A city flood simulation method based on a deep learning reconstructed image model

Abstract

The application discloses a kind of urban flood simulation methods based on deep learning reconstruction image model, the method comprises the following steps: step 1, obtain basic data of research area;Step 2, construct one-dimensional-two-dimensional water dynamic coupling model;Step 3, data preprocessing;Step 4, construct U-Net deep learning reconstruction image model;Step 5, the setting and training of U-Net deep learning reconstruction image model;Step 6, urban flood simulation.The method improves the response speed of urban flood simulation spatial resolution, and considers the simulation accuracy and the calculation efficiency;Realize that using limited training data, quickly and accurately obtain high-resolution floodwater simulation result, provide scientific support for urban flood disaster prevention and control.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of urban flood hydrology simulation, and particularly relates to a city flood simulation method based on a deep learning reconstructed image model. BACKGROUND

[0002] With the continuous acceleration of urbanization in China, urban flood disasters occur frequently, which poses a major threat to the safety of residents and the operation of cities. One of the effective means to cope with urban floods is to use hydrodynamic flood simulation to establish digital models and numerical calculation means to estimate the flood spread range and flow depth in a timely manner, and to provide scientific support for urban flood disaster prevention and control. However, the traditional high-resolution hydrodynamic simulation method relies on fine grid calculation, which can accurately reflect the flood process, but often requires a large amount of computing resources and a long running time, making it difficult to meet the real-time requirements of emergency warning. On the other hand, low-resolution simulation results can be processed by simple interpolation to improve the resolution, but it is difficult to reproduce details in complex terrain and local urban features, resulting in large flood depth prediction errors. SUMMARY

[0003] The purpose of the present application is to provide a city flood simulation method based on a deep learning reconstructed image model to solve the above technical problems.

[0004] To achieve the above purpose, the present application provides the following technical solutions:

[0005] The present application discloses a city flood simulation method based on a deep learning reconstructed image model, which comprises the following steps:

[0006] Step 1, obtaining basic data of the research area: obtaining basic data of the research area, including geographic digital elevation data, remote sensing image data, land use type data, drainage pipe network data, historical rainfall data and waterlogging monitoring data; and modifying the obtained geographic digital elevation data, remote sensing image data, land use type data and drainage pipe network data to the same coordinate system; in the same coordinate system, according to the remote sensing image data, frame the road and building contour as building contour data;

[0007] Step 2, constructing a one-dimensional-two-dimensional hydrodynamic coupling model: generating two high-precision grids and low-precision grids of different sizes in the flood accumulation grid generation tool in the study area, and constructing a one-dimensional-two-dimensional hydrodynamic coupling model using high-precision grids according to the obtained geographic digital elevation data, building contour data, land use type data, and drainage pipe network data; then, through historical rainfall data and waterlogging monitoring data, the one-dimensional-two-dimensional hydrodynamic coupling model simulation results are compared with the monitoring data to calibrate and verify; after the one-dimensional-two-dimensional hydrodynamic coupling model constructed by using high-precision grids is completed, the high-precision grids are replaced with low-precision grids, and other parameters remain unchanged, to obtain a one-dimensional-two-dimensional hydrodynamic coupling model of low-precision grids;

[0008] Step 3, data preprocessing: for the one-dimensional-two-dimensional hydrodynamic coupling model of high-precision and low-precision grids constructed, the multi-field design storm data generated according to the storm formula published by the local meteorological station of the study area and the historical rainfall data of the study area are input into the one-dimensional-two-dimensional hydrodynamic coupling model, and the floodwater distribution and peak water depth in the study area are simulated, then a plurality of urban floodwater simulation results of different rain types under multiple storm recurrence periods are obtained, including high-precision floodwater simulation results and low-precision floodwater simulation results, which are saved in the form of grid data; the low-precision floodwater simulation results are processed to obtain fine-size low-precision floodwater simulation results, and then the fine-size low-precision floodwater simulation results, high-precision floodwater simulation results, geographic digital elevation data, building contour data, and rainfall data are structured to store structured data; the rainfall data includes historical rainfall data and design storm data;

[0009] Step 4, constructing a U-Net deep learning reconstruction image model: constructing a U-Net deep learning reconstruction image model, which includes a rainfall data input layer, a hydrodynamic floodwater simulation result input layer, a geographic digital elevation data input layer, a building contour data input layer, a convolutional encoder, a rainfall feature reconstruction layer, a feature fusion layer, a convolutional decoder, a fully connected layer, and an output layer;

[0010] Step 5, setting and training of the U-Net deep learning reconstructed image model: structured data in step 3 is used to construct training samples, and the training samples are input into the U-Net deep learning reconstructed image model for training, so that the model can predict output data according to input data; the input data includes rainfall data, fine size low precision waterlogging simulation results, geographic digital elevation data and building contour data in the input period, and the output data is the peak value of waterlogging depth and waterlogging distribution in the input period corresponding to the sample area; the simulation results of the complete waterlogging depth peak value and waterlogging distribution of the study area are generated by recombining and splicing the samples; during the model training process, the model parameters are adjusted, and when the training round reaches the pre-set round or the validation loss tends to be stable and no longer decreases, the model training is completed;

[0011] Step 6, urban flood simulation: the low-precision waterlogging simulation results of the study area under the rainfall scenario to be predicted are obtained by a one-dimensional-two-dimensional hydrodynamic model, and the obtained low-precision waterlogging simulation results are input into the U-Net deep learning reconstructed image model which has been trained, and the high-precision floodwaterlogging simulation results of the study area under the corresponding rainfall scenario are output.

[0012] Further, the geographic digital elevation data in step 1 is elevation point data, and a rectangular grid digital elevation model DEM is constructed by interpolation in GIS software; the land use type is the underlying surface type of the study area, which is divided into building land, green land and traffic land; the drainage pipe network data includes the spatial distribution of the drainage pipe network, the length of the pipe, the cross-sectional shape, the pipe diameter, the distribution and depth of the inspection well, and the connection mode of the pipe and the inspection well; the historical rainfall data includes rainfall and rainfall duration; the historical waterlogging monitoring data includes floodwaterlogging distribution, waterlogging depth peak value, outlet flow and pipe network flow; wherein the waterlogging depth peak value refers to the maximum waterlogging depth in a given rainfall period.

[0013] Further, the high-precision grid in step 2 is defined as a grid division scheme with a spatial resolution of 10 meters or less, and the number of grids in the study area reaches more than 10,000 per square kilometer; otherwise, it is a low-precision grid.

[0014] Further, the specific process of the method of comparing the simulation results of the one-dimensional-two-dimensional hydrodynamic coupling model with the monitoring data for calibration and verification in step 2 is as follows: according to the input historical rainfall data, the one-dimensional-two-dimensional hydrodynamic coupling model simulates the simulation results corresponding to the input historical rainfall data, including floodwaterlogging distribution, waterlogging depth peak value, outlet flow and pipe network flow, and then compares the floodwaterlogging distribution, waterlogging depth peak value, outlet flow and pipe network flow of the simulation results and the monitoring data; the accuracy of the one-dimensional-two-dimensional hydrodynamic coupling model simulation is evaluated by using the Nash efficiency coefficient NSE, and the specific formula is as follows:

[0015]

[0016] In the formula: S i The simulated value at time i; O i The value monitored at time i; is the average value of the monitored values; n is the total number of monitored values; the range of NSE is (-∞, 1], and an NSE value close to 1 indicates that the simulation performance of the one-dimensional-two-dimensional hydrodynamic coupling model is better. When constructing the one-dimensional-two-dimensional hydrodynamic coupling model, a threshold is set. If it is lower than the set threshold, the model needs to be reconstructed.

[0017] Furthermore, the rainstorm formula mentioned in step 3 is:

[0018]

[0019] In the formula: i is the intensity of the rainstorm, in mm / min; P is the return period of the rainstorm, in years; t is the duration of rainfall, in min; parameters A, B, C, and n are empirical parameters, fitted according to the regional rainstorm statistics published by the local meteorological department.

[0020] Furthermore, the specific process of processing the low-precision water accumulation simulation results to obtain the fine-size low-precision water accumulation simulation results in step 3 is as follows: Nearest neighbor upsampling is performed on the low-precision water accumulation simulation results. Specifically, each pixel of the low-precision water accumulation simulation results is directly mapped to an equal-value sub-block region in the target image through spatial expansion, making its grid size the same as the high-precision water accumulation simulation results, thereby obtaining the fine-size low-precision water accumulation simulation results. The specific operation is as follows:

[0021] For any pixel I in the original data coarse (i, j), whose value is copied and filled into the corresponding s×s sub-block in the target data:

[0022] I fine (si+m,sj+n)=I coarse (i, j) (3)

[0023] In the formula: I coarse is the pixel value in the i-th row and j-th column of the original low-resolution data; s is the expansion factor; I fine This represents the pixel value in the (si+m)th row and (sj+n)th column of the high-resolution data, copied from the corresponding I. coarse (i, j); m and n are I coarse The original data size, and m, n∈{0,1,2,…,s-1}.

[0024] Furthermore, the specific process of structuring the fine-scale low-precision water accumulation simulation results, high-precision water accumulation simulation results, geographic digital elevation data, building outline data, and rainfall data in step 3 is as follows: extract the spatial index and attribute information of the raster cells, and organize them into structured data according to the preset field order.

[0025] Furthermore, in step 4, the rainfall data input layer is used to input vector data of time-series rainfall data. A fully connected layer follows the input layer, gradually expanding and reshaping the rainfall data into a three-dimensional tensor, which is then fused with other spatial features. The hydrodynamic water accumulation simulation result input layer is used to input representative, low-precision water accumulation simulation results. Geographic digital elevation data is input from the geographic digital elevation data input layer. Building outline data is input from the building outline data input layer. The hydrodynamic water accumulation simulation result input layer, geographic digital elevation input layer, and building outline data input layer are connected by a splicing layer to form a three-dimensional tensor, which serves as the input to the convolutional encoder. The convolutional encoder includes 12 convolutional layers, 3 pooling layers, and 3 deconvolutional layers, sequentially extracting the hydrodynamic simulation results, geographic digital elevation data, and building outline data. Specifically, it includes: a first convolutional layer, a second convolutional layer, a first pooling layer, a third convolutional layer, and a fourth convolutional layer. The system consists of a convolutional layer, a second pooling layer, a fifth convolutional layer, a sixth convolutional layer, and a third pooling layer. The rainfall feature reconstruction layer comprises multiple fully connected layers that progressively expand the rainfall vector into a tensor matching the output size of the convolutional encoder and concatenate it with the bottleneck layer features of the convolutional encoder. The feature fusion layer is used to fuse the spatial features output by the convolutional encoder with the expanded rainfall information tensor along the channel dimension to form a comprehensive feature representation. The convolutional decoder includes deconvolutional layers and convolutional layers, which progressively improve spatial resolution through layer-by-layer upsampling and feature recovery. Specifically, it includes: a first deconvolutional layer, a seventh convolutional layer, an eighth convolutional layer, a second deconvolutional layer, a ninth convolutional layer, a tenth convolutional layer, a third deconvolutional layer, an eleventh convolutional layer, and a twelfth convolutional layer. The output layer maps the features of the last layer of the decoder to a single-channel output through a linear convolution, which is used to predict the flood distribution and peak water depth of the study area.

[0026] Add a Dropout layer to the bottleneck layer of the convolutional encoder to prevent overfitting during training and improve the model's generalization ability.

[0027] An activation function, specifically the ReLU function, is connected after each convolutional layer to enhance nonlinear fitting capabilities. The specific mathematical expression for the ReLU function is as follows:

[0028] R (x) =max(0,x) (4)

[0029] In the formula: R (x)The value represents the output of the activation function; x represents the input value; when x is greater than 0, the output is x; when x is less than or equal to 0, the output is 0.

[0030] Furthermore, the specific process of constructing training samples from the structured data in step 3, as described in step 5, is as follows: Considering hardware performance and the number of research samples, a slider slicing method is used to divide the research area into several small samples. The data information contained in each small sample is consistent with the corresponding location data information in the original sample. Specifically, in the script, the fine-size low-precision water accumulation simulation results, geographic digital elevation data, and building outline data are divided into several small samples in groups according to a 128×128 rectangle and a stride of 16. The 128×128×1 small samples generated in each group are stitched together into 128×128×3 small samples through a stitching layer. Then, through data augmentation, the small samples are randomly rotated by 90°, rotated by 180°, and mirrored to obtain more small samples, thus forming secondary construction samples, which are then used in subsequent processes.

[0031] Furthermore, the specific process of adjusting the model parameters in step 5 is as follows: using the weighted mean squared error Loss as the loss function of the U-Net deep learning image reconstruction model, and adjusting the model parameters based on the fine-grid flood prediction results generated during model training. The mathematical expression of the loss function is:

[0032] Loss = Mean[(y pred -y) 2 ×(1+α·y)×(1+β·(1-y DEM (5)

[0033] In the formula: y is the peak value of the actual water depth on the grid; y pred The peak value of the predicted water depth on the grid; α and β are hyperparameters used to adjust the weight magnitude; y DEM The value is the normalized geographic digital elevation of the grid, with a higher weight for lower terrain; Mean is the average of all pixels in the entire batch to obtain the overall loss value.

[0034] The beneficial effects of this invention are: the method described in this invention improves the response speed in terms of spatial resolution for urban flood simulation, while taking into account both simulation accuracy and computational efficiency; it enables the rapid and accurate acquisition of high-resolution flood water accumulation simulation results using only limited training data, providing scientific support for urban flood disaster prevention and control.

[0035] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0036] Figure 1This is a schematic diagram of the method flow described in this invention;

[0037] Figure 2 This is a schematic diagram of the process of constructing a U-Net deep learning image reconstruction model in Example 1. Detailed Implementation

[0038] This invention discloses a method for simulating urban flooding based on deep learning-reconstructed image models, such as... Figure 1 As shown, the method includes the following steps:

[0039] Step 1: Obtain basic data for the study area:

[0040] Basic data for the study area were obtained from relevant local departments, including geographic digital elevation data, remote sensing image data, land use type data, drainage network data, historical rainfall data, and urban flooding monitoring data. The obtained geographic digital elevation data, remote sensing image data, land use type data, and drainage network data were modified to the same coordinate system. Under the same coordinate system, the outlines of roads and buildings were roughly drawn based on the remote sensing image data as building outline data.

[0041] Among them, the geographic digital elevation data consists of elevation point data, which is constructed using interpolation in GIS software to create a rectangular grid digital elevation model (DEM); the land use type refers to the underlying surface type of the study area, generally divided into building land, green space, and transportation land; the drainage network data includes the spatial distribution of the drainage network, pipe length, cross-sectional shape, pipe diameter, distribution and depth of manholes, and the connection method between pipes and manholes; historical rainfall data includes rainfall amount and duration; historical urban flooding monitoring data includes flood water distribution, peak water depth, outflow rate, and network flow rate; among them, the peak water depth refers to the maximum water depth within a given rainfall period, rather than the water depth at a specific moment.

[0042] Step 2: Construct a one-dimensional-two-dimensional hydrodynamic coupling model:

[0043] The flood water accumulation grid generation tool generates two types of two-dimensional grids with significant size differences for the study area: high-precision grids and low-precision grids. The grids are manually checked to prevent over-stretching or sharpening. High-precision grids are defined according to the Basic Geographic Information Data Quality Requirements and Assessment (GB / T 41149-2021), characterized by a spatial resolution of 10 meters or less and a grid density of at least 10,000 grids per square kilometer within the study area. Grids that do not meet the high-precision grid standard are considered low-precision grids.

[0044] Based on the acquired geographic digital elevation data, building outline data, land use type data, and drainage network data, a one-dimensional-two-dimensional hydrodynamic coupling model is constructed using a high-precision grid.

[0045] Then, using historical rainfall data and urban flooding monitoring data, the simulation results of the one-dimensional-two-dimensional hydrodynamic coupling model were compared with the monitoring data for calibration and verification to ensure the rationality of the one-dimensional-two-dimensional hydrodynamic coupling model. The specific process is as follows: Based on the input historical rainfall data, the one-dimensional-two-dimensional hydrodynamic coupling model generates simulation results corresponding to the input historical rainfall data, including flood distribution, peak water depth, outlet flow, and pipe network flow. Then, the simulation results are compared with the monitoring data for flood distribution, peak water depth, outlet flow, and pipe network flow. The Nash efficiency coefficient (NSE) is used to evaluate the accuracy of the one-dimensional-two-dimensional hydrodynamic coupling model simulation. The specific formula is as follows:

[0046]

[0047] In the formula: S i The simulated value at time i; O i The value monitored at time i; is the average value of the monitored values; n is the total number of monitored values; the value range of NSE is (-∞, 1]. An NSE value close to 1 indicates that the simulation performance of the one-dimensional-two-dimensional hydrodynamic coupling model is better. When constructing the one-dimensional-two-dimensional hydrodynamic coupling model, a threshold is set (generally lower than 0.75). If it is lower than the set threshold, the model needs to be reconstructed.

[0048] After constructing the one-dimensional-two-dimensional hydrodynamic coupling model using a high-precision mesh, the high-precision mesh is replaced with a low-precision mesh, while other parameters remain unchanged, to obtain the one-dimensional-two-dimensional hydrodynamic coupling model using a low-precision mesh.

[0049] Step 3, Data Preprocessing:

[0050] For the constructed one-dimensional-two-dimensional hydrodynamic coupling model with high and low precision grids, multi-field design storm data conforming to the rainfall characteristics of the study area, generated based on the storm formula published by the local meteorological station, and historical rainfall data of the study area are input into the one-dimensional-two-dimensional hydrodynamic coupling model to simulate the distribution and peak depth of floodwater in the study area. This yields several sets of urban floodwater simulation results for different rainfall patterns under multiple storm return periods, including high-precision and low-precision simulation results, which are saved in raster data form. Different rainfall patterns include single-peak pre-peak, single-peak mid-peak, single-peak post-peak, double-peak pre-peak, double-peak mid-peak, double-peak post-peak, uniform, wave, Chicago I, Chicago II, and Chicago III, etc. The storm formula is:

[0051]

[0052] In the formula: i is the intensity of the rainstorm, in mm / min; P is the return period of the rainstorm, in years; t is the duration of rainfall, in min; parameters A, B, C, and n are empirical parameters, fitted according to the regional rainstorm statistics published by the local meteorological department.

[0053] The process involves nearest-neighbor upsampling of the low-precision grid urban flooding simulation results. Specifically, each pixel in the low-precision simulation result is directly mapped to an equal-value sub-block region in the target image through spatial expansion, making its grid size the same as that of the high-precision grid urban flooding simulation results. This yields a fine-sized low-precision flooding simulation result. The specific steps are as follows:

[0054] For any pixel I in the original data coarse (i, j), whose value is copied and filled into the corresponding s×s sub-block in the target data:

[0055] I fine (si+m,sj+n)=I coarse (i, j) (3)

[0056] In the formula: I coarse is the pixel value in the i-th row and j-th column of the original low-resolution data; s is the expansion factor; I fine This represents the pixel value in the (si+m)th row and (sj+n)th column of the high-resolution data, copied from the corresponding I. coarse (i, j); m and n are I coarse The original data size, and m, n∈{0,1,2,…,s-1}.

[0057] For raster cell data such as flood and waterlogging simulation results, geographic digital elevation data, building outline data, and rainfall data (including historical rainfall data and design storm data), a pre-defined program is used for structured processing. Specifically, this includes extracting the spatial index and attribute information of the raster cells, organizing them into structured data according to a preset field order, and finally outputting and storing them as structured files for subsequent applications. Each data set includes low-precision, fine-scale flood and waterlogging simulation results, high-precision raster elevation data, building outline data, and rainfall data, all stored in structured data format.

[0058] Step 4: Construct the U-Net deep learning image reconstruction model:

[0059] A U-Net deep learning image reconstruction model is constructed, which includes a rainfall data input layer, a hydrodynamic water accumulation simulation result input layer, a geographic digital elevation data input layer, a building outline data input layer, a convolutional encoder, a rainfall feature reconstruction layer, a feature fusion layer, a convolutional decoder, a fully connected layer, and an output layer.

[0060] The rainfall data input layer is used to input vector data of time-series rainfall data. After the input layer, a fully connected layer is connected to gradually expand and reshape the rainfall data into a three-dimensional tensor, which is then fused with other spatial features. The hydrodynamic water accumulation simulation result input layer is used to input the simulation results of fine-scale, low-precision water accumulation. The geographic digital elevation data is input by the geographic digital elevation data input layer. The building outline data is input by the building outline data input layer. The hydrodynamic water accumulation simulation result input layer, the geographic digital elevation input layer, and the building outline data input layer are connected by a concatenation layer to form a three-dimensional tensor, which serves as the input to the convolutional encoder. The convolutional encoder comprises 12 convolutional layers, 3 pooling layers, and 3 deconvolutional layers, sequentially extracting hydrodynamic simulation results, geographic digital elevation data, and building outline data. Specifically, it includes: the first convolutional layer, the second convolutional layer, the first pooling layer, the third convolutional layer, the fourth convolutional layer, the second pooling layer, the fifth convolutional layer, the sixth convolutional layer, and the third pooling layer. The rainfall feature reconstruction layer comprises multiple fully connected layers, progressively expanding the rainfall vector into a tensor matching the output size of the convolutional encoder, and concatenating it with the bottleneck layer features of the convolutional encoder. The feature fusion layer is used to fuse the spatial features output by the convolutional encoder with the expanded rainfall information tensor along the channel dimension to form a comprehensive feature representation. The convolutional decoder comprises deconvolutional layers and convolutional layers, progressively improving spatial resolution through layer-by-layer upsampling and feature recovery. Specifically, it includes: the first deconvolutional layer, the seventh convolutional layer, the eighth convolutional layer, the second deconvolutional layer, the ninth convolutional layer, the tenth convolutional layer, the third deconvolutional layer, the eleventh convolutional layer, and the twelfth convolutional layer.

[0061] The output layer maps the features from the last layer of the decoder to a single-channel output through a linear convolution, which is used to predict the distribution of floodwater and the peak depth of floodwater in the study area.

[0062] Adding a Dropout layer to the bottleneck layer of the convolutional encoder prevents overfitting during training and improves the model's generalization ability.

[0063] An activation function, specifically the ReLU function, is connected after each convolutional layer to enhance nonlinear fitting capabilities. The specific mathematical expression for the ReLU function is as follows:

[0064] R (x) =max(0,x) (4)

[0065] In the formula: R (x)The value represents the output of the activation function; x represents the input value; when x is greater than 0, the output is x; when x is less than or equal to 0, the output is 0.

[0066] Step 5: Setting up and training the U-Net deep learning image reconstruction model:

[0067] Training samples are constructed using the structured data from step 3, and then input into the U-Net deep learning reconstructed image model in groups for training. The specific process for constructing training samples is as follows: Considering hardware performance and the number of research samples, a slider slicing method is used to divide the research area into several small samples. The data information contained in each small sample is consistent with the corresponding location data information in the original sample. Specifically, in the script, the fine-scale low-precision water accumulation simulation results, geographic digital elevation data, and building outline data are divided into several small samples in groups of 128×128 rectangles with a stride of 16. Each group of 128×128×1 small samples is then stitched together into 128×128×3 small samples through a stitching layer. Then, through data augmentation, the small samples are randomly rotated by 90°, rotated by 180°, and mirrored to obtain more small samples, thus forming secondary constructed samples, which are then input into subsequent processes.

[0068] The reconstructed samples are used as training samples and input into the U-Net deep learning image reconstruction model for training and testing, enabling the model to predict output data based on the input data. Input data includes water accumulation and topography-related tensors (including fine-scale, low-precision water accumulation simulation results and geographic digital elevation data), building outline tensors (including building outline data), and rainfall sequence tensors (including historical rainfall data and designed storm data). Output data is a water accumulation depth tensor (including peak water accumulation depth and water accumulation distribution). The samples are then reassembled and stitched together to generate complete simulation results of peak water accumulation depth and water accumulation distribution for the study area.

[0069] In the input data, the building outline data is processed with edge enhancement, and the rainfall sequence data is used to describe the rainfall changes within the input time period; the output data represents the peak water depth and water distribution of the corresponding sample area within the input time period.

[0070] We use the weighted mean squared error (MSE) as the loss function for the U-Net deep learning image reconstruction model. The model parameters are adjusted based on the fine-grid flood prediction results generated during model training. The mathematical expression of the loss function is:

[0071] Loss = Mean[(y pred -y) 2 ×(1+α·y)×(1+β·(1-y DEM (5)

[0072] In the formula: y is the peak value of the actual water depth on the grid; y pred The peak value of the predicted water depth on the grid; α and β are hyperparameters used to adjust the weight magnitude; y DEM The value is the normalized geographic digital elevation of the grid, with a higher weight for lower terrain; Mean is the average of all pixels in the entire batch to obtain the overall loss value.

[0073] When the training rounds reach the preset number of rounds, or when the validation loss stabilizes and no longer decreases, the model training is complete, and the model parameters are stored in a file with the ".h5" suffix for inference and prediction.

[0074] Step 6: Urban Flood Simulation

[0075] Low-precision waterlogging simulation results for the required rainfall scenarios in the study area are obtained through a one-dimensional-two-dimensional hydrodynamic model. The obtained low-precision waterlogging simulation results are used as input data to train the U-Net deep learning reconstructed image model, and high-precision flood and waterlogging simulation results for the corresponding rainfall scenarios in the study area are output.

[0076] Example 1

[0077] This embodiment is an application example of the above method.

[0078] This embodiment discloses a method for simulating urban flooding based on deep learning-reconstructed image models, the steps of which are as follows:

[0079] Step 1: Obtain basic data for the study area:

[0080] This case study selects the Minzhi area of ​​Shenzhen as the research area. This area has a high degree of urbanization, dense buildings, and complex transportation facilities, and has typical flood risk characteristics. Moreover, the geographic information data and rainfall and meteorological data are relatively complete, providing ideal conditions for carrying out urban flood simulation and rapid prediction research.

[0081] Basic data for the study area were obtained from relevant local departments, including geographic digital elevation data, remote sensing image data, land use type data, drainage network data, historical rainfall data, and urban flooding monitoring data. The obtained geographic digital elevation data, remote sensing image data, land use type data, and drainage network data were modified to the same coordinate system. Under the same coordinate system, the outlines of roads and buildings were roughly drawn based on the remote sensing image data as building outline data.

[0082] Among them, the geographic digital elevation data is elevation point data, which is constructed into a rectangular grid digital elevation model (DEM) in GIS software using interpolation; the land use type is the underlying surface type of the study area, generally divided into building land, green space and transportation land; the drainage network data includes the spatial distribution of drainage network, pipe length, cross-sectional shape, pipe diameter, distribution and depth of manholes, and connection method between pipes and manholes; the historical rainfall data includes rainfall amount and rainfall duration; the historical waterlogging monitoring data includes the distribution of floodwater, peak water depth, outflow rate and network flow rate.

[0083] Step 2: Construct a one-dimensional-two-dimensional hydrodynamic coupling model:

[0084] In the flood water accumulation grid generation tool, high-precision grids (10m) and low-precision grids (200m) are generated for the study area, and the grids are manually checked to avoid overstretching or sharpening.

[0085] Based on the acquired geographic digital elevation data, building outline data, land use type data, and drainage network data, a one-dimensional-two-dimensional hydrodynamic coupling model is constructed using a high-precision grid.

[0086] Then, using historical rainfall data and urban flooding monitoring data, the simulation results of the one-dimensional-two-dimensional hydrodynamic coupling model were compared with the monitoring data for calibration and verification to ensure the rationality of the one-dimensional-two-dimensional hydrodynamic coupling model. The specific process is as follows: Based on the input historical rainfall data, the one-dimensional-two-dimensional hydrodynamic coupling model generates simulation results corresponding to the input historical rainfall data, including flood distribution, peak water depth, outlet flow, and pipe network flow. Then, the simulation results are compared with the monitoring data for flood distribution, peak water depth, outlet flow, and pipe network flow. The Nash efficiency coefficient (NSE) is used to evaluate the accuracy of the one-dimensional-two-dimensional hydrodynamic coupling model simulation. The specific formula is as follows:

[0087]

[0088] In the formula: S i The simulated value at time i; O i The value monitored at time i; is the average value of the monitored values; n is the total number of monitored values; the NSE value ranges from (-∞, 1], and an NSE value close to 1 indicates better simulation performance of the one-dimensional-two-dimensional hydrodynamic coupling model. In this embodiment, the NSE threshold is set to 0.65; if it is lower than this threshold, the model needs to be rebuilt.

[0089] After constructing the one-dimensional-two-dimensional hydrodynamic coupling model using a high-precision mesh, the high-precision mesh is replaced with a low-precision mesh, while other parameters remain unchanged, to obtain the one-dimensional-two-dimensional hydrodynamic coupling model using a low-precision mesh.

[0090] The one-dimensional to two-dimensional hydrodynamic coupling model in this embodiment was constructed using IFMS / Urban software. For specific operating steps, please refer to the software's user manual.

[0091] Step 3, Data Preprocessing:

[0092] For the constructed one-dimensional-two-dimensional hydrodynamic coupling model with high and low precision grids, multiple design storm data that conform to the rainfall characteristics of the study area, generated according to the storm formula published by the local meteorological station, and historical rainfall data of the study area are input into the one-dimensional-two-dimensional hydrodynamic coupling model to simulate the distribution of floodwater and peak water depth in the study area. This yields several sets of urban floodwater simulation results for different rainfall types under multiple storm return periods, including high-precision and low-precision floodwater simulation results, which are saved in the form of raster data. The storm formula published by the local meteorological station in the study area in this embodiment is:

[0093]

[0094] In the formula: i is the intensity of the rainstorm, in mm / min; P is the return period of the rainstorm, in years; t is the duration of rainfall, in min.

[0095] The process involves nearest-neighbor upsampling of the low-precision grid urban flooding simulation results. Specifically, each pixel in the low-precision simulation result is directly mapped to an equal-value sub-block region in the target image through spatial expansion, making its grid size the same as that of the high-precision grid urban flooding simulation results. This yields a fine-sized low-precision flooding simulation result. The specific steps are as follows:

[0096] For any pixel I in the original data coarse (i, j), whose value is copied and filled into the corresponding s×s sub-block in the target data:

[0097] I fine (si+m,sj+n)=I coarse (i, j) (3)

[0098] In the formula: I coarse is the pixel value in the i-th row and j-th column of the original low-resolution data; s is the expansion factor; I fine This represents the pixel value in the (si+m)th row and (sj+n)th column of the high-resolution data, copied from the corresponding I. coarse (i, j); m and n are I coarse The original data size, and m, n∈{0,1,2,…,s-1}.

[0099] For raster cell data such as flood and waterlogging simulation results, geographic digital elevation data, building outline data, and rainfall data (including historical rainfall data and design storm data), a pre-defined program is used for structured processing. Specifically, this includes extracting the spatial index and attribute information of the raster cells, organizing them into structured data according to a preset field order, and finally outputting and storing them as structured files for subsequent applications. Each data set includes low-precision, fine-scale flood and waterlogging simulation results, high-precision raster elevation data, building outline data, and rainfall data, all stored in structured data format.

[0100] Step 4: Construct the U-Net deep learning image reconstruction model:

[0101] Constructing a U-Net deep learning image reconstruction model, such as Figure 2 As shown, the model includes a rainfall data input layer, a hydrodynamic water accumulation simulation result input layer, a geographic digital elevation data input layer, a building outline data input layer, a convolutional encoder, a rainfall feature reconstruction layer, a feature fusion layer, a convolutional decoder, a fully connected layer, and an output layer.

[0102] The rainfall data input layer is used to input vector data representing time-series rainfall data, with a size of 1×25. The hydrodynamic water accumulation simulation result input layer is used to input small-scale, low-precision water accumulation simulation results, with a size of 128×128×1. Geographic digital elevation data is input from the geographic digital elevation data input layer, with a size of 128×128×1. Building outline data is input from the building outline input layer. The hydrodynamic water accumulation simulation result input layer, geographic digital elevation data input layer, and building outline data input layer are connected by a concatenation layer to form a tensor with an input size of 128×128×3, which serves as the input to the convolutional encoder. The rainfall data input layer is processed by a fully connected layer, which consists of three Dense layers (output dimensions of 128, 256, and 512, respectively). Finally, after being expanded to 256 nodes by the Dense layer, it is reconstructed into a tensor with a size of 16×16×1 for fusion with spatial features.

[0103] The convolutional encoder includes, in sequence:

[0104] The first convolutional layer (Conv1) takes a 128×128×3 tensor as input and outputs a 128×128×32 tensor.

[0105] The second convolutional layer (Conv2) takes a 128×128×32 tensor as input and outputs a 128×128×32 tensor.

[0106] The first pooling layer (Maxpool1) takes a 128×128×32 tensor as input and outputs a 64×64×32 tensor.

[0107] The third convolutional layer (Conv3) takes a 64×64×32 tensor as input and outputs a 64×64×64 tensor.

[0108] The fourth convolutional layer (Conv4) takes a 64×64×64 tensor as input and outputs a 64×64×64 tensor.

[0109] The second pooling layer (Maxpool2) takes a 64×64×64 tensor as input and outputs a 32×32×64 tensor.

[0110] The fifth convolutional layer (Conv5) takes a 32×32×64 tensor as input and outputs a 32×32×128 tensor.

[0111] The sixth convolutional layer (Conv6) takes a 32×32×128 tensor as input and outputs a 32×32×128 tensor.

[0112] The third pooling layer (Maxpool3) takes a 32×32×128 tensor as input and outputs a 16×16×128 tensor.

[0113] The 16×16×1 tensor output by the rainfall feature reconstruction layer is connected to the 16×16×128 tensor output by the convolutional encoder through a concatenate layer and spliced ​​along the channel dimension to obtain a tensor of size 16×16×129.

[0114] The convolutional decoder part includes:

[0115] The first deconvolutional layer (Deconv1) takes a 16×16×129 tensor as input and outputs a 32×32×256 tensor.

[0116] Feature maps of the same scale as the encoder are skipped and then concatenated.

[0117] The seventh convolutional layer (Conv7) takes a 32×32×256 tensor as input and outputs a 32×32×128 tensor.

[0118] The eighth convolutional layer (Conv8) takes a 32×32×128 tensor as input and outputs a 32×32×128 tensor.

[0119] The second deconvolutional layer (Deconv2) takes a 32×32×128 tensor as input and outputs a 64×64×128 tensor.

[0120] After the skip connection, the ninth convolutional layer (Conv9) and the tenth convolutional layer (Conv10) are performed;

[0121] The third deconvolutional layer (Deconv3) takes a 64×64×64 tensor as input and outputs a 128×128×64 tensor.

[0122] After the skip connection, the eleventh convolutional layer (Conv11) and the twelfth convolutional layer (Conv12) are performed.

[0123] Finally, through the output layer, a convolution operation with a 1×1 kernel maps the 128×128×32 tensor to a 128×128×1 tensor, outputting the predicted flood distribution and peak water depth.

[0124] Add a Dropout layer to the bottleneck part of the convolutional encoder with a dropout rate of 0.5 to prevent overfitting.

[0125] All convolutional and deconvolutional layers in the model are followed by the ReLU activation function, the expression of which is:

[0126] R (x) =max(0,x) (4)

[0127] In the formula: R (x) The value represents the output of the activation function; x represents the input value. When x is greater than 0, the output is x; when x is less than or equal to 0, the output is 0.

[0128] Step 5: Setting up and training the U-Net deep learning image reconstruction model:

[0129] Training samples are constructed using the structured data from step 3, and these samples are then grouped and input into the U-Net deep learning image reconstruction model for training. The specific process for constructing training samples is as follows: The study area is divided into several small samples using a sliding slicing method. The data information contained in each small sample is consistent with the corresponding data information in the original sample. Then, through data augmentation, the small samples are randomly rotated by 90°, rotated by 180°, and mirrored to obtain more small samples, thus forming secondary constructed samples, which are then used in subsequent processes.

[0130] The reconstructed samples are used as training samples and input into the U-Net deep learning image reconstruction model for training and testing, enabling the model to predict output data based on the input data. Then, the samples are recombined and stitched together to generate a complete simulation of the peak water depth and water distribution in the study area.

[0131] In this embodiment, the entire study area is divided into sliding blocks according to a fixed step size, and multiple secondary construction samples with a size of 128×128 pixels are extracted. Each sample contains input data and output data. The input data includes a water accumulation and terrain-related tensor with a size of 128×128×2 (including fine-scale low-precision water accumulation simulation results and geographic digital elevation data), a building outline tensor with a size of 128×128×1 (including building outline data), and a rainfall sequence tensor with a size of 1×25 (including historical rainfall data and design rainstorm data). The output data is a water accumulation depth tensor with a size of 128×128×1 (including peak water accumulation depth and water accumulation distribution).

[0132] In this embodiment, the input data includes building outline data that has undergone edge enhancement processing, and rainfall sequence data is a set of rainfall data for 25 input time periods, used to describe the rainfall change process within the input time period; the output data represents the peak water depth and water distribution of the corresponding sample area within the input time period.

[0133] We use the weighted mean squared error (MSE) as the loss function for the deep learning model. The model parameters are adjusted based on the fine-grid flood prediction results generated during model training. The mathematical expression of the loss function is as follows:

[0134] Loss = Mean[(y Pred -y) 2 ×(1+α·y)×(1+β·(1-y DEM (5)

[0135] In the formula: y is the peak value of the actual water depth on the grid; y pred The peak value of the predicted water depth on the grid; α and β are hyperparameters used to adjust the weight magnitude; y DEM The value is the normalized geographic digital elevation of the grid, with a higher weight for lower terrain; Mean is the average of all pixels in the entire batch to obtain the overall loss value.

[0136] When the training rounds reach the preset number of rounds, or when the validation loss stabilizes and no longer decreases, the model training is complete, and the model parameters are stored in a file with the ".h5" suffix for inference and prediction.

[0137] Step 6: Urban Flood Simulation

[0138] Low-precision waterlogging simulation results for the required rainfall scenarios in the study area are obtained through a one-dimensional-two-dimensional hydrodynamic model. The obtained low-precision waterlogging simulation results are used as input data to train the U-Net deep learning reconstructed image model, and high-precision flood and waterlogging simulation results for the corresponding rainfall scenarios in the study area are output.

[0139] Finally, it should be noted that the above description is only used to illustrate the technical solution of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A city flood simulation method based on a deep learning reconstructed image model, characterized in that, The method comprises the following steps: Step 1, obtaining basic data of the research area: obtaining basic data of the research area, including geographic digital elevation data, remote sensing image data, land use type data, drainage pipe network data, historical rainfall data and waterlogging monitoring data; and modifying the obtained geographic digital elevation data, remote sensing image data, land use type data and drainage pipe network data into the same coordinate system; under the same coordinate system, the road and building contour are framed according to the remote sensing image data as building contour data; Step 2, constructing a one-dimensional-two-dimensional hydrodynamic coupling model: generating two different high-precision grids and low-precision grids in the flood accumulation grid generation tool in the research area, the high-precision grid is defined as a grid division scheme with a spatial resolution of 10 meters or less, and the number of grids in the research area reaches more than 10,000 per square kilometer; the low-precision grid does not meet the high-precision grid standard; a one-dimensional-two-dimensional hydrodynamic coupling model is constructed using high-precision grids according to the obtained geographic digital elevation data, building contour data, land use type data and drainage pipe network data; then, by comparing the simulation results of the one-dimensional-two-dimensional hydrodynamic coupling model with the monitoring data, the model is calibrated and verified; after the one-dimensional-two-dimensional hydrodynamic coupling model using high-precision grids is constructed, the high-precision grids are replaced with low-precision grids, and other parameters remain unchanged, to obtain a one-dimensional-two-dimensional hydrodynamic coupling model using low-precision grids; Step 3, data preprocessing: for the one-dimensional-two-dimensional hydrodynamic coupling model of high-precision and low-precision grids constructed, the multi-field design storm data generated according to the storm formula published by the local meteorological station of the research area and the historical rainfall data of the research area are input into the one-dimensional-two-dimensional hydrodynamic coupling model, the floodwater distribution and peak water depth in the research area are simulated, and then a plurality of urban floodwater simulation results under different rain types of multiple storm recurrence periods are obtained, including high-precision floodwater simulation results and low-precision floodwater simulation results, which are saved in the form of grid data; the low-precision floodwater simulation results are processed to obtain fine-size low-precision floodwater simulation results, and then the fine-size low-precision floodwater simulation results, high-precision floodwater simulation results, geographic digital elevation data, building contour data and rainfall data are structurally processed and stored as structured data; the rainfall data includes historical rainfall data and design storm data; The specific process of processing the low-precision floodwater simulation results to obtain fine-size low-precision floodwater simulation results is as follows: nearest neighbor up-sampling is performed on the low-precision floodwater simulation results, specifically, each pixel of the low-precision floodwater simulation results is directly mapped to an equivalent sub-block region in the target image through spatial expansion, so that the grid size is the same as that of the high-precision floodwater simulation results, thereby obtaining fine-size low-precision floodwater simulation results, and the specific operation is as follows: For any pixel I coarse (i, j) in the original data, its value is copied and padded to the corresponding s x s sub-block in the target data: I fine (si+m, sj+n) = I coarse (i, j) (3) wherein: I coarse is the pixel value of the i-th row and j-th column in the original low-resolution data; s is the expansion multiple; I fine is the pixel value of the si+m-th row and sj+n-th column in the high-resolution data, which is copied from the corresponding I coarse (i,j); m and n are the original data sizes of I coarse , and m, n ∈ {0, 1, 2, …, s-1}. Step 4, constructing a U-Net deep learning reconstructed image model: constructing a U-Net deep learning reconstructed image model, which comprises a rainfall data input layer, a water dynamic waterlogging simulation result input layer, a geographic digital elevation data input layer, a building contour data input layer, a convolutional encoder, a rainfall feature reconstruction layer, a feature fusion layer, a convolutional decoder, a fully connected layer, and an output layer; Step 5, setting and training of the U-Net deep learning reconstructed image model: training samples are constructed based on the structured data in step 3, and the training samples are input into the U-Net deep learning reconstructed image model for training, so that the model can predict output data according to input data; the input data includes rainfall data, fine-size low-precision waterlogging simulation results, geographic digital elevation data, and building contour data within an input period, and the output data is the peak waterlogging depth and waterlogging distribution of the corresponding sample area within the input period; the simulation results of the complete peak waterlogging depth and waterlogging distribution of the study area are generated by recombining and splicing the samples; during the model training process, the model parameters are adjusted, and when the training round reaches the pre-set round or the validation loss tends to be stable and no longer decreases, the model training is completed; Step 6, urban flood simulation: obtaining low-precision waterlogging simulation results of the study area under the rainfall scenario to be predicted by a one-dimensional-two-dimensional water dynamic model, and inputting the obtained low-precision waterlogging simulation results into the trained U-Net deep learning reconstructed image model to output high-precision floodwaterlogging simulation results of the study area under the corresponding rainfall scenario.

2. The urban flood simulation method based on the deep learning reconstructed image model according to claim 1, wherein, The geographic digital elevation data in step 1 is elevation point data, and a rectangular grid digital elevation model DEM is constructed by interpolation in GIS software; the land use type is the underlying surface type of the study area, which is divided into building land, green land, and transportation land; the drainage pipe network data includes the spatial distribution of the drainage pipe network, the length of the pipe, the cross-sectional shape, the pipe diameter, the distribution and depth of the inspection well, and the connection mode of the pipe and the inspection well; the historical rainfall data includes rainfall and rainfall duration; the historical waterlogging monitoring data includes floodwaterlogging distribution, peak waterlogging depth, outlet flow, and pipe network flow; wherein the peak waterlogging depth refers to the maximum waterlogging depth within a given rainfall period. 3.The urban flood simulation method based on deep learning reconstruction image model according to claim 2, wherein, The specific process of the method of calibrating and verifying by comparing the simulation results of the one-dimensional-two-dimensional water dynamic coupling model with the monitoring data in step 2 is as follows: according to the input historical rainfall data, the one-dimensional-two-dimensional water dynamic coupling model simulates the simulation results corresponding to the input historical rainfall data, including floodwaterlogging distribution, peak waterlogging depth, outlet flow, and pipe network flow, and then compares the simulation results with the monitoring data of floodwaterlogging distribution, peak waterlogging depth, outlet flow, and pipe network flow; the accuracy of the one-dimensional-two-dimensional water dynamic coupling model simulation is evaluated by using the Nash efficiency coefficient NSE, and the specific formula is as follows: wherein: S i is the simulation value at time i; O i is the monitoring value at time i; is the average of the monitoring values; n is the total number of monitoring values; the value range of NSE is (-∞, 1], the closer the NSE value is to 1, the better the simulation performance of the one-dimensional-two-dimensional hydrodynamic coupling model, and a threshold is set for the one-dimensional-two-dimensional hydrodynamic coupling model when it is constructed, and if the threshold is lower, the model needs to be reconstructed. 4.The urban flood simulation method based on deep learning reconstruction image model according to claim 1, wherein, The storm formula in step 3 is as follows: In the formula, i is the storm intensity, with the unit of mm / min; P is the storm return period, with the unit of year; t is the rainfall duration, unit is min; parameters A, B, C, n are empirical parameters, which are fitted according to the regional rainstorm statistical data published by the local meteorological department. 5.The urban flood simulation method based on deep learning reconstructed image model according to claim 1, wherein, The specific process of the structured processing of the low-precision fine-size waterlogging simulation result, the high-precision waterlogging simulation result, the geographic digital elevation data, the building contour data and the rainfall data in step 3 is as follows: the spatial index and attribute information of the grid unit are extracted, and are organized into structured data according to the preset field order.

6. The urban flood simulation method based on the deep learning image reconstruction model according to claim 1, wherein The rainfall data input layer in step 4 is used for inputting the vector data of the time series rainfall data, and is connected with the full connection layer after the input layer, so as to gradually expand and reshape the rainfall data into a three-dimensional tensor and fuse with other spatial features; the hydrodynamic waterlogging simulation result input layer is used for inputting the representative fine-size low-precision waterlogging simulation result; the geographic digital elevation data is input by the geographic digital elevation data input layer; the building contour data is input by the building contour data input layer; the hydrodynamic waterlogging simulation result input layer, the geographic digital elevation input layer and the building contour data input layer are connected through a splicing layer to form a three-dimensional tensor as the input of the convolutional encoder; the convolutional encoder includes 12 convolutional layers, 3 pooling layers and 3 inverse convolutional layers, which sequentially extract the hydrodynamic simulation result, the geographic digital elevation data and the building contour data, and specifically include: a first convolutional layer, a second convolutional layer, a first pooling layer, a third convolutional layer, a fourth convolutional layer, a second pooling layer, a fifth convolutional layer, a sixth convolutional layer and a third pooling layer; the rainfall feature reconstruction layer includes multiple full connection layers, which gradually expand the rainfall vector into a tensor matching the output size of the convolutional encoder, and splice with the bottleneck layer features of the convolutional encoder; the feature fusion layer is used for fusing the spatial features output by the convolutional encoder with the expanded rainfall information tensor along the channel dimension to form a comprehensive feature representation; the convolutional decoder includes inverse convolutional layers and convolutional layers, which realize the step-by-step improvement of the spatial resolution through layer-by-layer upsampling and feature recovery, and specifically include: a first inverse convolutional layer, a seventh convolutional layer, an eighth convolutional layer, a second inverse convolutional layer, a ninth convolutional layer, a tenth convolutional layer, a third inverse convolutional layer, an eleventh convolutional layer and a twelfth convolutional layer; the output layer maps the features of the last layer of the decoder to a single-channel output through a linear convolution, which is used for predicting the floodwater distribution and the peak value of the waterlogging depth in the study area; A Dropout layer is added to the bottleneck layer part of the convolutional encoder to prevent overfitting during the training process and improve the generalization ability of the model; An activation function is connected after each convolutional layer, and the activation function adopts a ReLU function to enhance the nonlinear fitting ability, and the specific mathematical expression of the ReLU function is as follows: R (x) = max(0, x) (4) wherein: R (x) represents an activation function output value; x represents an input value; when x is greater than 0, the output is x; when x is less than or equal to 0, the output is 0.

7. The urban flood simulation method based on the deep learning reconstructed image model according to claim 1, wherein, The specific process of constructing the training sample with the structured data in step 3 in step 5 is as follows: considering the hardware performance and the number of research samples, the research area is divided into a plurality of small samples by using the sliding window slicing, and the data information contained in each small sample is consistent with the data information at the corresponding position in the original sample; specifically, in the script, the fine size low precision waterlogging simulation result, the geographic digital elevation data and the building contour data are divided into a plurality of small samples by using the script, the 128*128 rectangle and the 16 step, and each group of 128*128*1 small samples generated is spliced into 128*128*3 small samples through a splicing layer; then, through data enhancement, the small samples are randomly rotated by 90°, rotated by 180° and mirror flipped to obtain more small samples, that is, secondary construction samples are formed, and are put into the subsequent process. 8.The urban flood simulation method based on deep learning reconstruction image model according to claim 1, wherein, The specific process of adjusting the parameters of the model in step 5 is as follows: the loss function using the weighted mean square error Loss is used as the loss function of the U-Net deep learning reconstruction image model, and the parameters of the model are adjusted through the fine grid flood prediction result generated in the model training process, and the mathematical expression of the loss function is as follows: Loss = Mean[(y pred - y 2 ) x (1 + a · y) x (1 + b · (1 - y DEM )] (5) In the formula, y is the real peak value of the accumulated water depth on the grid; y pred is the predicted peak value of the accumulated water depth on the grid; a and b are hyperparameters used to regulate the weight amplitude; y DEM is the normalized value of the geographic digital elevation on the grid, and the lower the terrain, the greater the weight; Mean is the average of all pixel points in the entire batch to obtain the overall loss value.

Images