A flood peak prediction method and device

Abstract

The application discloses a flood peak prediction method and system, the method comprising: obtaining topographic elevation data of a target area and target rainfall pattern information; inputting the topographic elevation data and the target rainfall pattern information into a pre-trained flood peak prediction model to obtain a flood peak prediction result, the prediction result comprising a maximum water depth spatial distribution of the target area; wherein the flood peak prediction model comprises a generation network constructed based on a conditional generative adversarial network, the generation network introducing a topographic attention mechanism and a multi-scale rainfall encoding mechanism in the process of generating the maximum water depth spatial distribution, and the topographic feature being a feature representation obtained by feature extraction on the topographic elevation data. The application can directly output the maximum water depth spatial distribution in the prediction stage, reduce the dependence on numerical simulation iteration, shorten the result acquisition period, and enhance the adaptability and output stability under different rainfall pattern scenarios.

Description

Technical Field

[0001] This application relates to the field of rainfall-induced flood peak forecasting technology, and in particular to a flood peak prediction method and apparatus. Background Technology

[0002] With the intensification of global climate change and the continuous advancement of urbanization, the frequency and intensity of extreme rainfall events are on the rise. Floods, including urban flooding, flash floods, and river overflows caused by torrential rains, pose serious threats to human safety, transportation operations, and the stability of critical infrastructure. In flood control, disaster reduction, and emergency response scenarios, in addition to focusing on the temporal changes of flood events, management departments and emergency systems often pay more attention to the "peak results" of flood impacts, such as the maximum water depth / maximum water depth distribution that may be reached at various locations in a watershed (or urban area) under given rainfall conditions. This information is used to support decision-making in risk zoning, early warning issuance, evacuation route planning, and the deployment of rescue resources. Therefore, obtaining reliable spatial distribution predictions of flood peaks (maximum water depths) within a limited timeframe has become a crucial fundamental issue in flood early warning and rapid response.

[0003] In related technologies, flood or waterlogging prediction typically relies on physical mechanism models or numerical simulation methods. For example, two-dimensional hydrodynamic models based on partial differential equations such as the shallow water equation are used, employing grid discretization and numerical iteration to obtain hydrodynamic elements such as water depth and flow velocity. Alternatively, methods based on discrete grid-point cellular automata can be used to approximate the flood diffusion process through rule-driven iterative evolution. While these methods can reflect the water propagation mechanism to some extent, they often require relatively complete configuration of parameters such as topography, roughness, and boundary conditions, and depend on high computational resources and long simulation times. When rapid comparison of multiple rainfall scenarios or real-time simulations are needed in emergency warnings, they are easily limited by computational efficiency and modeling barriers. Meanwhile, some data-driven / machine learning methods have attempted to use historical samples or simulation results as supervised data to map inputs such as topography and rainfall to outputs such as water depth. However, in practical applications, there may still be problems such as insufficient utilization of external conditions (especially the spatial differences between rainfall patterns and topography), limited expressive power due to simple splicing of input features, and insufficient model reuse and adaptation capabilities when facing multiple rainfall scenarios. These problems affect the accuracy and stability of flood peak prediction and make it difficult to meet the engineering requirements of "fast, accurate, and stable".

[0004] Therefore, in flood peak prediction based on rainfall and topographic information, the long computation time of numerical simulation, the difficulty in fully characterizing the coupling effect of rainfall patterns and topographic spatial differences, and the insufficient model adaptation and generalization ability under multiple rainfall scenarios have become urgent problems to be solved. Summary of the Invention

[0005] This application provides a flood peak prediction method and apparatus, aiming to solve the problems of long numerical simulation calculation time, difficulty in fully characterizing the coupling effect of rainfall patterns and spatial differences in terrain, and insufficient model adaptation and generalization ability under multiple rainfall scenarios in the existing technology for flood peak prediction based on rainfall and topographic information.

[0006] In a first aspect, a flood peak prediction method, the method comprising: Acquire topographic elevation data and rainfall pattern information for the target area; The topographic elevation data and the target rainfall pattern information are input into a pre-trained flood peak prediction model to obtain the flood peak prediction result for the target area. The flood peak prediction result includes the spatial distribution of the maximum water depth in the target area. The flood peak prediction model includes a generator network based on a conditional generative adversarial network. In the process of generating the spatial distribution of the maximum water depth, the generator network introduces a terrain attention mechanism and a multi-scale rainfall coding mechanism. The terrain attention mechanism is used to weight terrain features, and the multi-scale rainfall coding mechanism is used to inject the target rainfall pattern information into the generation process in a multi-scale conditional form. The terrain features are feature representations obtained by feature extraction from the terrain elevation data.

[0007] Optionally, in the above scheme, the method further includes: The terrain elevation data is input into a terrain feature extraction network to obtain at least two layers of terrain feature representations at different scales; The terrain features are generated based on the terrain feature representations at at least two different scales.

[0008] Optionally, in the above scheme, the step of inputting the topographic elevation data and the target rainfall pattern information into a pre-trained flood peak prediction model to obtain the flood peak prediction result for the target area includes: The terrain elevation data is divided into multiple terrain data blocks; For any one of the multiple terrain data blocks, the terrain data block and the target rainfall pattern information are input into the flood peak prediction model to obtain the maximum water depth prediction data block corresponding to the terrain data block; Based on the spatial relationship of multiple maximum water depth prediction data blocks, the multiple maximum water depth prediction data blocks are reconstructed to obtain the spatial distribution of the maximum water depth.

[0009] Optionally, in the above scheme, the step of reconstructing multiple maximum water depth prediction data blocks based on their spatial positional relationships to obtain the spatial distribution of the maximum water depth includes: Determine the spatial relationship between the plurality of maximum water depth prediction data blocks in the target area; According to the spatial relationship, the multiple maximum water depth prediction data blocks are mapped to a unified grid position in the target area to obtain multiple mapped water depth layers; For any grid position in the unified grid positions, determine one or more mapped water depth layers covering the grid position, and obtain one or more water depth candidate values ​​corresponding to the grid position; A fusion process is performed on the one or more candidate water depth values ​​to obtain the reconstructed water depth value at the grid location; wherein the fusion process includes mean fusion, weighted fusion, and / or median fusion; The spatial distribution of the maximum water depth is obtained based on the reconstructed water depth values ​​of each grid position in the unified grid position.

[0010] Optionally, in the above scheme, the generating network introduces a terrain attention mechanism during the generation of the maximum water depth spatial distribution, including: Attention weights are generated based on the terrain features; The terrain features are weighted using the attention weights to obtain terrain enhancement features; Based on the terrain enhancement features, the spatial distribution of the maximum water depth is generated.

[0011] Optionally, in the above scheme, generating attention weights based on the terrain features includes: The terrain features are subjected to average pooling in the channel dimension to obtain the first converged feature; Max pooling is performed on the terrain features in the channel dimension to obtain the second converged feature; The attention weights are generated based on the first convergence feature and the second convergence feature.

[0012] Optionally, in the above scheme, the generating network introduces a multi-scale rainfall coding mechanism during the generation of the maximum water depth spatial distribution, including: The target rainfall pattern information is encoded to obtain rainfall condition features at least two different scales; The rainfall condition features are injected into at least two feature layers of different scales in the generation network, and the spatial distribution of maximum water depth is generated based on the generation process after the rainfall condition features are injected.

[0013] Optionally, in the above scheme, the flood peak prediction model is obtained by training a conditional generative adversarial network, and the training of the conditional generative adversarial network includes: For each training sample, the terrain elevation training data and rainfall pattern information corresponding to the training sample are input into the generation network, and the maximum water depth prediction result is output. The maximum water depth prediction result and the corresponding sample maximum water depth data are input into the discriminant network so that the discriminant network can learn to discriminate between the maximum water depth prediction result and the sample maximum water depth data. The discrimination network is configured as a patch discrimination network, which divides the maximum water depth data into multiple local patches and outputs local discrimination results for each patch; the maximum water depth data includes the maximum water depth prediction results and the sample maximum water depth data; The local discrimination results based on multiple local patches are used to generate a discrimination output for the maximum water depth data, thereby completing the adversarial training of the conditional generative adversarial network.

[0014] Optionally, in the above scheme, the training of the conditional generative adversarial network further includes: The discriminant network is used to output the rainfall pattern regression results corresponding to the training samples; A regression loss is constructed based on the regression results of the rainfall patterns and the rainfall pattern information corresponding to the training samples; The parameters of the generator network and / or the discriminator network are updated based on the regression loss to introduce rainfall pattern regression constraints to complete the training of the conditional generative adversarial network.

[0015] In a second aspect, a flood peak prediction device includes a processor and a memory, the memory storing a computer program that, when executed on the processor, causes the device to perform the flood peak prediction method described in the first aspect.

[0016] Compared with the prior art, this application has at least the following beneficial effects: Based on further analysis and research of existing technical problems, this application recognizes that existing technologies for flood peak prediction based on rainfall and topographic information suffer from drawbacks such as long computation time in numerical simulation, difficulty in fully characterizing the coupling effect of rainfall patterns and topographic spatial differences, and insufficient model adaptation and generalization ability under multiple rainfall scenarios. This application addresses these issues by acquiring topographic elevation data and target rainfall pattern information for the target area, and using these as inputs to a pre-trained flood peak prediction model. The model then directly outputs flood peak prediction results that include the spatial distribution of maximum water depth. The generative network of this model incorporates a topographic attention mechanism and a multi-scale rainfall encoding mechanism during the generation process. Furthermore, the terrain features are obtained by feature extraction from terrain elevation data. Based on the above mechanism, the model does not need to perform time-consuming numerical iteration solutions during the prediction stage. It can stably generate the maximum water depth spatial distribution that matches the terrain spatial differences and rainfall patterns under the same input conditions. Therefore, it can reduce the dependence on complex physical modeling and long-term simulation, shorten the chain of obtaining flood peak results, and improve the adaptability and output stability of different rainfall pattern scenarios. This solves the problems in the background technology of "long numerical simulation calculation time, difficulty in fully characterizing the coupling effect of rainfall patterns and terrain spatial differences, and insufficient model adaptability and generalization ability under multiple rainfall scenarios". Attached Figure Description

[0017] Figure 1 A flowchart illustrating a flood peak prediction method provided in one embodiment of this application; Figure 2 A basic framework diagram of a deep flood adversarial generative network provided in one embodiment of this application; Figure 3 A multi-scale rainfall coding network structure diagram provided in one embodiment of this application; Figure 4 Experimental results of real depth maps, predicted depth maps, and error maps on the Zurich dataset provided in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] In one embodiment, such as Figure 1 As shown, a flood peak prediction method is provided, including the following steps: Acquire topographic elevation data and rainfall pattern information for the target area; The topographic elevation data and the target rainfall pattern information are input into a pre-trained flood peak prediction model to obtain the flood peak prediction result for the target area. The flood peak prediction result includes the spatial distribution of the maximum water depth in the target area. The flood peak prediction model includes a generator network based on a conditional generative adversarial network. In the process of generating the spatial distribution of the maximum water depth, the generator network introduces a terrain attention mechanism and a multi-scale rainfall coding mechanism. The terrain attention mechanism is used to weight terrain features, and the multi-scale rainfall coding mechanism is used to inject the target rainfall pattern information into the generation process in a multi-scale conditional form. The terrain features are feature representations obtained by feature extraction from the terrain elevation data.

[0020] This implementation proposes a flood peak prediction method for predicting the spatial distribution of flood peak (maximum water depth) in a given target area. The target area can be an urban flood-prone area, a watershed catchment area, or a river network-affected area, etc. The topographic elevation data can be digital elevation model (DEM) raster data, recording the elevation value of each raster location within the target area. Target rainfall pattern information is used to characterize the rainfall scenario under the prediction conditions and can consist of one or more parameters, such as rainfall intensity level, duration, return period, rainfall type, or rainfall condition information encoded by time series / spatial distribution.

[0021] During method execution, topographic elevation data and target rainfall pattern information are input into a pre-trained flood peak prediction model to directly obtain the flood peak prediction result. The flood peak prediction result includes at least the spatial distribution of the maximum water depth in the target area. That is, under the same spatial coordinates and resolution as the topographic elevation data, a maximum water depth prediction value is output for each grid location, thereby forming a maximum water depth distribution map (peak water depth map).

[0022] The flood peak prediction model includes a generative network based on a conditional generative adversarial network. The generative network outputs the spatial distribution of maximum water depth, conditioned on "topography and rainfall patterns." To make the generative network more responsive to input conditions, a topographic attention mechanism and a multi-scale rainfall encoding mechanism are introduced during the generation process. Terrain attention mechanism: Weighting terrain features to form terrain enhancement features, which then participate in the generation of maximum water depth spatial distribution; Multi-scale rainfall coding mechanism: The target rainfall pattern information is injected into the generation process in the form of multi-scale conditions, so that the generation network obtains rainfall condition constraints in feature layers at different scales.

[0023] Among them, terrain features are feature representations obtained by extracting features from terrain elevation data. Terrain feature extraction can be understood as mapping the original elevation raster to a feature space that is more suitable for deep network processing. Features can simultaneously carry information on local elevation changes (such as slope / undulation) and regional morphology (such as water catchment trends / terrain connectivity), providing basic input for subsequent attention weighting and condition generation.

[0024] This embodiment uses topographic elevation data and target rainfall pattern information as input conditions, and directly outputs the spatial distribution of maximum water depth from a pre-trained flood peak prediction model. Simultaneously, attention weighting is applied to topographic features and multi-scale condition injection is performed on the rainfall pattern during the generation process. This application can obtain peak water depth distribution results in an end-to-end manner during the prediction stage, thereby reducing the computational time and deployment complexity caused by traditional numerical simulation iterative solutions. It also strengthens the coupled expression of rainfall patterns and topographic spatial differences in peak prediction, improving adaptability and output stability under multiple rainfall scenarios, thus solving the problems of "long computational time, insufficient coupling characterization, and insufficient scenario generalization and adaptability" in the background technology.

[0025] In this embodiment, the method further includes: The terrain elevation data is input into a terrain feature extraction network to obtain at least two layers of terrain feature representations at different scales; The terrain features are generated based on the terrain feature representations at at least two different scales.

[0026] This embodiment further provides a specific method for generating terrain features. First, terrain elevation data is input into a terrain feature extraction network. This network can be a multi-layer convolutional coding structure, or a feature extraction structure that includes downsampling, residual blocks, normalization, and nonlinear activation, used to convert the elevation raster into a multi-scale feature representation.

[0027] The terrain feature extraction network outputs at least two layers of terrain feature representations at different scales. For example, the first-scale feature can preserve high spatial resolution to express local elevation changes; the second-scale feature can obtain a larger receptive field through downsampling to express the regional terrain morphology and spatial connectivity. Feature representations at different scales can come from the outputs of different network layers or from the outputs of different branches of the pyramid structure.

[0028] After obtaining at least two layers of terrain feature representations at different scales, the terrain features are generated based on these representations. Specifically, this can be achieved by: aligning features at different scales to a unified spatial scale and then stitching them together, followed by convolutional mapping to obtain terrain features with a unified channel dimension; or by using a scale-wise weighted fusion to obtain the terrain features. The generated terrain features are then used for weighted processing in a terrain attention mechanism and for generating the spatial distribution of maximum water depth.

[0029] This embodiment extracts multi-scale features from terrain elevation data to form terrain features, enabling terrain representation to simultaneously include local and global terrain information. This provides a more sufficient and stable input basis for the subsequent weight generation of the terrain attention mechanism and the weighting of terrain features, thereby improving the expressibility of terrain conditions in the peak prediction process.

[0030] In this embodiment, the step of inputting the topographic elevation data and the target rainfall pattern information into a pre-trained flood peak prediction model to obtain the flood peak prediction result for the target area includes: The terrain elevation data is divided into multiple terrain data blocks; For any one of the multiple terrain data blocks, the terrain data block and the target rainfall pattern information are input into the flood peak prediction model to obtain the maximum water depth prediction data block corresponding to the terrain data block; Based on the spatial relationship of multiple maximum water depth prediction data blocks, the multiple maximum water depth prediction data blocks are reconstructed to obtain the spatial distribution of the maximum water depth.

[0031] For any one of multiple terrain data blocks, input the terrain data block along with the target rainfall pattern information into the flood peak prediction model to obtain the corresponding maximum water depth prediction data block. This maximum water depth prediction data block corresponds to the terrain data block in spatial range, and the output resolution can be the same as the input, or it can be restored to the target resolution through internal model upsampling.

[0032] Since each terrain data block corresponds to a different spatial sub-region within the target area, multiple maximum water depth prediction data blocks have a defined spatial relationship. Based on this spatial relationship, the multiple maximum water depth prediction data blocks are reconstructed, that is, each prediction data block is placed back into its corresponding spatial location, and overlapping / boundary areas are processed to finally obtain the spatial distribution of maximum water depth covering the target area.

[0033] This embodiment divides the terrain elevation data into blocks and infers from each block, then reconstructs the spatial distribution of the maximum water depth based on the spatial relationship. This embodiment can reduce the scale of a single model input without changing the prediction target (the spatial distribution of the maximum water depth), improve the processability and deployment flexibility of large-scale or high-resolution data, and facilitate parallel computing and engineering implementation.

[0034] In this embodiment, the process of reconstructing multiple maximum water depth prediction data blocks based on their spatial relationships to obtain the spatial distribution of the maximum water depth includes: Determine the spatial relationship between the plurality of maximum water depth prediction data blocks in the target area; According to the spatial relationship, the multiple maximum water depth prediction data blocks are mapped to a unified grid position in the target area to obtain multiple mapped water depth layers; For any grid position in the unified grid positions, determine one or more mapped water depth layers covering the grid position, and obtain one or more water depth candidate values ​​corresponding to the grid position; A fusion process is performed on the one or more candidate water depth values ​​to obtain the reconstructed water depth value at the grid location; wherein the fusion process includes mean fusion, weighted fusion, and / or median fusion; The spatial distribution of the maximum water depth is obtained based on the reconstructed water depth values ​​of each grid position in the unified grid position.

[0035] Based on spatial relationships, multiple maximum depth prediction data blocks are mapped to a unified raster location in the target area, resulting in multiple mapped depth layers. A "mapped depth layer" refers to the overlay layer formed after a prediction data block is backfilled into a unified raster coordinate system according to its spatial location. When overlapping areas exist, the same raster location may be covered by multiple mapped depth layers, thus creating multiple candidate depth value sources.

[0036] For any grid location within a unified grid, determine one or more mapped depth layers covering that grid location, and obtain one or more candidate depth values ​​corresponding to that grid location. If the grid location is covered by only one mapped depth layer, the candidate value is single; if it is covered by multiple mapped depth layers, the candidate value is multiple.

[0037] A fusion process is performed on one or more candidate water depth values ​​to obtain the reconstructed water depth value for that grid location. The fusion process can employ methods such as: mean fusion (averaging candidate values), weighted fusion (weighted summation based on preset weights or weights determined by factors such as distance from the block center), and median fusion (taking the median of candidate values). After calculating the reconstructed water depth values ​​for all grid locations, a maximum water depth spatial distribution is formed based on the reconstructed water depth values ​​of each grid location within a unified grid location.

[0038] This embodiment effectively addresses overlapping coverage and boundary discontinuities during the reconstruction phase of block inference by mapping predicted data blocks to a unified grid and fusing candidate water depth values ​​in overlapping areas. This improves the consistency and usability of the reconstructed maximum water depth spatial distribution in terms of spatial alignment and numerical continuity.

[0039] In this embodiment, the generation network introduces a terrain attention mechanism during the generation of the maximum water depth spatial distribution, including: Attention weights are generated based on the terrain features; The terrain features are weighted using the attention weights to obtain terrain enhancement features; Based on the terrain enhancement features, the spatial distribution of the maximum water depth is generated.

[0040] This implementation presents the implementation of a terrain attention mechanism in a generative network. The terrain attention mechanism takes terrain features as input and generates attention weights related to those features. These attention weights can be spatial weight maps, channel weight vectors, or a combination of both, used to characterize the differences in contribution of different locations / channels within the terrain features to the generation of maximum water depth.

[0041] After generating attention weights based on terrain features, these weights are used to weight the terrain features, resulting in terrain enhancement features. Weighting can be achieved using element-wise multiplication, such as multiplying spatial attention weights pixel-by-pixel with the terrain features, or multiplying channel attention weights channel-by-channel with the terrain features. These terrain enhancement features serve as input to subsequent layers of the generative network, participating in the generation of the maximum water depth spatial distribution.

[0042] During the generation process of the network, terrain enhancement features can be fused with intermediate features in the network, or they can work together with rainfall condition features in the generation process, thereby enhancing key terrain information while maintaining the terrain representation.

[0043] This embodiment introduces a terrain attention mechanism into the generator network and weights terrain features to form terrain enhancement features. This application can make differentiated use of terrain information when generating the maximum water depth spatial distribution, enhance the ability of the generation process to express the spatial heterogeneity of terrain, and thus improve the matching degree between the peak water depth distribution and terrain conditions and the generation stability.

[0044] In this embodiment, generating attention weights based on the terrain features includes: The terrain features are subjected to average pooling in the channel dimension to obtain the first converged feature; Max pooling is performed on the terrain features in the channel dimension to obtain the second converged feature; The attention weights are generated based on the first convergence feature and the second convergence feature.

[0045] Average pooling is performed on the terrain features in the channel dimension to obtain the first converged feature, which reflects the average response intensity of each spatial location in the channel dimension; mean pooling is performed on the terrain features in the channel dimension to obtain the second converged feature, which reflects the significant response intensity of each spatial location in the channel dimension.

[0046] Attention weights are generated based on the first and second convergent features. Specifically, the two types of convergent features can be concatenated or added together, then input into a convolutional mapping layer or a fully connected mapping layer, and range-controlled attention weights are obtained through normalized activation (e.g., sigmoid). The generated attention weights are then used in the terrain feature weighting step.

[0047] In implementation, the converged features obtained by average pooling and max pooling can also be independently mapped and then fused to generate attention weights, in order to adapt to different network structure designs.

[0048] This embodiment generates attention weights by simultaneously utilizing two types of convergence features obtained from average pooling and max pooling. This application can take into account both overall response and significant response information in the calculation of attention weights, so that the obtained attention weights can more comprehensively reflect the different response forms of terrain features, thereby supporting a more reasonable terrain feature weighting process.

[0049] In this embodiment, the generating network introduces a multi-scale rainfall coding mechanism during the generation of the maximum water depth spatial distribution, including: The target rainfall pattern information is encoded to obtain rainfall condition features at least two different scales; The rainfall condition features are injected into at least two feature layers of different scales in the generation network, and the spatial distribution of maximum water depth is generated based on the generation process after the rainfall condition features are injected.

[0050] This implementation elaborates on the multi-scale rainfall encoding mechanism in the generative network. First, the target rainfall model information is encoded to obtain rainfall condition features at least two different scales. The encoding can employ embedding mapping, fully connected layer concatenation, or multi-branch mapping structures, so that the same rainfall model information is mapped to condition feature representations at different levels / scales.

[0051] Rainfall condition features are injected into at least two feature layers of different scales in the generative network. These different scale feature layers can correspond to different resolution levels of the generative network, such as different layers in the encoder and decoder, or different jumper layers in a U-shaped structure. The injection can be implemented by: expanding the rainfall condition features to match the spatial size of the target feature layer and then concatenating the channels; or mapping the rainfall condition features to additive modulation parameters of the feature layer to achieve affine modulation; or using a cross-attention structure to fuse the rainfall condition features with intermediate features.

[0052] After conditional injection is completed in at least two feature layers of different scales, the generation network integrates terrain enhancement features and multi-scale rainfall condition features in the subsequent generation process to output the spatial distribution of maximum water depth, thereby realizing conditional peak generation under different rainfall scenarios.

[0053] By encoding the target rainfall pattern information into multi-scale rainfall condition features and injecting them into different scale feature layers of the generation network, this application can achieve cross-scale rainfall condition constraints during the generation process, making the generation network's response to different rainfall scenarios more hierarchical and consistent, thereby improving the adaptability and stability of the generation of maximum water depth spatial distribution under multiple scenario inputs.

[0054] In this embodiment, the flood peak prediction model is obtained by training a conditional generative adversarial network (GAN), and the training of the GAN includes: For each training sample, the terrain elevation training data and rainfall pattern information corresponding to the training sample are input into the generation network, and the maximum water depth prediction result is output. The maximum water depth prediction result and the corresponding sample maximum water depth data are input into the discriminant network so that the discriminant network can learn to discriminate between the maximum water depth prediction result and the sample maximum water depth data. The discrimination network is configured as a patch discrimination network, which divides the maximum water depth data into multiple local patches and outputs local discrimination results for each patch; the maximum water depth data includes the maximum water depth prediction results and the sample maximum water depth data; The local discrimination results based on multiple local patches are used to generate a discrimination output for the maximum water depth data, thereby completing the adversarial training of the conditional generative adversarial network.

[0055] In this embodiment, the training of the conditional generative adversarial network further includes: The discriminant network is used to output the rainfall pattern regression results corresponding to the training samples; A regression loss is constructed based on the regression results of the rainfall patterns and the rainfall pattern information corresponding to the training samples; The parameters of the generator network and / or the discriminator network are updated based on the regression loss to introduce rainfall pattern regression constraints to complete the training of the conditional generative adversarial network.

[0056] In one embodiment, a method for intelligent forecasting of rainfall-induced flood peaks based on generative adversarial networks (GANs) is provided. This embodiment innovatively combines conditional generative networks with rainfall-induced flood prediction. The problem of predicting and regressing the maximum water depth at various points in the watershed, inputting rainfall patterns and topographic DEM maps, is transformed into an image translation problem. This allows for the generation of a maximum water depth prediction map based on an advanced GAN. The learning objective is the result of numerical simulation based on a two-dimensional cellular automata model under corresponding rainfall and watershed topographic inputs. Specifically, the simulation results of the two-dimensional cellular automata model CADDIES are used as the data driver for the deep learning network, i.e., the objective value. The study investigates the maximum water depth results for different watersheds under different rainfall patterns. The topographic elevation DEM data and the maximum water depth simulation results are both 1-meter resolution ASCII raster data. Because this patent focuses on the maximum water depth prediction problem under specific watershed and different rainfall input conditions, it does not involve generalization and transfer problems between watersheds; therefore, it falls within the scope of single-watershed supervised learning.

[0057] like Figure 2As shown, the topographic and water level data preprocessing module uses flood simulation (maximum water depth) datasets under different rainfall conditions in different watersheds as the raw data. Since watershed topographic images are usually large in size, they need to be randomly cut into smaller pieces to improve efficiency, which are much smaller than the size of the entire watershed map. This represents the number of channels corresponding to each terrain patch. Each terrain input patch is "image translated" into its corresponding water depth estimation patch. Since only the maximum water depth within this stage is ultimately output, the output water depth estimation patch has 1 channel. Finally, all predicted water depth estimation patches are collected. For water depth estimation patches with overlapping parts, merging methods such as averaging are used for processing. Finally, the maximum water depth map of the entire watershed under the corresponding rainfall pattern is constructed and output.

[0058] This embodiment presents a deep flood adversarial generative network (GDN) framework. It applies the Pix2Pix network structure to the problem of predicting the maximum flood peak caused by rainfall. This framework is highly suitable for real-world scenarios where only the maximum water depth of the basin needs to be considered, without delving into the changes in hydrological elements during the evolution process. Examples include designing urban pipeline layouts and constructing flood control systems, where the maximum water depth under different rainfall intensities needs to be considered to anticipate extreme disasters and improve the system's response capabilities. Therefore, this embodiment reduces the complex process of iteratively solving a series of equations using traditional hydrodynamic models, or the high-cost work of discrete iterative state updates using cellular automata to simulate complex physical phenomena, to a simple image pair translation task with a specified image input and output. The proposed deep flood adversarial generative network not only meets the needs of such real-world scenarios but also combines with already mature generative networks to achieve intelligent prediction of rainfall-induced flood peaks.

[0059] The multi-layered terrain attention module shows that the model’s attention to terrain features of different sizes and locations varies spatially. That is, within a watershed, due to the different surface properties at each location, the hydrological elements of low-lying areas and flat areas show significantly different trends. Therefore, different spatial regions require the model to give different attention.

[0060] Specifically, firstly, average pooling (AvgPool) and max pooling (MaxPool) operations are performed on each terrain feature map along the channel dimension. That is, the maximum and average values ​​of the input terrain features are calculated along the channel dimension. Each operation targets feature values ​​at the same spatial location in different channels. Max pooling can obtain the most salient activated parts of the features, while average pooling can smooth the features and obtain global information. The combination of these two different pooling methods can enhance each other.

[0061] As shown in formula (1.1) (1.1) like Figure 3As shown, the multi-scale rainfall coding module treats rainfall as a type of "style information" in the problem of predicting rainfall-induced flood peaks, influencing the "appearance" of the final depth map distribution. In the rainfall-induced flood problem, rainfall pattern vectors not only guide the maximum depth at local locations but also the global depth distribution. Therefore, introducing multi-level rainfall "style" information from coarse to fine granular levels during the generation process can more accurately guide the generation of maximum depth estimation maps under different return periods of rainfall. In summary, this embodiment proposes a multi-scale rainfall coding module to fully explore the effect of rainfall input on the final depth, from global (coarse-grained) to local (fine-grained), helping the model understand the response changes of water level under different rainfall patterns, thereby making the prediction results more accurate and improving the ability to identify and respond to extreme rainfall conditions.

[0062] The rainfall pattern regression module, in its original Generative Adversarial Network (GAN) for the "image translation" task, only performs a single style transfer, specifying the input and output domains. If image translation needs to be performed in a different domain, a new GAN needs to be trained. This is problematic for tasks involving image translation. Migration between domains requires generating ( 1) Multiple Generators. In real-world applications, to simulate potential extreme weather disasters, a large number of different rainfall patterns are needed for testing. Even slight numerical variations between rainfall patterns can lead to differences in the final maximum water depth map. Therefore, repeatedly generating multiple generators is impractical for rainfall-induced flooding problems. To facilitate transformation between multiple domains, rainfall patterns can also be viewed as labels. This section adopts the idea of ​​adversarial network generative game theory to improve the discriminator and generator's recognition of different "rainfall labels." It proposes a rainfall regression loss as an additional form of supervision besides adversarial loss, assisting the game process between the discriminator and generator.

[0063] Formula 1.2 is the regression loss for the discriminator, and Formula 1.3 is the regression loss for the generator.

[0064] (1.2) (1.3) Adversarial generation and discrimination module, generator Water depth patches are predicted using given terrain patches and rainfall pattern vectors. Inspired by Pix2Pix, the discriminator in this embodiment... The PatchGAN approach is used, where the ground truth depth estimate block or the depth estimate block generated by the generator before the discriminator is divided into several sub-blocks smaller than the original depth block. Each sub-block corresponds to a region in the input image and has its own discriminative output. Compared to the original generative adversarial network discriminator, which directly maps the input features to a real number representing the true probability of the input, PatchGAN maps the input to a multi-dimensional matrix, where each element represents the true probability of a feature within its corresponding receptive field. The model ultimately determines the authenticity of the input by averaging the probability values ​​of all sub-block outputs. PatchGAN's segmentation followed by merging operation allows the model to focus on detailed information in the image. The discriminator uses a fully convolutional approach; as the convolutional layers deepen, the receptive field gradually increases, thus allowing for a larger receptive field space for each sub-block. Averaging all sub-blocks provides a global judgment, improving the reasonableness of the output discrimination.

[0065] Compared to existing methods using deep learning to accelerate the rapid forecasting of rainfall-induced flood peaks, this embodiment fully leverages the importance of external input conditions. Instead of directly concatenating rainfall and topographic features—two cross-domain features—it employs a multi-scale topographic attention mechanism and rainfall guidance, allowing regional topographic features and statistical rainfall characteristics to play their respective roles in the generation of flood level maps at different scales. This significantly utilizes external conditions, resulting in test results from the deep learning-trained model that better match simulation results. By combining deep learning with numerical simulation methods, models learned on the training set can be transferred to the test set without simulating the evolution process from scratch, thus accelerating forecasting speed and significantly improving forecast accuracy while maintaining timeliness.

[0066] like Figure 4As shown, the flood prediction method based on Generative Adversarial Networks (MAE) proposed in this embodiment exhibits significant advantages in technical performance. Quantitative evaluation shows that the method achieves the lowest prediction error and best prediction accuracy on both the Portugal and Zurich datasets, with an R² exceeding 80%, indicating high model reliability and excellent fitting performance. Compared with traditional deep learning methods, MAE reduces error by 50%-72%, improves R² by 20%-81%, and increases prediction accuracy for high-risk areas and the entire watershed by 20%-81% and 11%-28%, respectively. Compared with image translation methods, MAE reduces error by 73%-75%, improves R² by 79%-97%, and increases prediction accuracy for high-risk and the entire watershed by 106%-127% and 110%-111%, respectively. Furthermore, this method demonstrates excellent generalization ability and stable performance across different datasets, while the comparative methods show greater fluctuations. By introducing a multi-layered terrain attention mechanism and a multi-scale rainfall encoding module, the model effectively integrates terrain and rainfall features, improving feature utilization and prediction interpretability. Ablation experiments confirm that each module significantly enhances performance. In terms of time performance, this method requires only tens of seconds for inference, far exceeding the computational cost of several hours for traditional numerical simulation methods, achieving efficient and low-cost flood peak forecasting.

[0067] In one embodiment, a flood peak prediction device is provided, the device including a processor and a memory, the memory storing a computer program, which, when executed on the processor, causes the device to perform the flood peak prediction method.

[0068] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for predicting flood peak value, characterized in that, The method includes: Acquire topographic elevation data and rainfall pattern information for the target area; The topographic elevation data and the target rainfall pattern information are input into a pre-trained flood peak prediction model to obtain the flood peak prediction result for the target area. The flood peak prediction result includes the spatial distribution of the maximum water depth in the target area. The flood peak prediction model includes a generator network based on a conditional generative adversarial network. In the process of generating the spatial distribution of the maximum water depth, the generator network introduces a terrain attention mechanism and a multi-scale rainfall coding mechanism. The terrain attention mechanism is used to weight terrain features, and the multi-scale rainfall coding mechanism is used to inject the target rainfall pattern information into the generation process in a multi-scale conditional form. The terrain features are feature representations obtained by feature extraction from the terrain elevation data.

2. The method according to claim 1, characterized in that, The method further includes: The terrain elevation data is input into a terrain feature extraction network to obtain at least two layers of terrain feature representations at different scales; The terrain features are generated based on the terrain feature representations at at least two different scales.

3. The method according to claim 1, characterized in that, The step of inputting the topographic elevation data and the target rainfall pattern information into a pre-trained flood peak prediction model to obtain the flood peak prediction result for the target area includes: The terrain elevation data is divided into multiple terrain data blocks; For any one of the multiple terrain data blocks, the terrain data block and the target rainfall pattern information are input into the flood peak prediction model to obtain the maximum water depth prediction data block corresponding to the terrain data block; Based on the spatial relationship of multiple maximum water depth prediction data blocks, the multiple maximum water depth prediction data blocks are reconstructed to obtain the spatial distribution of the maximum water depth.

4. The method according to claim 3, characterized in that, The method of reconstructing the spatial distribution of the maximum water depth based on the spatial relationship of multiple maximum water depth prediction data blocks to obtain the maximum water depth spatial distribution includes: Determine the spatial relationship between the plurality of maximum water depth prediction data blocks in the target area; According to the spatial relationship, the multiple maximum water depth prediction data blocks are mapped to a unified grid position in the target area to obtain multiple mapped water depth layers; For any grid position in the unified grid positions, determine one or more mapped water depth layers covering the grid position, and obtain one or more water depth candidate values ​​corresponding to the grid position; A fusion process is performed on the one or more candidate water depth values ​​to obtain the reconstructed water depth value at the grid location; wherein the fusion process includes mean fusion, weighted fusion, and / or median fusion; The spatial distribution of the maximum water depth is obtained based on the reconstructed water depth values ​​of each grid position in the unified grid position.

5. The method according to claim 1, characterized in that, The generating network introduces a terrain attention mechanism during the generation of the maximum water depth spatial distribution, including: Attention weights are generated based on the terrain features; The terrain features are weighted using the attention weights to obtain terrain enhancement features; Based on the terrain enhancement features, the spatial distribution of the maximum water depth is generated.

6. The method according to claim 5, characterized in that, The generation of attention weights based on the terrain features includes: The terrain features are subjected to average pooling in the channel dimension to obtain the first converged feature; Max pooling is performed on the terrain features in the channel dimension to obtain the second converged feature; The attention weights are generated based on the first convergence feature and the second convergence feature.

7. The method according to claim 1, characterized in that, The generating network introduces a multi-scale rainfall coding mechanism in the process of generating the maximum water depth spatial distribution, including: The target rainfall pattern information is encoded to obtain rainfall condition features at least two different scales; The rainfall condition features are injected into at least two feature layers of different scales in the generation network, and the spatial distribution of maximum water depth is generated based on the generation process after the rainfall condition features are injected.

8. The method according to claim 1, characterized in that, The flood peak prediction model is obtained by training a conditional generative adversarial network (GAN), the training of which includes: For each training sample, the terrain elevation training data and rainfall pattern information corresponding to the training sample are input into the generation network, and the maximum water depth prediction result is output. The maximum water depth prediction result and the corresponding sample maximum water depth data are input into the discriminant network so that the discriminant network can learn to discriminate between the maximum water depth prediction result and the sample maximum water depth data. The discrimination network is configured as a patch discrimination network, which divides the maximum water depth data into multiple local patches and outputs local discrimination results for each patch; the maximum water depth data includes the maximum water depth prediction results and the sample maximum water depth data; The local discrimination results based on multiple local patches are used to generate a discrimination output for the maximum water depth data, thereby completing the adversarial training of the conditional generative adversarial network.

9. The method according to claim 8, characterized in that, The training of the conditional generative adversarial network also includes: The discriminant network is used to output the rainfall pattern regression results corresponding to the training samples; A regression loss is constructed based on the regression results of the rainfall patterns and the rainfall pattern information corresponding to the training samples; The parameters of the generator network and / or the discriminator network are updated based on the regression loss to introduce rainfall pattern regression constraints to complete the training of the conditional generative adversarial network.

10. A flood peak prediction device, characterized in that, The device includes a processor and a memory, wherein the memory stores a computer program that, when executed on the processor, causes the device to perform the flood peak prediction method according to any one of claims 1 to 9.

Images