Urban-scale hourly water depth prediction method using spatio-temporal state space deep learning

Abstract

The application discloses a kind of urban scale hourly waterlogging depth prediction method of spatiotemporal state space deep learning, its characteristics are that hourly water depth field generated by two-dimensional Rain-on-Grid hydrodynamic model is used as supervision label, combined with rainfall time sequence, hydraulic consistency ground elevation and spatial distributed ground parameter, event level deep learning proxy model is constructed, and urban scale hourly waterlogging depth fast prediction is realized, and specific steps include: physical model supervision sample and event level sample construction, spatial feature extraction, multi-scale context aggregation and model training and prediction output etc..Compared with prior art, the application has the advantages that spatial control and time dependence of physical model are embedded into deep learning prediction process in the absence of detailed drainage pipe network data, time cost of repeated two-dimensional hydrodynamic calculation is significantly reduced, and the application can be used for urban storm waterlogging fast warning, scenario simulation and providing hourly water depth product, and has good application prospect and value of popularization and use.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing surface modeling and urban stormwater flooding prediction, specifically to an hourly urban-scale water depth prediction method that uses a two-dimensional hydrodynamic model as the supervision source and rainfall time series and static urban surface parameters as inputs through spatiotemporal state-space deep learning. Background Technology

[0002] Urban flooding caused by rainstorms is jointly controlled by rainfall intensity, urban micro-topography, surface impermeability, building obstruction, road confluence channels, infiltration losses, and water body connectivity. Two-dimensional hydrodynamic models can effectively describe the process of rainfall transforming into surface runoff, local water storage, and two-dimensional overflow. However, under large-scale, high-resolution, multi-event, and multi-scenario simulations, computational costs increase rapidly with the number of spatial grids, event duration, and scenarios, making it difficult to meet the needs of rapid prediction and frequent scenario analysis. While existing deep learning-based flood prediction methods can improve inference speed, relying directly on training with limited observed water depths often faces problems such as scarce samples, insufficient spatial coverage, and weak cross-event generalization ability. Furthermore, static input-output fitting alone is insufficient to represent the event-level temporal processes of pre-rainfall, water storage lag, and water receding.

[0003] In summary, existing urban flood prediction technologies suffer from high computational costs, scarce observation samples, weak cross-event generalization of deep learning models, and insufficient physical consistency in two-dimensional hydrodynamic models. Therefore, there is an urgent need for a rapid deep learning prediction method that uses process data generated by physical hydrodynamic models as supervised samples, complete rainstorm events as training units, and simultaneously expresses spatial heterogeneity and event-level temporal dependencies. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a spatiotemporal state-space deep learning-based method for predicting hourly water depth at the urban scale. This method uses an hourly water depth field generated by a physical model as a supervision label, taking rainfall time series and static surface parameters as input. It outputs an hourly water depth field through convolutional spatial encoding, cyclic state-space temporal propagation, multi-scale context aggregation, and U-shaped decoding, achieving rapid urban-scale flood prediction. This method concatenates dynamic rainfall fields with static urban surface descriptors, extracts spatial features such as topography, infiltration, roads, buildings, and runoff paths through a convolutional encoding module, propagates the temporal state of rainstorm events through a cyclic state-space module, and outputs an hourly water depth raster through multi-scale context aggregation, U-shaped decoding, and Softplus non-negative constraints. This invention significantly reduces the time cost of repetitive hydrodynamic simulations while maintaining the main spatial control and temporal response of the physical model, and provides hourly water depth products for urban flood warning and scenario simulation, demonstrating promising application prospects and value for widespread use.

[0005] The specific technical solution to achieve the purpose of this invention is: a method for predicting urban rainstorm flooding using physical model supervision and spatiotemporal state-space networks, characterized by the following steps:

[0006] Step 1: Construct supervised samples for the physical model

[0007] Using a two-dimensional Rain-on-Grid hydrodynamic model, several historical rainstorm events were simulated under the condition of no one-dimensional drainage network coupling. The output hourly water depth field was used as a supervision label to construct a physical model supervision sample.

[0008] Step 2: Construct model input data

[0009] Acquire rainfall time series, hydraulically consistent surface elevation, surface roughness, impermeability, maximum deficit, initial deficit, constant rate and static surface parameters, and unify them into spatially consistent raster input to construct model input data.

[0010] Step 3: Construct event-level samples

[0011] Training samples are organized according to complete rainstorm events. Each rainstorm event is divided into several spatial slice sequences, preserving the temporal order and cyclic state transmission relationship within the event to construct event-level samples.

[0012] Step 4: Spatial Feature Extraction

[0013] The rainfall input at the current moment or in historical time periods is spliced ​​with static surface attributes, and multi-scale spatial features are extracted by a convolutional coding module.

[0014] Step 5: Event-level timing modeling

[0015] The encoded features are input into the cyclic state space module, the spatial features are projected into potential patch tokens, and the hidden states are updated in chronological order to construct an event-level time series model that expresses the impact of previous rainfall, accumulated water storage, and hydraulic connectivity evolution on subsequent water accumulation.

[0016] Step 6: Multi-scale context aggregation

[0017] The dilated convolution context aggregation module integrates spatial features at different scales and expands the model's receptive field through multi-dilation rate convolution, thereby enhancing the model's ability to express terrain undulations, rainfall response, and water accumulation distribution patterns at different scales. The spatial features include: local neighborhood features, nearby region features, and larger area features.

[0018] Step 7: Depth Field Decoding Output

[0019] The original spatial resolution of the water depth field is restored by upsampling, convolutional reconstruction and skip connections. The water depth output is constrained by Softplus non-negative activation to output the hourly water depth field.

[0020] Step 8: Model Training and Predicted Output

[0021] Using the hourly water depth field of a two-dimensional hydrodynamic model as the supervised target, the model is trained using a composite loss function. During the inference phase, new rainfall sequences and fixed surface parameters are input, and the hourly water depth raster of urban areas is output. The predicted output is then converted into water accumulation range, maximum water depth, road flooding status, risk level, and emergency mapping products.

[0022] The two-dimensional Rain-on-Grid hydrodynamic model is the HEC-RAS two-dimensional model or other two-dimensional surface runoff models based on shallow water equations. The hourly water depth field output by the two-dimensional Rain-on-Grid hydrodynamic model is the supervision label of the deep learning surrogate model.

[0023] The static surface parameters include one or more of the following: hydraulically consistent digital surface model, land cover category, Manning roughness coefficient, impermeability ratio, maximum deficit, initial deficit, constant infiltration rate, building constraints, road constraints, and water body constraints.

[0024] The construction of the event-level samples is carried out by dividing the training, verification and testing into complete rainstorm events, with a division ratio of 6:1:3.

[0025] The convolutional coding module includes: two-dimensional convolution, group normalization, and SiLU activation.

[0026] The cyclic state space module converts spatial features into a spatiotemporal token sequence and performs event-level state propagation, which includes: patch projection, selective state space update, back projection, and residual fusion steps.

[0027] The dilated convolution context aggregation module includes: parallel convolution branches with different dilation rates and pooling context branches.

[0028] The composite loss function includes: wet area weighted Smooth-L1 loss, full-pixel RMSE loss, and dry area false alarm penalty loss, which are used to constrain the depth accuracy of water accumulation grids, global spatial consistency, and false positive water depth in non-water accumulation areas, respectively.

[0029] The predicted output includes one or more of the following: hourly water depth, maximum water depth, waterlogged area, road flooding status, and urban flooding risk level.

[0030] Compared with the prior art, the present invention has the following beneficial technical effects and significant technical progress:

[0031] 1) Physical supervision enhances generalization ability: Supervision samples are generated through a two-dimensional hydrodynamic model to alleviate the problem of scarce measured water depth samples for urban flooding, and the deep learning prediction process inherits the topographic, infiltration and hydraulic connectivity constraints in the physical model.

[0032] 2) Event-level time series modeling capability: The model is trained with complete rainstorm events rather than isolated time slices. The cyclic state space module can propagate the implicit state of the event, enabling the model to express the early rainfall, water storage lag and water receding process.

[0033] 3) Urban spatial heterogeneity representation capability: The convolutional encoder, combined with hydraulically consistent digital surface models, land cover and distributed infiltration parameters, can identify low-lying areas, road confluence channels, building barriers and differences in surface parameters.

[0034] 4) Multi-scale confluence relationship expression capability: The dilated convolution context aggregation module expands the effective receptive field, expressing cross-neighborhood confluence contributions while preserving street-scale details.

[0035] 5) Improved physical rationality of water depth prediction: Softplus non-negative output avoids negative water depth, and the composite loss function simultaneously constrains wet grid accuracy, global consistency and false alarms in dry areas, reducing non-physical shallow water noise.

[0036] 6) Rapid prediction and business application capabilities: Compared to repeatedly running two-dimensional hydrodynamic models, the trained surrogate model can quickly output hourly water depth products at the city scale, which are suitable for rainstorm warnings, emergency dispatching, risk map production and scenario analysis. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the process of Embodiment 1 of the present invention;

[0038] Figure 2 This is a schematic diagram of the Huayu-Urban deep learning agent model structure;

[0039] Figure 3 A schematic diagram of the event-level time-series state space update module;

[0040] Figure 4 This diagram illustrates event-level training, composite loss, and fast prediction output. Detailed Implementation

[0041] The present invention specifically includes the following steps:

[0042] Step 1: Construction of Physically Supervised Samples

[0043] A two-dimensional Rain-on-Grid hydrodynamic model was used to simulate historical rainstorm events hourly, and the water depth grid at each moment was output as a supervision label. The preferred hydrodynamic model was the HEC-RAS v6.6 two-dimensional Rain-on-Grid model, which was not coupled to the one-dimensional drainage network. The model was driven by rainfall input, hydraulically consistent digital surface model and distributed surface parameters.

[0044] Step 2: Input Data Organization

[0045] The rainfall time series, static surface properties, and effective pixel mask are unified into a spatially consistent raster sequence. The static surface properties include: hydraulically consistent digital surface model, roughness, impermeability, maximum deficit, initial deficit, constant rate, and other surface parameters.

[0046] Step 3: Constructing Event-Level Slice Samples

[0047] Using complete rainstorm events as the basic sample unit, and maintaining the temporal order within each event, the urban area is divided into a spatial slice sequence. Each slice contains inputs and water depth monitoring labels in the complete time dimension.

[0048] Step 4: Convolutional Spatial Encoding

[0049] The dynamic rainfall channel and the static surface channel are spliced ​​together and then input into a convolutional encoder. Spatial features such as terrain depressions, road channels, building barriers, infiltration differences and local confluence paths are extracted through two-dimensional convolution, group normalization and SiLU activation.

[0050] Step 5: Cyclic State-Space Temporal Modeling

[0051] A cyclic state space module is introduced in each layer of the encoder to project spatial features into potential patch tokens, update the hidden state in chronological order, and restore it into a spatial feature map through back projection and residual fusion to express the evolution of early rainfall, accumulated water storage and water surface connectivity in rainstorm events.

[0052] Step 6: Multi-scale context aggregation

[0053] Parallel dilated convolution branches and pooling context branches are used to integrate spatial information at different scales, enhancing the model's ability to express cross-neighborhood confluence and larger catchment contribution areas.

[0054] Step 7: Depth Decoding and Non-negative Output

[0055] The spatial resolution is restored step by step through upsampling and convolution using a U-shaped decoding structure, and shallow detail features are fused through skip connections. Finally, the non-negative water depth is output after Softplus activation.

[0056] Step 8: Composite Loss Training

[0057] The model is trained using a combination of wet area weighted Smooth-L1 loss, full-pixel RMSE loss, and dry area false alarm penalty loss, which respectively constrain the depth error of water accumulation area, global spatial consistency, and false water accumulation in dry area.

[0058] Step 9: Rapid Reasoning and Product Output

[0059] Once the model is trained, inputting a new rainfall sequence and fixed surface parameters will quickly generate an hourly water depth field, and further generate water accumulation range, maximum water depth, and risk classification maps.

[0060] The wetland-weighted Smooth-L1 loss refers to a smoothed absolute error loss that assigns higher weights to waterlogged areas, used to enhance the model's constraint on water depth errors in waterlogged pixels. Let the first... The output water depth of the two-dimensional hydrodynamic model with effective pixels is... The deep learning model predicts the water depth as The error between the two is:

[0061]

[0062] The Smooth-L1 loss can be expressed as:

[0063]

[0064]

[0065] in, This is the error smoothing threshold. For cases where the actual water depth is greater than the wetland determination threshold... Pixels with a high water depth are assigned a higher weight; pixels with a true water depth no greater than this threshold are assigned a normal weight. The wetland weighted Smooth-L1 loss can be expressed as:

[0066]

[0067] in, For the number of effective pixels, For cell weights. When hour, ;when hour, . This is the weighting coefficient for wet areas. This loss reduces the dominant role of a large number of dry area pixels in model training, improving the accuracy of water depth prediction in waterlogged areas.

[0068] The full-pixel RMSE loss refers to the root mean square error loss calculated across all valid pixels, used to constrain the overall consistency between the predicted water depth and the output water depth of the two-dimensional hydrodynamic model. It can be expressed as:

[0069]

[0070] The false alarm penalty loss for dry areas refers to the penalty imposed on pixels whose actual water depth is less than or equal to the dry area determination threshold, but whose model-predicted water depth exceeds the allowable threshold, in order to reduce false shallow water predictions in non-waterlogged areas. It can be expressed as:

[0071]

[0072] in, For the effective set of pixels in the dry area, The number of effective pixels in the dry area. This represents the permissible water depth threshold for the dry zone. The total loss during model training can be expressed as:

[0073]

[0074] in, , and These are the weighting coefficients for the three losses.

[0075] RMSE, or root mean square error, is used to evaluate the overall deviation between the predicted water depth and the water depth output by the two-dimensional hydrodynamic model. Its calculation formula is as follows:

[0076]

[0077] in, The number of valid pixels participating in the evaluation. For the first The predicted water depth per pixel, For the first The supervised water depth is measured in pixels. RMSE is more sensitive to larger errors and can be used to reflect the model's ability to control local deep water accumulation or anomalous errors.

[0078] MAE, or Mean Absolute Error, is used to evaluate the average absolute deviation between predicted and monitored water depths. Its calculation formula is as follows:

[0079]

[0080] MAE reflects the average water depth error of the model across all valid pixels. It is less sensitive to a small number of extreme errors than RMSE and can be used in conjunction with RMSE to evaluate the stability and overall accuracy of the model's predictions.

[0081] The present invention will be further described in detail below using a central urban area of ​​a city as an example. Unless otherwise specified, the software, hardware, and conventional data processing steps involved in the embodiments can be implemented using methods commonly used in the art. This embodiment is only used to explain the present invention and should not be construed as limiting the scope of protection of the present invention.

[0082] Example 1

[0083] See Figure 1 A rapid prediction method for urban rainstorm flooding based on physical model supervision and spatiotemporal state space network is proposed, and the specific steps are as follows.

[0084] Step 1: Construction of Physical Supervision Data

[0085] 1-1: Study Area and Basic Data

[0086] See Figure 1 The example uses the central urban area of ​​a city as the implementation area, and uses topographic surface parameters (hydraulic consistent digital surface model, infiltration parameters) and historical rainfall sequences as inputs for the two-dimensional model.

[0087] 1-2: Rainfall-driven

[0088] See Figure 1 The historical rainfall sequence uses historical rainfall records from 11 rain gauge stations in the city. Spatial partitioning is performed using Thiessen polygons, and the time series of each rain gauge station is assigned to the corresponding polygon area to form a spatially variable rainfall input.

[0089] 1-3: Two-dimensional hydrodynamic model simulation

[0090] A two-dimensional Rain-on-Grid model based on HEC-RAS v6.6 was used for stormwater simulation, decoupled from the one-dimensional drainage network. The two-dimensional grid contained 1,478,300 elements with an average face length of 20 m, and no break lines were set. The calculation time step was 1 hour, and the intervals for hydrological processes, maps, and detailed outputs were 1 hour. Hourly water depth fields were output for ten historical typhoon-related rainstorm events, with Bebinca-Prasang treated as a composite rainfall event. The hourly water depth fields served as hourly water depth labels.

[0091] Step 2: Event-level sample slicing

[0092] 2-1: Dynamic input includes: historical rainfall sequence, cumulative rainfall at multiple times and hourly water depth labels, where cumulative rainfall at multiple times is used to express the impact of previous rainfall.

[0093] 2-2: Static inputs include: topographic surface parameters (hydraulic consistent digital surface model, infiltration parameters).

[0094] 2-3: See reference Figure 4Using complete rainstorm events as units, dynamic and static inputs are used as model inputs, and hourly water depth labels are used as supervision data to construct training, validation, and test samples. The training set uses 6 rainfall events, the validation set uses 1 independent event, and the test set uses 3 independent events, which are further divided into 300×300 pixel spatial slice sequences.

[0095] Step 3: Building the Huayu-Urban Model

[0096] 3-1: Convolutional Coding Module

[0097] See Figure 2 Historical rainfall sequences, cumulative rainfall over multiple time periods, and topographic surface parameters are input into the feedforward layer and processed by convolution and normalization to form initial features. These initial features are then sequentially fed into coding layer 1, coding layer 2, and coding layer 3. Each coding layer extracts spatial features at different scales through convolution and normalization.

[0098] 3-2: Cyclic State Space Module

[0099] See Figure 2 The features output by each coding layer are input into the corresponding state layer 1, state layer 2 and state layer 3 respectively, and the temporal state is updated through the selective state update module.

[0100] See Figure 3 The input feature map is first processed by patch projection to divide and map spatial features into feature representations suitable for temporal modeling. Then, token generation is performed to obtain a feature sequence input to the selective state space update module. This module combines continuous state information from times T-1, T, and T+1 to update the temporal state of the current features, characterizing the state propagation relationships during hourly water depth changes. The updated features are then processed by backprojection and residual fusion. Backprojection restores the token features to a spatial feature map, while residual fusion preserves the original spatial information in the input features. The final output feature map serves as the input for subsequent decoding and water depth prediction.

[0101] 3-3: Dilated Convolution

[0102] See Figure 2 The output of state layer 3 is further input into the dilated convolution module to aggregate spatial context information and obtain high-level spatiotemporal features.

[0103] 3-4: Depth Field Decoding and Softplus Activation

[0104] See Figure 2The high-level spatiotemporal features are sequentially processed through decoding layer 2 and decoding layer 1 for spatial feature recovery. During the decoding process, the output of state layer 2 is input into decoding layer 2 via a skip connection, and the output of state layer 1 is input into decoding layer 1 via a skip connection. This enables the decoding process to fuse spatiotemporal state features at different scales and restore spatial resolution step by step through upsampling and convolution. Finally, the decoding result is input into the prediction head, and after being constrained by the Softplus function, the dynamic result, i.e., the hourly water depth prediction result, is output.

[0105] 3-5: Composite Loss

[0106] See Figure 4 The model training process employs a composite training objective for constraint, which includes a wet-area weighted Smooth_L1 loss, a full-pixel RMSE loss, and a dry-area false alarm penalty term. The wet-area weighted Smooth_L1 loss is used to improve the accuracy of water depth prediction, the full-pixel RMSE loss is used for model selection during the validation phase, and the dry-area false alarm penalty term is used to suppress misjudgments in non-waterlogged areas, thereby improving the stability and reliability of hourly water depth prediction results.

[0107] Step 4: Implementation of the Proxy Model

[0108] 4-1: The RMSE and MAE were evaluated using three independent events: Bebinca-Prasan, Plum Blossom, and Wimbiya. The results were also used to calculate the global single-time step inference time.

[0109] 4-2: After training is completed, input a new rainfall sequence and fixed surface parameters to output hourly water depth raster. The prediction results can be further converted into water accumulation range, maximum water accumulation depth, road flooding status, risk level and emergency mapping products.

[0110] In this embodiment, the prediction accuracy and computational efficiency of the independent test events are shown in Table 1 below:

[0111] Table 1 Implementation Data and Predicted Results

[0112]

[0113] Where RMSE and MAE are in units of m, and tstep represents the average inference time per time step across the entire domain.

[0114] The above embodiments are merely illustrative of the present invention and are not intended to limit the invention. All equivalent transformations, substitutions, or combinations made based on the principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A spatiotemporal state-space deep learning method for predicting hourly water depth at the urban scale, characterized in that, The method includes the following steps: Step 1: Construct supervised samples for the physical model Using a two-dimensional Rain-on-Grid hydrodynamic model, several historical rainstorm events were simulated under the condition of no one-dimensional drainage network coupling. The output hourly water depth field was used as a supervision label to construct a physical model supervision sample. Step 2: Construct model input data Acquire rainfall time series, hydraulically consistent surface elevation, surface roughness, impermeability, maximum deficit, initial deficit, constant rate and static surface parameters, and unify them into spatially consistent raster input to construct model input data; Step 3: Construct event-level samples Training samples are organized according to complete rainstorm events. Each rainstorm event is divided into several spatial slice sequences, preserving the temporal order and cyclic state transmission relationship within the event to construct event-level samples. Step 4: Spatial Feature Extraction The rainfall input at the current moment or in historical time periods is spliced ​​with static surface attributes, and multi-scale spatial features are extracted through a convolutional coding module; Step 5: Event-level timing modeling The encoded features are input into the cyclic state space module, the spatial features are projected into potential patch tokens, and the hidden state is updated in chronological order to construct an event-level time series model that expresses the impact of previous rainfall, accumulated water storage, and hydraulic connectivity evolution on subsequent water accumulation. Step 6: Multi-scale context aggregation A dilated convolution context aggregation module is used to integrate spatial features at different scales. Multi-dilation rate convolution is used to expand the receptive field of the model and enhance the model's ability to express terrain undulations, rainfall response and water accumulation distribution patterns at different scales. The spatial features include: local neighborhood features, nearby region features and range region features. Step 7: Depth Field Decoding Output The original spatial resolution of the water depth field is restored by upsampling, convolutional reconstruction and skip connections. The water depth output is constrained by Softplus non-negative activation and outputs the time-by-time water depth field. Step 8: Model Training and Predicted Output Using the hourly water depth field of a two-dimensional hydrodynamic model as the supervised target, the model is trained using a composite loss function. During the inference phase, new rainfall sequences and fixed surface parameters are input, and the hourly water depth raster of urban areas is output. The predicted output is then converted into water accumulation range, maximum water depth, road flooding status, risk level, and emergency mapping products.

2. The urban-scale hourly water depth prediction method based on spatiotemporal state spatial deep learning according to claim 1, characterized in that, The two-dimensional Rain-on-Grid hydrodynamic model is either the HEC-RAS two-dimensional model or a two-dimensional surface runoff model based on shallow water equations. The hourly water depth field output by the two-dimensional Rain-on-Grid hydrodynamic model serves as the supervision label for the deep learning proxy model.

3. The urban-scale hourly water depth prediction method based on spatiotemporal state spatial deep learning according to claim 1, characterized in that, The static surface parameters include one or more of the following: hydraulically consistent digital surface model, land cover category, Manning roughness coefficient, impermeability ratio, maximum deficit, initial deficit, constant infiltration rate, building constraints, road constraints, and water body constraints.

4. The urban-scale hourly water depth prediction method based on spatiotemporal state deep learning according to claim 1, characterized in that, The construction of the event-level samples is carried out by dividing the training, verification and testing into complete rainstorm events, with a division ratio of 6:1:

3.

5. The urban-scale hourly water depth prediction method based on spatiotemporal state spatial deep learning according to claim 1, characterized in that, The convolutional coding module includes: two-dimensional convolution, group normalization, and SiLU activation.

6. The urban-scale hourly water depth prediction method based on spatiotemporal state spatial deep learning according to claim 1, characterized in that, The cyclic state space module converts spatial features into a spatiotemporal token sequence and performs event-level state propagation, which includes: patch projection, selective state space update, back projection, and residual fusion steps.

7. The urban-scale hourly water depth prediction method based on spatiotemporal state spatial deep learning according to claim 1, characterized in that, The dilated convolution context aggregation module includes: parallel convolution branches with different dilation rates and pooling context branches.

8. The urban-scale hourly water depth prediction method based on spatiotemporal state spatial deep learning according to claim 1, characterized in that, The composite loss function includes: wet area weighted Smooth-L1 loss, full-pixel RMSE loss, and dry area false alarm penalty loss, which are used to constrain the depth accuracy of water accumulation grids, global spatial consistency, and false positive water depth in non-water accumulation areas, respectively.

9. The urban-scale hourly water depth prediction method based on spatiotemporal state spatial deep learning according to claim 1, characterized in that, The predicted output includes one or more of the following: hourly water depth, maximum water depth, waterlogged area, road flooding status, and urban flooding risk level.

Images