training, and the loss function combines adversarial loss, terrain reconstruction loss, and terrain structure gradient loss. Through adversarial training between the generator and the discriminator, existing terrain data can be used to intelligently fill in and correct missing or inaccurate terrain parts in P missing feature vectors, generating elevation data that better matches the actual terrain, improving the completeness and accuracy of the terrain data, and providing a high-quality terrain foundation for the subsequent construction of hydrodynamic models.

[0065] PatchGAN is a Generative Adversarial Network (GAN) architecture widely used in image generation and inpainting. Instead of generating the entire image, PatchGAN generates small patches within it. The judgment result of each patch is used to evaluate the overall image quality. In image inpainting, PatchGAN improves the accuracy of the inpainting through local discrimination, resulting in more detailed and realistic results. For sample inpainting tasks, the discriminator judges whether the generated inpainted data matches the real terrain data. A key feature of the PatchGAN architecture is that it divides the image into multiple patches and judges the realism of each patch, rather than judging the entire image. Through PatchGAN, each local region of the inpainted data (e.g., a small patch of terrain data) is judged by the discriminator, ensuring that the inpainting of each patch is reasonable.

[0066] Standard inpainting models include U-Net and PFRNet. U-Net is a deep learning architecture for image inpainting, segmentation, and reconstruction, using an encoder-decoder structure to restore image details. In terrain data inpainting, U-Net is used to repair holes. The encoder part of U-Net extracts features from the input data, compressing the spatial dimension of the input data layer by layer while preserving deep information. The decoder part upsamples the features extracted by the encoder, restoring the original spatial dimensions. In this process, U-Net uses skip connections to combine features from the encoder and decoder, thus better restoring image details. PFRNet is a network that improves image quality by progressively adjusting image features. Its key feature is progressively recalibrating the image's feature maps, thereby improving the detail quality of the generated image. By progressively improving image features, PFRNet can finely repair missing parts of an image. For example, when repairing holes, PFRNet can progressively adjust image features to avoid obvious differences between the repaired area and the surrounding area. PFRNet not only learns features but also improves image repair performance through recalibration of feature maps. By adjusting the features, PFRNet is able to better preserve the natural transitions and details of the image.

[0067] Based on the P missing feature vectors, if the problem is identified as a small-scale local area requiring correction, U-Net is selected as the generator structure in the adversarial network; if the problem is identified as a large-scale continuous area requiring correction, PFRNet is selected. PatchGAN is used as the discriminator for local authenticity determination. The processing flow with U-Net as the backbone is as follows: the input DEM data is downsampled by the encoder convolution to extract multi-scale features; the encoded features of each scale are passed to the corresponding decoding layer through skip connections; in the decoding stage, the received encoded features and the upper-layer decoding output are upsampled and fused, and the fused features are input into the U-Net module at the key layer; finally, after processing and upsampling by the decoding layer containing U-Net, the high-resolution DEM forgery sample is output by the convolutional layer.

[0068] The RFRNet module iteratively optimizes feature representation and refines terrain structure information through a recurrent inference mechanism. Its core lies in using recurrent units to establish long-range dependencies between spatial locations within the feature map, refining the feature map F in the t-th iteration. t The update process can be represented as: h t ,F t+1 =RFRCell(F t ,h t-1 ,C), where h t It is the hidden state of the current iteration, h t-1It represents the hidden state of the previous iteration, C is optional context information, RFRCell represents the computation process of the loop unit, and F... t+1 This is the feature map from the previous iteration; after T iterations, the optimized feature map F is output. t .

[0069] A predefined multi-scale missing vector is a vector containing missing feature descriptions at multiple scales, used to define terrain holes of different sizes. Each scale corresponds to a range proportion threshold used to determine the size of the hole. The multi-scale range proportion threshold is used to distinguish terrain holes of different sizes, determining the scale at which repair should be performed based on the size and extent of the missing region, thereby optimizing the infilling strategy. Based on the multi-scale range proportion thresholds of the predefined multi-scale missing vector, multiple sample infilling models are obtained by calling them in the standard infilling model.

[0070] After accessing multiple complete DEM data samples locally, invalid regions are manually constructed based on multi-scale range ratio thresholds, generating multi-scale sample terrain cavity sets. The size of the missing regions is selected according to threshold rules. Regions larger than the threshold (e.g., area greater than 100 raster cells) are considered large-scale cavity regions, while regions smaller than the threshold (e.g., area less than 50 raster cells) are considered small-scale cavity regions. Based on existing complete DEM data, missing or invalid regions are manually inserted to simulate missing parts in actual data. Multiple cavity sets are constructed using invalid regions at different scales, each cavity set containing missing region samples at a specific scale.

[0071] Multiple multi-scale sample sets of terrain holes and multiple complete DEM data samples are used as training data. Each training sample includes a DEM data with holes and the target repair portion (i.e., the filling region). For each sample in training, the filling error of the hole region is adjusted using a weighted loss function. Larger hole regions may be assigned higher weights, focusing on filling large holes. Through iterative optimization, multiple filling models are trained using the weighted loss function, allowing each model to be optimized according to features such as the size and shape of the hole region, achieving higher repair accuracy. A terrain filling model library is generated from the trained multiple filling models, containing multiple models optimized for different missing region features (such as size, edge gradient, etc.). During the training process using the weighted loss function, multiple filling models for different hole features are generated, and each trained filling model is stored in the terrain filling model library according to its applicable hole features.

[0072] Using P missing feature vectors, the terrain filling model library is traversed, and the most suitable filling model is selected based on a vector matching mechanism. The features of each hole are represented by a multi-dimensional vector (such as extent ratio, edge gradient, etc.), and each missing feature vector is matched with a model in the terrain filling model library. Based on the similarity of the feature vectors, the most suitable filling model for the current hole is selected. Once a matching model is found, it can be used to fill the hole area.

[0073] The P adaptive filling models are driven to perform multi-threaded parallel processing of the P terrain hole missing parts, and P corrected terrain structures are output.

[0074] In the standardized historical flood dataset, the pixel-level replacement of the P missing terrain holes is performed using the P corrected terrain structures to obtain corrected DEM data.

[0075] Specifically, P adaptive infilling models are driven to process P missing terrain holes in parallel using multi-threaded methods. The repair task for each hole is assigned to different threads for parallel execution, with each thread responsible for repairing one hole independently. Through multi-threaded parallel execution, all holes are repaired, and corresponding corrected terrain structures are output. Each repaired terrain region contains complete data; elevation information is filled, and the hole area no longer exists. The P corrected terrain structures are the terrain data structures repaired by the adaptive infilling models; the hole areas are filled, the data no longer contains holes, and the terrain structure is complete. In the standardized historical flood dataset, P corrected terrain structures are used to perform pixel-level replacement of the P missing terrain holes, that is, using the corrected terrain structures to replace the missing areas in the terrain data. Each corrected terrain structure represents a filled region, and the infilling model has generated appropriate repair data based on the characteristics of the hole area. After repair, all hole areas have been replaced with valid terrain data, thus obtaining corrected DEM data. Corrected digital elevation model (DEM) data refers to terrain data that has been repaired and filled in, eliminating missing or voided areas. Pixel-level replacement refers to operating on each pixel individually in an image or raster data, meaning that for each missing raster cell (i.e., each pixel), the missing value is filled in using corrected terrain data. By replacing pixels one by one, missing data in voided areas is filled in, ensuring that the terrain information in the entire DEM dataset is free of missing or voided areas.

[0076] After constructing a two-dimensional hydrodynamic model based on the corrected DEM data, the two-dimensional hydrodynamic model is driven to perform flood simulation by setting a hydrodynamic boundary control set, and the historical flood dynamic evolution process is output.

[0077] Furthermore, this application also includes the following steps: performing coordinate unification processing on the multivariate topographic impact data to generate a river channel associated region; dividing the river channel associated region into a main channel sub-region and a floodplain sub-region based on topographic complexity; using the river channel associated region as a spatial alignment reference, subdividing the main channel sub-region with a 5-10m high-resolution unstructured grid and the floodplain sub-region with a 10-50m medium-resolution unstructured grid to obtain an initial river channel model; loading the river cross-section data into the initial river channel model, and then using bilinear interpolation to map the corrected DEM data to the grid nodes of the initial river channel model to generate the two-dimensional hydrodynamic model.

[0078] Specifically, coordinate unification processing is performed on the multi-dimensional topographic impact data, transforming geographic data from different sources into a unified coordinate system. This allows the data to be aligned and compared, resulting in the river-related region. The river-related region refers to the river and its surrounding related areas, including the river channel itself and the floodplains that may be submerged. Based on the complexity of the terrain, the river-related region is further divided into the main channel sub-region and the floodplain sub-region. The main channel is identified, typically the area with concentrated and fastest flow, resulting in the main channel sub-region. The main channel is usually narrower, but the water depth and velocity are greater. The floodplain is divided into multiple sub-regions, typically including areas with slower flow, resulting in the floodplain sub-regions, which may be submerged during floods but are shallower and have lower flow velocities.

[0079] Based on the different characteristics of the river channel's associated areas, the main channel sub-region and the floodplain sub-region were meshed. For the main channel sub-region, a high-resolution unstructured mesh of 5–10 meters was used; for the floodplain sub-region, a medium-resolution unstructured mesh of 10–50 meters was used. Because the river channel is complex and has a high flow velocity, a higher spatial resolution is needed to capture the dynamics of the water flow. High-resolution meshes can accurately represent the river's geometry and flow changes. Floodplains are typically vast and have slower flow rates; therefore, lower-resolution meshes can be used to reduce computational load while ensuring accurate simulation of flood spread and distribution.

[0080] River cross-section data is loaded, and the corrected DEM data is mapped onto the grid nodes of the initial river model. Bilinear interpolation is used to map the corrected DEM data onto the grid nodes, generating a two-dimensional hydrodynamic model. The river cross-section data includes cross-sectional information such as water depth and flow velocity, used to further describe changes in water flow within the river channel. The corrected DEM data includes data for filling voids. Bilinear interpolation is used to map the DEM data onto the river grid nodes, thus generating a two-dimensional hydrodynamic model with realistic topography. The two-dimensional hydrodynamic model simulates the evolution of floodwaters within the study area by solving the fundamental equations of water flow, including dynamic characteristics such as the expansion of the flood inundation area, changes in water depth, and flow velocity distribution. The final generated dynamic inundation process of the study area visually demonstrates the entire process of flooding from occurrence to recede.

[0081] When the terrain along the river channel is below the highest water level, a smaller scale topographic map (1:50,000 recommended) and a larger calculation grid (no larger than 0.5 km recommended) should be used to determine the calculation range. 2Using a selected flood analysis model, a rough calculation is performed on the inundation range of the maximum-scale flood and the levee breach. The final calculation range is then determined based on the calculation results. The water level of the maximum-scale flood can be determined using a steady non-uniform flow method. When determining the calculation range of the river channel, multiple factors need to be considered comprehensively. All main and tributary rivers that have a significant impact on the flood inflow of the designated flood protection area should be included. When determining the upper boundary of the river channel calculation range, hydrological control stations upstream of each relevant river channel are selected. If there is no hydrological control station upstream, the calculation range needs to be expanded upstream. One approach is to select control sections where the flood inflow process can be determined by engineering control; another is to select sections that can access the output results of the upstream hydrological model. Both can serve as the upper boundary. Simultaneously, the distance from the selected upper boundary of the river channel to the upstream end of the calculation range of the designated area should preferably exceed 10 times the river channel width. For the lower boundary of the river channel calculation range, downstream hydrological control stations, controlling hydraulic structures, or large water bodies such as reservoirs, lakes, and sea areas are selected. If these conditions are not met, approximate methods can be used to determine the lower boundary. The lower boundary is approximated using the Manning formula, requiring that the last measured cross-section downstream of the river be located in a straight section, and that the distance from this cross-section to the downstream end of the calculation area of ​​the compilation zone should exceed 10 times the river width. Then, based on the shape of the last river cross-section and the water surface gradient of the last two measured cross-sections under constant flow conditions in the floodplain, the lower boundary is extended downstream by more than 5 times the river width. The catchment areas on both sides of the river calculation zone belong to the interval calculation zone, and their runoff can be calculated using hydrological methods. Their runoff will flow into the corresponding river channel and / or the compilation zone. Historical flood inundation dynamics are reconstructed using a hydrodynamic model. Model construction includes two-dimensional grid generation, importing river cross-section data, and two-dimensional grid elevation interpolation.

[0082] The rationality and usability of the two-dimensional hydrodynamic model were verified. This included checking for oscillations in the calculated water level and flow processes; the reasonableness of the ratio of river flow to breach flow; any anomalies in the river surface line; any interruptions in the flood inundation area; the reasonableness of the flood arrival time distribution; any anomalies in the flow field distribution; the occurrence of negative water depths during the calculation; and the model's ability to reasonably reflect characteristics such as bridge and culvert water passage, linear landform obstruction, and internal canal diversion and flood discharge within the modeling area. The model was also verified to ensure that flood control and drainage projects were properly scheduled according to regulations, and that the water levels and flow rates upstream and downstream of these projects were reasonable. Finally, the model was assessed for the reasonableness of the inundation area and water depth distribution, and whether the flood (waterlogging) inundation characteristics were similar to those of historical floods (waterlogging) of similar magnitude. After verification, the two-dimensional hydrodynamic model was obtained.

[0083] A hydrodynamic boundary control set is established, including factors such as water inflow, outflow, velocity, flow rate, precipitation, evaporation, and soil infiltration, to ensure that the 2D hydrodynamic model can perform flood simulations under specific boundary conditions. The 2D hydrodynamic model is then driven to perform flood simulations, calculating the dynamic evolution of the flood based on the input boundary conditions, topographic data, and other hydrodynamic parameters (such as velocity and water level). Specifically, the 2D hydrodynamic model is initialized based on the input topographic data, boundary conditions, and climate data. The simulation is divided into multiple time steps, and the 2D hydrodynamic model calculates changes in water level and velocity based on the flow conditions at each time step. The calculation result for each time step represents the state of the flood at that moment. At each time step, the 2D hydrodynamic model updates the boundary conditions based on the established hydrodynamic boundary control set (such as flow rate and precipitation), influencing the flow changes within the simulation area. The 2D hydrodynamic model outputs the historical dynamic evolution of the flood, including its initiation, development, overflow, spread, and recede. The output typically includes information such as water level, velocity, flow rate, and submersion depth at different time points. The charts or animations can be used to display changes in water level and flow velocity, show the areas covered by floods, and analyze the impact range of floods at different points in time.

[0084] In summary, the deep learning-based method for dynamic reconstruction of historical flood inundation provided in this application has the following beneficial effects: by locating terrain voids on a standardized historical flood dataset, identifying missing areas, and matching suitable terrain filling models for multi-threaded parallel repair, the accuracy and adaptability of void repair are improved. The corrected terrain structure is accurately applied to the location of terrain voids for pixel-level replacement. A two-dimensional hydrodynamic model is constructed for flood simulation, which improves the accuracy and generalization ability of flood simulation, thereby enhancing the accuracy of dynamic flood reconstruction.

[0085] Example 2: Based on the same inventive concept as the deep learning-based historical flood inundation dynamic reconstruction method in Example 1, this application also provides a deep learning-based historical flood inundation dynamic reconstruction system. Please refer to the appendix. Figure 2The deep learning-based historical flood inundation dynamic reconstruction system includes: a coordinate unification and fusion module 11, used to unify and fuse the collected historical multivariate flood data and topographic multivariate impact data to obtain a standardized historical flood dataset; a missing location module 12, used to locate P missing topographic cavities in the standardized historical flood dataset, wherein the P missing topographic cavities are identified by P missing feature vectors; a model matching module 13, used to match and reuse P adaptive filling models in a topographic filling model library based on the P missing feature vectors; a multi-threaded parallel module 14, used to drive the P adaptive filling models to perform multi-threaded parallel processing of the P missing topographic cavities and output P corrected topographic structures; a pixel-level replacement module 15, used to perform pixel-level replacement of the P missing topographic cavities in the standardized historical flood dataset using the P corrected topographic structures to obtain corrected DEM data; and a flood simulation module 16, used to construct a two-dimensional hydrodynamic model based on the corrected DEM data, and drive the two-dimensional hydrodynamic model to perform flood simulation by setting a hydrodynamic boundary control set, outputting the historical flood dynamic evolution process.

[0086] Furthermore, the coordinate unification and fusion module 11 in the deep learning-based historical flood inundation dynamic reconstruction system is also used to: drive the confidence review engine to traverse and perform three-dimensional confidence reviews on the historical multivariate flood data and topographic multivariate impact data; if the review is passed, coordinate unification and fusion are performed after spatiotemporally aligning the historical multivariate flood data and topographic multivariate impact data to obtain the standardized historical flood dataset.

[0087] Furthermore, the coordinate unification fusion module 11 in the deep learning-based historical flood inundation dynamic reconstruction system is also used to: if the review fails, separate a multivariate defect dataset from the historical multivariate flood data and the topographic multivariate impact data; perform bidirectional correction on the multivariate defect dataset to obtain a multivariate corrected dataset; and after the defect data fusion and replacement of the historical multivariate flood data and the topographic multivariate impact data is performed using the multivariate corrected dataset, perform coordinate unification data fusion to output the standardized historical flood dataset.

[0088] Furthermore, the coordinate unification and fusion module 11 in the deep learning-based historical flood inundation dynamic reconstruction system is also used for: aggregating the multivariate defect dataset based on defect type, and outputting an observation defect set, a mutation defect set, and an event defect set; pre-constructing a multi-channel defect correction model, wherein the multi-channel defect correction model includes a parallel architecture of an observation correction channel, a mutation correction channel, and an event supplementary channel; loading the observation defect set, the mutation defect set, and the event defect set into the multi-channel defect correction model, and driving the observation correction channel, the mutation correction channel, and the event supplementary channel to perform bidirectional defect correction in parallel, outputting an observation correction set, a mutation correction set, and an event supplementary set; and obtaining the multivariate correction dataset by restoring the data structure of the observation correction set, the mutation correction set, and the event supplementary set.

[0089] Furthermore, the coordinate unification and fusion module 11 in the deep learning-based historical flood inundation dynamic reconstruction system is also used for: step a: driving the observation correction channel to call the same source data of neighboring stations to perform spatiotemporal interpolation of the observation defect set, and obtaining the observation correction set; step b: driving the mutation correction channel to apply the double cumulative curve method to perform systematic deviation correction on the mutation defect set, and obtaining the mutation correction set; step c: driving the event supplementation channel to perform extrapolation and supplementation on the event defect set based on the rainfall-runoff model, and obtaining the event supplementation set.

[0090] Furthermore, the missing location module 12 in the deep learning-based historical flood inundation dynamic reconstruction system is also used for: performing connected component analysis on the standardized historical flood dataset to locate multiple missing terrain cavities; selecting P missing terrain cavities with ≥100 grid cells from the multiple missing terrain cavities; extracting the first range ratio, first continuity indicator, first edge gradient, and first cavity type of the first missing terrain cavity; vectorizing and concatenating the first range ratio, first continuity indicator, first edge gradient, and first cavity type to output a first missing feature vector; and, after constructing the P missing feature vectors by analogy, associating the P missing terrain cavities and the P missing feature vectors through key-value pair mapping to generate a cavity location index matrix.

[0091] Furthermore, the model matching module 13 in the deep learning-based historical flood inundation dynamic reconstruction system is also used for: calling a standard imputation model based on a predefined multi-scale range ratio threshold of the multi-scale missing vectors to obtain multiple sample imputation models; after locally calling multiple complete DEM data samples, manually constructing invalid regions of the multiple complete DEM data samples based on the multi-scale range ratio threshold of the multi-scale missing vectors to obtain multiple multi-scale sample topographic cavity sets; using the multiple multi-scale sample topographic cavity sets and the multiple complete DEM data samples as training data, training with a weighted loss function, mapping and training the multiple sample imputation models, and obtaining the topographic imputation model library by associating and storing the multi-scale missing vectors and the multiple sample imputation models; and performing vector matching by traversing the topographic imputation model library based on the P missing feature vectors to reuse the P adapted imputation models.

[0092] Furthermore, the model matching module 13 in the deep learning-based historical flood inundation dynamic reconstruction system is also used to configure the discriminators of the multiple sample imputation models as PatchGAN architecture, and the standard imputation models include U-Net model and PFRNet model.

[0093] Furthermore, the flood simulation module 16 in the deep learning-based historical flood inundation dynamic reconstruction system is also used for: performing coordinate unification processing on the topographic multivariate impact data to generate a river channel associated region; dividing the river channel associated region into a main channel sub-region and a floodplain sub-region based on topographic complexity; using the river channel associated region as a spatial alignment reference, subdividing the main channel sub-region with a 5-10m high-resolution unstructured grid and the floodplain sub-region with a 10-50m medium-resolution unstructured grid to obtain an initial river channel model; loading the river channel cross-section data into the initial river channel model, and using bilinear interpolation to map the corrected DEM data to the grid nodes of the initial river channel model to generate the two-dimensional hydrodynamic model.

[0094] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Figure 1 The deep learning-based historical flood inundation dynamic reconstruction method and specific examples in Example 1 are also applicable to the deep learning-based historical flood inundation dynamic reconstruction system in this example.

[0095] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A deep learning-based method for dynamic reconstruction of historical flood inundation, characterized in that, include: By unifying and fusing the collected historical multivariate flood data and topographic multivariate impact data with coordinates, a standardized historical flood dataset is obtained. Locate P missing topographic cavities in the standardized historical flood dataset, wherein each of the P missing topographic cavities is identified by P missing feature vectors, including: Perform connected component analysis on the standardized historical flood dataset to locate multiple missing terrain holes; From the plurality of terrain cavity missing elements, filter out the P terrain cavity missing elements with ≥100 grid cells; Extract the first extent ratio, first continuity marker, first edge gradient, and first cavity type of the first missing terrain cavity; The first range ratio, the first continuity indicator, the first edge gradient, and the first hole type are vectorized and concatenated to output the first missing feature vector; After constructing the P missing feature vectors by analogy, the P missing terrain cavities and the P missing feature vectors are associated by key-value pair mapping to generate a cavity location index matrix. Based on the P missing feature vectors, P suitable infill models are matched and reused in the terrain infill model library, including: Based on the predefined multi-scale range ratio threshold of the multi-scale missing vector, the standard imputation model is called to obtain multiple sample imputation models; After calling multiple complete DEM data samples locally, invalid regions of the multiple complete DEM data samples are manually constructed according to the multi-scale range ratio threshold of the multi-scale missing vectors, resulting in multiple multi-scale sample terrain cavity sets. The terrain hole sets and complete DEM data of multiple samples are used as training data. A weighted loss function is used for training. After the multiple sample filling models are trained by mapping, the terrain filling model library is obtained by storing the multi-scale missing vectors and multiple sample filling models in association. Based on the P missing feature vectors, the terrain filling model library is traversed for vector matching to reuse the P fitting filling models; Drive the P adaptive filling models to perform multi-threaded parallel processing of the P terrain hole missing parts, and output P corrected terrain structures; In the standardized historical flood dataset, the P modified terrain structures are used to replace the P missing terrain holes at the pixel level to obtain modified DEM data; After constructing a two-dimensional hydrodynamic model based on the corrected DEM data, the two-dimensional hydrodynamic model is driven to perform flood simulation by setting a hydrodynamic boundary control set, and the historical flood dynamic evolution process is output.

2. The method for dynamic reconstruction of historical flood inundation based on deep learning as described in claim 1, characterized in that, A two-dimensional hydrodynamic model is constructed based on the corrected DEM data, including: The topographic multivariate impact data is processed to unify coordinates, generating river-related regions; Based on the complexity of the terrain, the river-related area is divided into the main channel sub-area and the floodplain sub-area. Using the river channel associated area as the spatial alignment reference, the main channel sub-area is divided into 5-10m high-resolution unstructured grids, and the floodplain sub-area is divided into 10-50m medium-resolution unstructured grids to obtain the initial river channel model. After loading the river cross-section data into the initial river model, the modified DEM data is mapped to the grid nodes of the initial river model using bilinear interpolation to generate the two-dimensional hydrodynamic model.

3. The method for dynamic reconstruction of historical flood inundation based on deep learning as described in claim 1, characterized in that, By unifying and fusing the collected historical multivariate flood data and topographic multivariate impact data with coordinates, a standardized historical flood dataset is obtained, including: The confidence review engine is driven to perform three-dimensional confidence reviews on the historical multivariate flood data and topographic multivariate impact data. If the review is approved, the historical multivariate flood data and topographic multivariate impact data will be spatiotemporally aligned and then fused with coordinate unification to obtain the standardized historical flood dataset.

4. The method for dynamic reconstruction of historical flood inundation based on deep learning as described in claim 3, characterized in that, Also includes: If the review fails, a multivariate defect dataset will be separated from the historical multivariate flood data and topographic multivariate impact data. The multivariate defect dataset is bidirectionally corrected to obtain a multivariate corrected dataset; After defective data fusion and replacement of the historical multivariate flood data and topographic multivariate impact data are performed using the multivariate corrected dataset, coordinate unified data fusion is performed to output the standardized historical flood dataset.

5. The deep learning-based method for dynamic reconstruction of historical flood inundation as described in claim 4, characterized in that, The multivariate defect dataset is bidirectionally corrected to obtain a multivariate corrected dataset, including: Aggregate the multivariate defect dataset based on defect type, and output the observation defect set, mutation defect set, and event defect set; A pre-constructed multi-channel defect correction model is provided, wherein the multi-channel defect correction model includes a parallel architecture observation correction channel, a mutation correction channel, and an event supplementation channel; After loading the observation defect set, mutation defect set, and event defect set into the multi-channel defect correction model, the observation correction channel, mutation correction channel, and event supplement channel are driven to perform bidirectional defect correction in parallel, and the observation correction set, mutation correction set, and event supplement set are output. The multivariate correction dataset is obtained by restoring the data structures of the observation correction set, mutation correction set, and event supplement set.

6. The method for dynamic reconstruction of historical flood inundation based on deep learning as described in claim 5, characterized in that, The observation correction channel, mutation correction channel, and event supplementation channel are driven to perform bidirectional defect correction in parallel, outputting observation correction sets, mutation correction sets, and event supplementation sets, including: Step a: Drive the observation correction channel to call the same source data from neighboring stations to perform spatiotemporal interpolation of the observation defect set, and obtain the observation correction set; Step b: Drive the mutation correction channel to apply the double cumulative curve method to perform systematic bias correction on the mutation defect set to obtain the mutation correction set; Step c: Drive the event supplement channel to extrapolate and supplement the event defect set based on the rainfall-runoff model to obtain the event supplement set.

7. The method for dynamic reconstruction of historical flood inundation based on deep learning as described in claim 1, characterized in that, The discriminators of the multiple sample imputation models are all configured with the PatchGAN architecture, and the standard imputation models include the U-Net model and the PFRNet model.

8. A deep learning-based dynamic reconstruction system for historical flood inundation, characterized in that, The step of implementing the deep learning-based historical flood inundation dynamic reconstruction method according to any one of claims 1 to 7, wherein the deep learning-based historical flood inundation dynamic reconstruction system comprises: The coordinate unification and fusion module is used to unify and fuse the collected historical multivariate flood data and topographic multivariate impact data to obtain a standardized historical flood dataset. The missing location module is used to locate P missing topographic cavities in the standardized historical flood dataset, wherein the P missing topographic cavities are identified by P missing feature vectors. The model matching module is used to match and reuse P suitable filling models in the terrain filling model library based on the P missing feature vectors. A multi-threaded parallel module is used to drive the P adaptive filling models to perform multi-threaded parallel processing of the P terrain hole missing parts, and output P corrected terrain structures. A pixel-level replacement module is used to perform pixel-level replacement of the P missing topographic holes in the standardized historical flood dataset using the P corrected topographic structures to obtain corrected DEM data. The flood simulation module is used to construct a two-dimensional hydrodynamic model based on the corrected DEM data, and then drive the two-dimensional hydrodynamic model to perform flood simulation by setting a hydrodynamic boundary control set, and output the historical flood dynamic evolution process.

Images