Multi-scale progressive short-term flood intelligent early warning method, device, equipment and medium

Abstract

This application relates to a multi-scale progressive short-term intelligent flood early warning method, device, equipment, and medium. The method includes: acquiring rainfall forecast data for a target area within a preset time period, driving a low-resolution numerical model to perform numerical calculations to obtain a first-resolution flood flow field prediction result, using a preset model for partitioned parallel calculations to obtain a second-resolution flood flow field prediction result, then using another preset model to enhance the local flood flow field resolution to obtain a third-resolution flood flow field prediction result, and determining the current warning level of the target area based on pixel range, maximum water depth information, and maximum flow velocity information, thereby issuing a flood warning based on the current warning level. This solves the problems of low computational efficiency, insufficient predictability, and insufficient data fusion capability in flood early warning. Addressing the dynamic needs of flood scenarios, it realizes a progressive system of localized refined flood early warning, rapid trend analysis, accurate disaster alerts, and targeted early warning dissemination.

Description

Technical Field

[0001] This application relates to the field of flood disaster prediction technology, and in particular to a multi-scale progressive short-term intelligent flood early warning method, device, equipment and medium. Background Technology

[0002] Urban stormwaters pose a global threat to people and infrastructure, and this threat is expected to intensify as floods become more frequent and the world’s population grows, with 68% of the global population projected to live in urban areas by 2050. From an economic perspective, the concentration of wealth in urban areas and the accelerated development of infrastructure have led to a significant increase in economic losses from floods, and these losses are expected to increase further globally.

[0003] The flood early warning system in related technologies faces the following technical bottlenecks:

[0004] (1) Low computational efficiency of refined models: Flood early warning systems in related technologies mainly rely on hydrological and hydrodynamic coupled numerical models. When the numerical solution uses relatively fine grid cells, the solution time increases sharply due to the increase in the number of grid cells and the decrease in the allowable time step. This creates a trade-off between simulation resolution and computing resources, making it difficult for related models to be effectively extended to large-scale flood early warning applications. Therefore, how to improve the computational efficiency of flood models and effectively balance the relationship between computational accuracy and computational efficiency is a hot demand for realizing intelligent disaster prevention and mitigation.

[0005] (2) Insufficient predictive power of intelligent models: In order to improve the computational efficiency of flood models, early warning systems in related technologies often use intelligent models based on deep neural networks for rapid prediction. However, due to the constraints of training memory usage, deep learning models in related technologies often adopt an autoregressive paradigm (i.e., using the historical flood inundation distribution of several steps and the current rainfall to predict the flood inundation distribution of the current step) for learning. As a result, the error is prone to accumulate seriously due to the increase in the number of iterations during the rolling forecast process, which makes the predicted flood inundation distribution appear fuzzy and dissipated. It is impossible to accurately analyze the short-term (0-24 hours) flood inundation distribution with high spatiotemporal resolution (flood inundation spatial resolution of not less than 10 minutes and spatial resolution of not less than 5m), making it difficult for the prediction results of deep learning models in related technologies to provide forward-looking guidance for flood control work.

[0006] (3) Insufficient ability to integrate and utilize multi-source, multi-scale elevation data: Surface elevation is a key element controlling the evolution of floods. Considering the accuracy requirements of flood warning, flood warning systems in related technologies mostly use sub-meter-level elevation data based on LiDAR measurements to construct flood models. However, due to the difficulty in obtaining LiDAR measurement permissions and the high cost of measurement, related technologies cannot achieve full coverage of sub-meter-level elevation data, making the integration and utilization of multi-scale elevation data obtained by multiple measurement methods an inevitable trend in the development of flood modeling technology. However, flood modeling in related technologies mainly uses a single-resolution regular grid, which cannot fully leverage the efficiency advantages of multi-scale modeling. This often results in existing flood warning systems facing a trade-off between the overall flood evolution situation awareness capability and the accuracy of local flood disaster detail analysis, making it impossible to coordinate and achieve coordinated early warning of cross-scale flood disaster events in megacities. Summary of the Invention

[0007] This application provides a multi-scale progressive short-term intelligent flood early warning method, device, equipment and medium to solve the problems of low computational efficiency, insufficient predictability and insufficient data fusion and application capability of flood early warning systems. It can effectively balance the relationship between computational efficiency and computational accuracy, and realize multi-source and multi-scale data fusion and rapid short-term flood early warning.

[0008] The first aspect of this application provides a multi-scale progressive short-term flood intelligent early warning method, including the following steps:

[0009] Obtain rainfall forecast data for the target area within a preset time period;

[0010] Based on the rainfall forecast data within the preset time period, a low-resolution numerical model is driven to perform numerical calculations to obtain the first-resolution flood flow field prediction results.

[0011] Based on the flood flow field prediction results at the first resolution, the second resolution flood flow field prediction results are obtained by performing partitioned parallel computation using the preset LSR-UNet-L1 model.

[0012] Based on the second resolution flood flow field prediction results, the local flood flow field resolution is enhanced by using the preset LSR-UNet-L2 model to obtain the third resolution flood flow field prediction results;

[0013] The pixel range is determined based on the maximum water depth information in the third-resolution flood flow field prediction results. Based on the pixel range, the current warning level of the target area is determined according to the maximum water depth information and the maximum flow velocity information in the third-resolution flood flow field prediction results, so as to issue a flood warning according to the current warning level.

[0014] Optionally, in some embodiments, before performing partitioned parallel computation using the preset LSR-UNet-L1 model based on the first-resolution flood flow field prediction result to obtain the second-resolution flood flow field prediction result, the method further includes:

[0015] The region to be simulated is determined, and the region to be simulated is divided into multiple sub-regions, and the constraints of each sub-region are determined.

[0016] Based on the constraints of each sub-region, the rainfall-flood flow field dataset of each sub-region is extracted from the preset flood flow field sequence;

[0017] Based on a preset loss function, a preset neural network is trained using the rainfall-flood flow field dataset of each sub-region to obtain the preset LSR-UNet-L1 model.

[0018] Optionally, in some embodiments, before extracting the rainfall-flood flow field dataset for each sub-region from the preset flood flow field sequence based on the constraints of each sub-region, the method further includes:

[0019] Obtain downscaled products based on satellite precipitation datasets;

[0020] Based on a preset random rainstorm displacement method, multiple sets of random rainfall scenarios are generated according to the downscaling product, and the preset random driving dataset is generated according to the multiple sets of random rainfall scenarios.

[0021] Based on pre-set fused DEM (Digital Elevation Model) data, a high-resolution flood numerical model and a low-resolution flood numerical model are constructed. The high-resolution flood numerical model is obtained based on unstructured grid discretization and hydrodynamic coupling calculation.

[0022] Based on multiple random rainfall scenarios in the preset random-driven dataset, numerical calculations are performed using the low-resolution flood numerical model and the high-resolution flood numerical model to obtain the preset flood flow field sequence, which satisfies the preset time resolution.

[0023] Optionally, in some embodiments, before constructing the high-resolution flood numerical model and the low-resolution flood numerical model based on the preset fused DEM data, the method further includes:

[0024] Based on the preset requirements for urban flood numerical simulation, raster datasets and spatial geographic datasets are obtained;

[0025] Based on the raster dataset and the spatial geographic dataset, and using the preset DEM-LR as a reference, the preset DEM-HR is fused and aligned to the preset DEM-LR to obtain the preset fused DEM data.

[0026] Optionally, in some embodiments, the step of fusing and aligning the preset DEM-HR to the preset DEM-LR based on the preset DEM-LR to obtain the fused DEM data includes:

[0027] Based on the preset DEM-HR measurement boundary line, determine the pixel mean offset between the preset DEM-HR and the preset DEM-LR along the measurement boundary line.

[0028] Based on the preset measurement boundary line of the DEM-HR and the first preset multiple, a buffer is generated according to the preset DEM-HR cell size;

[0029] Based on the boundaries of the buffer, the preset DEM-HR is mosaicked into the preset DEM-LR, and based on the preset buffer smoothing strategy, the preset fused DEM data is obtained according to the mosaicking result.

[0030] Optionally, in some embodiments, the preset LSR-UNet-L2 model is:

[0031] ;

[0032] in, To balance the weighting coefficients of flood flow field variable prediction errors, To address the mean square error in the prediction of flood flow field variables, This refers to the binarized cross-entropy error for predicting the extent of flooding.

[0033] A second aspect of this application provides a multi-scale progressive short-term flood intelligent early warning device, comprising:

[0034] The acquisition module is used to acquire rainfall forecast data for the target area within a preset time period;

[0035] The first prediction module is used to drive a low-resolution numerical model to perform numerical calculations based on rainfall forecast data within the preset time period to obtain the first-resolution flood flow field prediction results.

[0036] The second prediction module is used to perform partitioned parallel computation based on the first resolution flood flow field prediction results and using a preset LSR-UNet-L1 model to obtain the second resolution flood flow field prediction results.

[0037] The third prediction module is used to enhance the local flood field resolution based on the second resolution flood field prediction results using a preset LSR-UNet-L2 model to obtain the third resolution flood field prediction results.

[0038] The early warning module is used to determine the pixel range based on the maximum water depth information in the third-resolution flood flow field prediction results, and based on the pixel range, determine the current early warning level of the target area according to the maximum water depth information and the maximum flow velocity information in the third-resolution flood flow field prediction results, so as to issue a flood early warning according to the current early warning level.

[0039] Optionally, in some embodiments, before performing partitioned parallel computation using the preset LSR-UNet-L1 model based on the first-resolution flood flow field prediction result to obtain the second-resolution flood flow field prediction result, the second prediction module further includes:

[0040] A partitioning unit is used to determine the region to be simulated, divide the region to be simulated into multiple sub-regions, and determine the constraints of each sub-region;

[0041] An extraction unit is used to extract the rainfall-flood flow field dataset of each sub-region from a preset flood flow field sequence based on the constraints of each sub-region;

[0042] The generation unit is used to train a preset neural network based on a preset loss function and the rainfall-flood flow field dataset of each sub-region to obtain the preset LSR-UNet-L1 model.

[0043] Optionally, in some embodiments, before extracting the rainfall-flood flow field dataset for each sub-region from the preset flood flow field sequence based on the constraints of each sub-region, the extraction unit further includes:

[0044] The acquisition sub-unit is used to acquire downscaled products based on satellite precipitation datasets;

[0045] A sub-unit is generated to generate multiple sets of random rainfall scenarios based on a preset random rainstorm displacement method and the downscaling product, and to generate the preset random driving dataset based on the multiple sets of random rainfall scenarios.

[0046] Sub-units are constructed to build high-resolution and low-resolution flood numerical models based on preset fused DEM data. The high-resolution flood numerical model is obtained based on unstructured grid discretization and hydrological-hydrodynamic coupling calculation.

[0047] The computational subunit is used to perform numerical calculations based on multiple sets of random rainfall scenarios in the preset random driving dataset, using the low-resolution flood numerical model and the high-resolution flood numerical model respectively, to obtain the preset flood flow field sequence, wherein the preset flood flow field sequence satisfies the time resolution of a preset duration.

[0048] Optionally, in some embodiments, before constructing the high-resolution flood numerical model and the low-resolution flood numerical model based on the preset fused DEM data, the construction sub-unit further includes:

[0049] The acquisition component is used to acquire raster datasets and spatial geographic datasets based on preset urban flood numerical simulation requirements.

[0050] The fusion component is used to align the preset DEM-HR to the preset DEM-LR based on the raster dataset and the spatial geographic dataset, using the preset DEM-LR as a reference, to obtain the preset fused DEM data.

[0051] Optionally, in some embodiments, the fusion sub-component includes:

[0052] A sub-component is defined to determine the pixel mean offset between the preset DEM-HR and the preset DEM-LR along the measurement boundary line, based on the preset DEM-HR measurement boundary line.

[0053] The first generation sub-component is used to generate a buffer based on the preset measurement boundary line of the DEM-HR and a first preset multiple, according to the preset DEM-HR cell size;

[0054] The second generated component, based on the boundary of the buffer, embeds the preset DEM-HR into the preset DEM-LR, and obtains the preset fused DEM data according to the embedding result based on the preset buffer smoothing strategy.

[0055] Optionally, in some embodiments, the preset LSR-UNet-L2 model is:

[0056] ;

[0057] in, To balance the weighting coefficients of flood flow field variable prediction errors, To address the mean square error in the prediction of flood flow field variables, This refers to the binarized cross-entropy error for predicting the extent of flooding.

[0058] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-scale progressive short-term flood intelligent early warning method as described in the above embodiments.

[0059] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the multi-scale progressive short-term flood intelligent early warning method as described in the above embodiments.

[0060] Therefore, this application has at least the following beneficial effects:

[0061] (1) This application constructs a hierarchical intelligent computing structure of “full-domain partitioned parallel computing - limited-area super-resolution computing”, which effectively takes into account the cross-scale computing needs of full-domain trend inference and local detail analysis, so that users can flexibly adjust the corresponding partitioned computing and super-resolution computing settings according to their own computing resources, data availability and target warning objectives, and effectively balance the relationship between computing efficiency and computing accuracy.

[0062] (2) This application constructs a data-intelligence fusion computing strategy that guides deep learning intelligent computing with low-resolution numerical simulation. It utilizes the robust analytical capability of numerical models to analyze the overall evolution trend of flood flow fields, overcomes the fuzzy dissipation problem faced by intelligent models during multi-step iteration, and on this basis, leverages the powerful learning capability of intelligent models for local high-frequency nonlinear features to overcome the computational efficiency bottleneck faced by numerical models when conducting high-resolution simulations, and achieves rapid short-term flood forecasts with a forecast period of up to 24 hours and a time resolution of no less than 10 minutes.

[0063] (3) This application can align and integrate multi-source and multi-scale basic data, enabling the intelligent model to gradually focus on flood warning areas under the most consistent geographical coordinate system, improve the accuracy of flood flow field analysis, and realize large-scale progressive and refined flood risk warning covering the entire area of ​​megacities and with a spatial resolution of no less than 5 meters for key areas.

[0064] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0065] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0066] Figure 1 This is a schematic diagram illustrating the principle of an LSR-UNet model provided according to an embodiment of this application;

[0067] Figure 2 This is a flowchart of a multi-scale progressive short-term flood intelligent early warning method provided according to an embodiment of this application;

[0068] Figure 3 This is a block diagram of a multi-scale progressive short-term flood intelligent early warning device provided according to an embodiment of this application;

[0069] Figure 4 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0070] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0071] The following description, with reference to the accompanying drawings, describes a multi-scale progressive short-term flood intelligent early warning method, apparatus, equipment, and medium according to embodiments of this application. To address the problems of low computational efficiency, insufficient predictability, and inadequate data fusion and utilization capabilities in flood early warning systems mentioned in the background, this application provides a multi-scale progressive short-term intelligent flood early warning method. In this method, rainfall forecast data for the target area within a preset time period is acquired. Based on the rainfall forecast data within the preset time period, a low-resolution numerical model is driven to perform numerical calculations to obtain a first-resolution flood flow field prediction result. Based on the first-resolution flood flow field prediction result, a preset LSR-UNet-L1 model is used for partitioned parallel computation to obtain a second-resolution flood flow field prediction result. Then, based on the second-resolution flood flow field prediction result, a preset LSR-UNet-L2 model is used to enhance the local flood flow field resolution to obtain a third-resolution flood flow field prediction result. Finally, based on the maximum water depth information in the third-resolution flood flow field prediction result, the pixel range is determined. Based on the pixel range, the current warning level of the target area is determined according to the maximum water depth information and the maximum flow velocity information in the third-resolution flood flow field prediction result, so as to issue a flood warning based on the current warning level. This solves the problems of low computational efficiency, insufficient predictability, and inadequate data fusion and application capabilities in flood early warning systems. It can effectively balance the relationship between computational efficiency and computational accuracy, and realize multi-source, multi-scale data fusion and rapid short-term flood early warning.

[0072] Before introducing the multi-scale progressive short-term flood intelligent early warning method of this application embodiment, let's first introduce the LSR-UNet model.

[0073] Specifically, such as Figure 1As shown, the LSR-UNet model uses the classic UNet network from the field of computer vision as its basic structure. The UNet network is a deep learning model based on a convolutional neural network structure, consisting of an encoder, a decoder, and skip connections.

[0074] (1) Encoder module: It belongs to the input branch of the UNet network and is formed by a series of DownConv units. By continuously applying 3×3 convolution and nonlinear activation functions, it gradually extracts the distribution features of the input image from the local to the global and saves them as feature images of the corresponding scale.

[0075] (2) Decoder module: It belongs to the output branch of the UNet network and is formed by a series of UpConv units. It continuously applies bilinear interpolation and DoubleConv (consisting of two consecutive 3×3 convolutions and nonlinear activation), and splices it with the feature map of the corresponding level to form a high-level feature information representation. Finally, the target image is output by the OutConv unit (consisting of one 1×1 convolution and nonlinear activation).

[0076] (3) Skip connection module: It is mainly used for the connection of feature map information between the decoder and the encoder, thereby allowing the decoder to directly obtain low-level feature information such as local edges and textures extracted by the encoder, which facilitates the UNet model to fully integrate and utilize multi-scale features, reduce the difficulty of network training and improve its simulation accuracy in flood flow field inference.

[0077] In practical use, the input image channel can be defined by flood flow field data, geographic data, etc.

[0078] Figure 2 This is a flowchart illustrating a multi-scale progressive short-term flood intelligent early warning method provided in an embodiment of this application.

[0079] like Figure 2 As shown, this multi-scale progressive short-term flood intelligent early warning method includes the following steps:

[0080] In step S101, rainfall forecast data for the target area within a preset time period is obtained.

[0081] The preset duration can be set by the user, obtained through a limited number of experiments, or obtained through a limited number of computer simulations; no specific limitation is made here.

[0082] Specifically, when conducting multi-scale progressive early warning for short-term flood events within 0-24 hours, the embodiments of this application can obtain short-term rainfall forecast data for the target area within 0-24 hours.

[0083] In step S102, based on rainfall forecast data within a preset time period, a low-resolution numerical model is driven to perform numerical calculations to obtain the first-resolution flood flow field prediction results.

[0084] The first resolution can be preset by the user, obtained through a limited number of experiments, or obtained through a limited number of computer simulations. No specific limitation is made here. Preferably, the first resolution is 300 meters.

[0085] For example, embodiments of this application acquire 0-24 hour short-term rainfall forecast data, thereby driving a low-resolution numerical model. Numerical calculations were performed, and resampling was used to generate flood flow field prediction results with a resolution of 300 meters.

[0086] In step S103, based on the first resolution flood flow field prediction results, the preset LSR-UNet-L1 model is used to perform partitioned parallel computation to obtain the second resolution flood flow field prediction results.

[0087] The second resolution can be preset by the user, obtained through a limited number of experiments, or obtained through a limited number of computer simulations. No specific limitation is made here. Preferably, the second resolution is 30 meters.

[0088] For example, based on the 300-meter resolution flood flow field prediction results, the pre-set LSR-UNet-L1 model is used for partitioned parallel computation to generate 30-meter resolution flood flow field prediction results.

[0089] Furthermore, in some embodiments, before obtaining the second-resolution flood field prediction result by performing partitioned parallel computation using a preset LSR-UNet-L1 model based on the first-resolution flood field prediction result, the method further includes: determining the region to be simulated, dividing the region to be simulated into multiple sub-regions, and determining the constraints of each sub-region; extracting the rainfall-flood field dataset of each sub-region from a preset flood field sequence based on the constraints of each sub-region; and training a preset neural network based on a preset loss function and the rainfall-flood field dataset of each sub-region to obtain a preset LSR-UNet-L1 model.

[0090] Specifically, the embodiments of this application employ a divide-and-conquer modeling approach, decomposing the region to be simulated into... Sub-regions Referring to the spatial resolution characteristics of commonly used publicly available elevation datasets, the resolution of the flood flow field in the sub-regions was set to 30 meters (corresponding to the global high-resolution numerical simulation resampling results) and 300 meters (corresponding to the global low-resolution numerical simulation resampling results). Based on this, to preserve the connectivity between sub-regions and to consider the boundary features of irregular simulation regions and the needs of subsequent intelligent model training, the following constraints were imposed on the sub-region decomposition:

[0091] (1) The overlap area between sub-regions shall not be less than 10%;

[0092] (2) The sub-region covers no less than 70% of the area within the study area;

[0093] (3) The length and width of the sub-region after the number of pixels in the flood flow field resampling according to the high / low resolution should be an integer multiple of 8 (e.g., 64×64, 128×128, 640×640).

[0094] To facilitate subsequent model calculations, a 30-meter resolution flood flow field was used as the base grid. The nearest neighbor sampling method was employed to interpolate the 300-meter flood flow field to each cell within the base grid, ensuring that the two maintain the same length and width.

[0095] Extract rainfall-flood flow field datasets from each sub-region to train the LSR-UNet-L1 intelligent model for parallel prediction of regional flood flow fields. LSR-UNet-L1 takes 300-meter resolution flow field variables (i.e., water depth, x-direction velocity, y-direction velocity), 30-meter elevation (interpolated to the base grid using nearest neighbor sampling), and rainfall distribution (interpolated to the base grid using nearest neighbor sampling) as input, and outputs 30-meter resolution flow field variables. The datasets are then split in a 70:15:15 ratio to form training, validation, and test datasets. Specific training steps include:

[0096] 1. Training with shared parameters among sub-regions, meaning all sub-regions jointly construct a single parameter. intelligent model The model needs to traverse all sub-region samples during training to minimize the loss function. In the formula, For intelligent calculation value and numerical simulation values Mean square error of the mask for pixel-by-pixel comparison (i.e., mean square error excluding pixels falling outside the study area).

[0097] 2. Independent parameter fine-tuning within sub-regions: After completing the training of shared parameters, the shared intelligent model is frozen within each sub-region. The remaining parameters, excluding the last output layer, are fine-tuned using samples from the corresponding sub-regions to ultimately form... The parameters are intelligent model .

[0098] Based on this, and considering the connection relationship between simulated values ​​in different sub-regions, the existing mean square error loss function is used... Based on this, a connection loss function between sub-regions and their adjacent regions is added, and it is agreed that for units with spatial overlap between sub-regions, the average value of the intelligent model of each sub-region is used as its simulation reference value: [The text abruptly ends here, so the translation stops as well.] and For example, the connection loss function based on traversing samples in overlapping regions. and It can be defined as:

[0099] ;

[0100] in, For the first loss function, For the second loss function, sub-region Mean square error of element simulation values ​​in overlapping regions sub-region Mean square error of the simulated values ​​of the elements in the overlapping region.

[0101] By minimizing the connection loss function and This can guide the intelligent models of each sub-region to form variable values ​​that are as consistent as possible at the connecting boundaries.

[0102] Optionally, in some embodiments, before extracting the rainfall-flood field dataset for each sub-region from the preset flood field sequence based on the constraints of each sub-region, the method further includes: obtaining a downscaling product based on a satellite precipitation dataset; generating multiple sets of random rainfall scenarios based on the downscaling product using a preset random rainstorm displacement method, and generating a preset random driving dataset based on the multiple sets of random rainfall scenarios; constructing a high-resolution flood numerical model and a low-resolution flood numerical model based on preset fused DEM data, wherein the high-resolution flood numerical model is obtained based on unstructured grid discretization and hydrodynamic coupling calculation; and performing numerical calculations using the low-resolution flood numerical model and the high-resolution flood numerical model respectively based on the multiple sets of random rainfall scenarios in the preset random driving dataset to obtain a preset flood field sequence, wherein the preset flood field sequence satisfies a preset time resolution.

[0103] Specifically, the embodiments of this application can generate random rainfall scenarios. To ensure that the deep learning model can effectively solve flood processes under different spatiotemporal rainfall-driven scenarios, downscaling products based on satellite precipitation datasets (such as the CMORPH dataset released by the US Climate Prediction Center) are collected (with spatiotemporal resolutions of not less than 1 hour and 1 kilometer, and a total number of rainfall events of not less than 100). Based on this, no less than 1,000 sets of random rainfall scenarios are generated using a preset random rainstorm displacement method, providing a random driving dataset for subsequent flood numerical simulation.

[0104] Based on this, a no-rainfall event and the corresponding no-flood map are introduced, thereby guiding the LSR-UNet model to correctly predict the surface water depth and velocity distribution under no-rainfall scenarios (i.e., zero water depth and zero velocity), and to avoid generating non-physical prediction results as much as possible.

[0105] Here, we will explain how to construct a high-resolution numerical model.

[0106] For the entire study area, a high-resolution numerical model of flooding is constructed based on unstructured grid discretization and coupled hydrological and hydrodynamic calculations. In the unstructured grid discretization process, the river network boundary line and road network boundary line are used as discretization control lines, allowing the unstructured grid to be generated by conforming to the corresponding control lines. This ensures that the subsequent flood numerical calculation process can fully represent the corresponding confluence boundary information. Based on this, for key areas covered by high-resolution elevation data, the maximum grid area constraint used for unstructured grid discretization is appropriately reduced, enabling the grid resolution in these areas to reach the meter level. This ensures that the numerical calculation results for these key areas can provide sufficiently accurate flood flow field information, supporting the assessment needs of relevant flood control and emergency response work.

[0107] This section explains how to construct a low-resolution numerical model.

[0108] For the entire study area, referring to the aforementioned high-resolution numerical model construction method, the control constraints based on river network boundaries and road network boundaries are retained, but the maximum area constraint used for unstructured grid discretization is appropriately increased, so that:

[0109] (1) The total number of grids is located in to ;

[0110] (2) The minimum grid area is not less than 10 .

[0111] This leads to the development of corresponding low-resolution flood numerical models. .

[0112] For random rainfall scenarios, high-resolution numerical models were used respectively. and low-resolution numerical models Numerical calculations were performed to obtain a flood flow field sequence with a time resolution of 10 minutes. Specifically, the sequence consists of flood flow fields updated every 10 minutes; the flow field information is expressed as water depth, x-direction velocity, and y-direction velocity variables stored in an unstructured grid.

[0113] After completing the numerical calculations of floods, the corresponding rainfall scenarios, high-resolution model calculation results, and low-resolution model calculation results are paired and compiled into an ordered flood flow field dataset, which is used to train a pre-set neural network to obtain the LSR-UNet model.

[0114] For ease of analysis, flow field variables stored in unstructured grids can be resampled to a regular grid with a fixed resolution (e.g., 30m, 5m, 2m, etc.) for a specific target area. The values ​​of the regular grid variables can be defined as the area-weighted average of the unstructured grid variables that intersect with the regular grid.

[0115] Optionally, in some embodiments, before constructing a high-resolution flood numerical model and a low-resolution flood numerical model based on preset fused DEM data, the method further includes: obtaining a raster dataset and a spatial geographic dataset based on preset urban flood numerical simulation requirements; and fusion and alignment of the preset DEM-HR to the preset DEM-LR based on the raster dataset and the spatial geographic dataset, using the preset DEM-LR as a reference, to obtain preset fused DEM data.

[0116] Specifically, in accordance with the requirements of urban flood numerical simulation, spatial geographic datasets such as digital elevation models (DEMs), land use types, and building area density raster datasets were collected. For multi-source, multi-scale DEMs acquired by different measurement methods (such as lidar measurements and satellite measurements) and covering different ranges, the satellite-measured DEM with the largest coverage and lowest spatial resolution (referred to as DEM-LR) was used as the benchmark, and the remaining DEMs with smaller coverage and higher spatial resolution (referred to as DEM-HR) were fused and aligned to DEM-LR.

[0117] Optionally, in some embodiments, using a preset DEM-LR as a reference, the preset DEM-HR is fused and aligned to the preset DEM-LR to obtain fused DEM data, including: determining the pixel mean offset between the preset DEM-HR and the preset DEM-LR along the measurement boundary line based on the measurement boundary line of the preset DEM-HR; generating a buffer based on the measurement boundary line of the preset DEM-HR and a first preset multiple, according to the preset DEM-HR pixel size; embedding the preset DEM-HR into the preset DEM-LR based on the boundary of the buffer; and obtaining the preset fused DEM data based on the embedding result according to a preset buffer smoothing strategy.

[0118] Specifically, in this embodiment of the application, the preset DEM-LR is used as a reference, and the preset DEM-HR is fused and aligned to the preset DEM-LR to obtain fused DEM data through the following steps.

[0119] (1) System error correction: Statistically analyze the pixel mean offset of DEM-HR and DEM-LR along the boundary line, and deduct the mean offset from each pixel value in DEM-HR.

[0120] (2) Boundary buffer smoothing: Along the boundary line of DEM-HR measurement, a buffer with a size 10 times that of DEM-HR pixels is generated. DEM-HR is then mosaicked into DEM-LR according to the boundary of the buffer. Along the boundary line, a two-dimensional Gaussian filter with a kernel size 5 times that of DEM-HR pixels is used to smooth the elevation data in the buffer, thereby obtaining the fused DEM data DEM-LR-HR covering the range of DEM-HR.

[0121] When constructing the flood numerical model, for model cells falling within the DEM-HR range, the DEM data resolution covering the area is used as the priority order, so that the cells can preferentially read DEM-LR-HR data from the highest spatial resolution.

[0122] In step S104, based on the second-resolution flood flow field prediction results, the local flood flow field resolution is enhanced using the preset LSR-UNet-L2 model to obtain the third-resolution flood flow field prediction results.

[0123] The third resolution can be preset by the user, obtained through a limited number of experiments, or obtained through a limited number of computer simulations. No specific limitation is made here. Preferably, the third resolution is 3 meters.

[0124] For example, based on the 30-meter resolution flood flow field prediction results, focusing on key areas, the LSR-UNet-L2 model is used to enhance the resolution of local flood flow fields and generate 3-meter resolution flood flow field prediction results.

[0125] Specifically, for key local areas with high-resolution basic information, referencing commonly used refined flood simulation resolutions, the high-resolution numerical simulation results in these areas are resampled to 3 meters and 30 meters, respectively, corresponding to the local high-resolution flood flow field simulation results. (wherein, serial number) The simulation results represent different types of flow field variables, namely water depth, x-direction velocity, and y-direction velocity, as well as local low-resolution flood flow field simulation results. Furthermore, it ensures that the length and width of the sub-region based on the pixel count after resampling are both integer multiples of 8 (e.g., 64×64, 128×128, 640×640). If the initial region cannot meet this dimensional condition, the region size is adjusted appropriately until the dimensional condition is met.

[0126] by Using the base grid, the nearest neighbor sampling method is employed to interpolate spatial geographic datasets such as surface elevation, roughness (converted from land use type tables), impermeability (converted from land use type tables), and building area ratio into each cell of the base grid, resulting in... High-resolution substrate data with consistent length and width Meanwhile, to allow LSR-UNet-L2 to reuse the model structure of LSR-UNet-L1 as much as possible, following the approach in step S103, the nearest neighbor sampling method is adopted. Interpolate to the base grid, making and Maintain a consistent length and width.

[0127] Here, we will explain how to train the preset LSR-UNet-L2.

[0128] Specifically, high- and low-resolution flood flow field datasets are extracted from key local areas, and an LSR-UNet-L2 intelligent model is trained to enhance the resolution of local flood flow fields. Specifically, the LSR-UNet-L2 intelligent model uses flow field variables at a resolution of 30 meters. and 3-meter resolution substrate data As input, output flow field variables with 3-meter resolution. The datasets were then split into training, validation, and test datasets in a 70:15:15 ratio, which were used for model parameter learning, optimal parameter selection, and resolution performance evaluation, respectively.

[0129] To address the significant spatial variability of meter-resolution flood flow fields due to local micro-topography, an additional output branch for flood extent prediction is introduced into the LSR-UNet-L2 intelligent model output. Correspondingly, the existing mask mean square error based on pixel-by-pixel comparison is optimized. (in the formula) Based on the weighting coefficients assigned to different variables, a binarized cross-entropy error is introduced to measure the accuracy of flood inundation range prediction.

[0130]

[0131] in, For mathematical expectation, The pixel flooding probability is predicted by the pre-set LSR-UNet-L2 intelligent model. This represents the submersion label value of a pixel based on high-resolution simulated water depth classification, specifically when the simulated water depth exceeds 0.15 meters. Conversely Therefore, the error function required for training the super-resolution model can be expressed as:

[0132]

[0133] Based on the above error function representation, the following two-stage training is conducted:

[0134] 1. Initial training based on adaptive weight parameters: This involves using a multi-task learning algorithm to expand the original error function as follows:

[0135]

[0136] in, To balance the weighting coefficients for prediction errors of flood flow field variables (i.e., water depth, x-direction velocity, and y-direction velocity), To address the mean square error in the prediction of flood flow field variables, This refers to the binarized cross-entropy error for predicting the extent of flooding.

[0137] Therefore, the parameters and weight coefficients of the LSR-UNet-L2 intelligent model are jointly trained (to ensure training stability, further replacements are possible). And obtain the optimal combination of weight coefficients. .

[0138] 2. Reinforcement training based on fixed weight parameters: Based on the initial training results, reinforcement training is performed using the optimal weight coefficients and the original error function.

[0139]

[0140] This yields the final parameters of the LSR-UNet-L2 intelligent model.

[0141] In step S105, the pixel range is determined based on the maximum water depth information in the third-resolution flood flow field prediction results, and based on the pixel range, the current warning level of the target area is determined according to the maximum water depth information and the maximum flow velocity information in the third-resolution flood flow field prediction results, so as to issue a flood warning according to the current warning level.

[0142] Specifically, the maximum water depth included in the flood flow field at a 3-meter resolution. Maximum flow rate Information, filtering for the maximum water depth of all samples. The pixel range, and within that range, according to lie in , , , The warning levels are assigned as low risk, medium risk, relatively high risk, and high risk, respectively, and flood warnings are issued based on the current warning level.

[0143] Therefore, this application's embodiment, based on a limited-area super-resolution UNet network, achieves real-time inference of urban flood evolution processes with spatiotemporal resolution down to the minute and meter level through an encoder-decoder architecture and dynamic weight allocation, thereby enabling localized, targeted, and precise early warning. By integrating a multi-scale progressive intelligent computing architecture that combines numerical simulation, deep learning, partitioned parallelism, and super-resolution analysis, it effectively balances computational efficiency and accuracy, providing new impetus for high-efficiency urban disaster prevention and mitigation.

[0144] According to the multi-scale progressive short-term intelligent flood early warning method proposed in this application, rainfall forecast data of the target area within a preset time period is obtained. Based on the rainfall forecast data within the preset time period, a low-resolution numerical model is driven to perform numerical calculations to obtain the first-resolution flood flow field prediction result. Based on the first-resolution flood flow field prediction result, a preset LSR-UNet-L1 model is used to perform partitioned parallel calculations to obtain the second-resolution flood flow field prediction result. Based on the second-resolution flood flow field prediction result, a preset LSR-UNet-L2 model is used to enhance the local flood flow field resolution to obtain the third-resolution flood flow field prediction result. Finally, the pixel range is determined based on the maximum water depth information in the third-resolution flood flow field prediction result. Based on the pixel range, the current warning level of the target area is determined according to the maximum water depth information and the maximum flow velocity information in the third-resolution flood flow field prediction result, so as to issue a flood warning according to the current warning level. This solves the problems of low computational efficiency, insufficient predictability, and inadequate data fusion and application capabilities in flood early warning systems. It can effectively balance the relationship between computational efficiency and computational accuracy, and realize multi-source, multi-scale data fusion and rapid short-term flood early warning.

[0145] Next, referring to the accompanying drawings, a multi-scale progressive short-term flood intelligent early warning device proposed according to an embodiment of this application is described.

[0146] Figure 3 This is a block diagram of a multi-scale progressive short-term flood intelligent early warning device according to an embodiment of this application.

[0147] like Figure 3 As shown, the multi-scale progressive short-term flood intelligent early warning device 10 includes: an acquisition module 100, a first prediction module 200, a second prediction module 300, a third prediction module 400, and an early warning module 500.

[0148] The acquisition module 100 is used to acquire rainfall forecast data for the target area within a preset time period.

[0149] The first prediction module 200 is used to drive a low-resolution numerical model to perform numerical calculations based on rainfall forecast data within a preset time period to obtain the first-resolution flood flow field prediction results.

[0150] The second prediction module 300 is used to perform partitioned parallel computation based on the first resolution flood flow field prediction results and using the preset LSR-UNet-L1 model to obtain the second resolution flood flow field prediction results.

[0151] The third prediction module 400 is used to enhance the local flood field resolution based on the second-resolution flood field prediction results using a preset LSR-UNet-L2 model to obtain the third-resolution flood field prediction results.

[0152] The early warning module 500 is used to determine the pixel range based on the maximum water depth information in the third-resolution flood flow field prediction results, and based on the pixel range, determine the current early warning level of the target area according to the maximum water depth information and the maximum flow velocity information in the third-resolution flood flow field prediction results, so as to issue flood warnings according to the current early warning level.

[0153] Optionally, in some embodiments, before performing partitioned parallel computation using a preset LSR-UNet-L1 model based on the first resolution flood flow field prediction results to obtain the second resolution flood flow field prediction results, the second prediction module 300 further includes: a partitioning unit, an extraction unit, and a generation unit.

[0154] The partitioning unit is used to determine the region to be simulated, divide the region to be simulated into multiple sub-regions, and determine the constraints of each sub-region.

[0155] The extraction unit is used to extract the rainfall-flood flow field dataset for each sub-region from the preset flood flow field sequence based on the constraints of each sub-region.

[0156] The generation unit is used to train a preset neural network based on a preset loss function and the rainfall-flood flow field dataset of each sub-region to obtain a preset LSR-UNet-L1 model.

[0157] Optionally, in some embodiments, before extracting the rainfall-flood flow field dataset of each sub-region from a preset flood flow field sequence based on the constraints of each sub-region, the extraction unit further includes: acquiring sub-units, generating sub-units, constructing sub-units, and calculating sub-units.

[0158] The acquisition sub-unit is used to acquire downscaled products based on satellite precipitation datasets.

[0159] The generation sub-unit is used to generate multiple sets of random rainfall scenarios based on the preset random rainstorm displacement method and the downscaling product, and to generate a preset random driving dataset based on the multiple sets of random rainfall scenarios.

[0160] Sub-units are constructed to build high-resolution and low-resolution flood numerical models based on pre-defined fused DEM data. The high-resolution flood numerical model is obtained by unstructured grid discretization and hydrological-hydrodynamic coupling calculation.

[0161] The computational subunit is used to perform numerical calculations based on multiple sets of random rainfall scenarios in a preset random driving dataset, using both low-resolution and high-resolution flood numerical models, to obtain a preset flood flow field sequence. The preset flood flow field sequence satisfies a preset time resolution.

[0162] Optionally, in some embodiments, before constructing a high-resolution flood numerical model and a low-resolution flood numerical model based on preset fused DEM data, the construction of sub-units may further include: an acquisition sub-component and a fusion sub-component.

[0163] The acquisition sub-component is used to acquire raster datasets and spatial geographic datasets based on preset urban flood numerical simulation requirements.

[0164] The fusion component is used to merge and align the preset DEM-HR to the preset DEM-LR based on the raster dataset and the spatial geographic dataset, using the preset DEM-LR as a reference, to obtain the preset fused DEM data.

[0165] Optionally, in some embodiments, the fusion sub-component includes: a determination sub-component, a first generation sub-component, and a second generation sub-component.

[0166] The determination sub-component is used to determine the pixel mean offset between the preset DEM-HR and the preset DEM-LR along the measurement boundary line based on the preset DEM-HR measurement boundary line.

[0167] The first generation sub-component is used to generate a buffer based on the preset measurement boundary line of the DEM-HR and a first preset multiple, according to the preset DEM-HR cell size.

[0168] The second generation sub-component is used to embed the preset DEM-HR into the preset DEM-LR based on the buffer boundary, and to obtain the preset fused DEM data based on the embedding result according to the preset buffer smoothing strategy.

[0169] Optionally, in some embodiments, the preset LSR-UNet-L2 model is:

[0170] ;

[0171] in, To balance the weighting coefficients for prediction errors of flood flow field variables (i.e., water depth, x-direction velocity, and y-direction velocity), To address the mean square error in the prediction of flood flow field variables, This refers to the binarized cross-entropy error for predicting the extent of flooding.

[0172] It should be noted that the foregoing explanation of the multi-scale progressive short-term flood intelligent early warning method embodiment also applies to the multi-scale progressive short-term flood intelligent early warning device of this embodiment, and will not be repeated here.

[0173] According to the multi-scale progressive short-term flood intelligent early warning device proposed in this application, rainfall forecast data of the target area within a preset time period is acquired. Based on the rainfall forecast data within the preset time period, a low-resolution numerical model is driven to perform numerical calculations to obtain a first-resolution flood flow field prediction result. Based on the first-resolution flood flow field prediction result, a preset LSR-UNet-L1 model is used to perform partitioned parallel calculations to obtain a second-resolution flood flow field prediction result. Based on the second-resolution flood flow field prediction result, a preset LSR-UNet-L2 model is used to enhance the local flood flow field resolution to obtain a third-resolution flood flow field prediction result. Finally, the pixel range is determined based on the maximum water depth information in the third-resolution flood flow field prediction result. Based on the pixel range, the current warning level of the target area is determined according to the maximum water depth information and the maximum flow velocity information in the third-resolution flood flow field prediction result, so as to issue a flood warning according to the current warning level. This solves the problems of low computational efficiency, insufficient predictability, and inadequate data fusion and application capabilities in flood early warning systems. It can effectively balance the relationship between computational efficiency and computational accuracy, and realize multi-source, multi-scale data fusion and rapid short-term flood early warning.

[0174] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0175] The memory 401, the processor 402, and the computer program stored on the memory 401 and capable of running on the processor 402.

[0176] When the processor 402 executes the program, it implements the multi-scale progressive short-term flood intelligent early warning method provided in the above embodiments.

[0177] Furthermore, electronic devices also include:

[0178] Communication interface 403 is used for communication between memory 401 and processor 402.

[0179] The memory 401 is used to store computer programs that can run on the processor 402.

[0180] The memory 401 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0181] If the memory 401, processor 402, and communication interface 403 are implemented independently, then the communication interface 403, memory 401, and processor 402 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0182] Optionally, in a specific implementation, if the memory 401, processor 402, and communication interface 403 are integrated on a single chip, then the memory 401, processor 402, and communication interface 403 can communicate with each other through an internal interface.

[0183] Processor 402 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0184] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described multi-scale progressive short-term flood intelligent early warning method.

[0185] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0186] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0187] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0188] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0189] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0190] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A multi-scale progressive short-term intelligent flood early warning method, characterized in that, Includes the following steps: Obtain rainfall forecast data for the target area within a preset time period; Based on the rainfall forecast data within the preset time period, a low-resolution numerical model is driven to perform numerical calculations to obtain the first-resolution flood flow field prediction results. Based on the flood flow field prediction results at the first resolution, the second resolution flood flow field prediction results are obtained by performing partitioned parallel computation using the preset LSR-UNet-L1 model. Based on the second resolution flood flow field prediction results, the local flood flow field resolution is enhanced by using the preset LSR-UNet-L2 model to obtain the third resolution flood flow field prediction results; The pixel range is determined based on the maximum water depth information in the third-resolution flood flow field prediction results. Then, based on this pixel range, and according to the maximum water depth information and the maximum flow velocity information in the third-resolution flood flow field prediction results, the current warning level for the target area is determined, so as to issue a flood warning based on the current warning level. Before performing partitioned parallel computation using the preset LSR-UNet-L1 model to obtain the second-resolution flood flow field prediction results, the region to be simulated is determined, decomposed into multiple sub-regions, and constraints are defined for each sub-region. Based on the constraints of each sub-region, rainfall-flood flow field datasets are extracted from the preset flood flow field sequence for each sub-region. Based on a preset loss function, the preset LSR-UNet-L1 model is trained using the rainfall-flood flow field datasets of each sub-region. The preset LSR-UNet-L1 model takes 300-meter resolution flow field variables, 30-meter elevation, and rainfall distribution as inputs, outputs 30-meter resolution flow field variables, and is divided into training datasets, validation datasets, and test datasets according to a first preset ratio. Training the preset LSR-UNet-L1 model includes: All sub-regions jointly construct a single parameter. intelligent model The intelligent model traverses the sub-regions during training to minimize the loss function. , wherein For intelligent calculation value and numerical simulation values Mean square error of the mask for pixel-by-pixel comparison; Freeze the smart model within each sub-region. The remaining parameters, excluding the last output layer, are fine-tuned using samples from the corresponding sub-regions to ultimately form... The parameters are intelligent model ; Furthermore, based on the overlapping region sample traversal connection loss function and , ； in, For the first loss function, For the second loss function, sub-region Mean square error of element simulation values ​​in overlapping regions sub-region Mean square error of the simulated values ​​of the elements in the overlapping region.

2. The method according to claim 1, characterized in that, Before extracting the rainfall-flood flow field dataset for each sub-region from the preset flood flow field sequence based on the constraints of each sub-region, the process further includes: Obtain downscaled products based on satellite precipitation datasets; Based on a preset random rainstorm displacement method, multiple sets of random rainfall scenarios are generated according to the downscaling product, and the preset random driving dataset is generated according to the multiple sets of random rainfall scenarios. Based on the pre-set fused digital elevation model (DEM) data, a high-resolution flood numerical model and a low-resolution flood numerical model are constructed. The high-resolution flood numerical model is obtained based on unstructured grid discretization and hydrodynamic coupling calculation. Based on multiple random rainfall scenarios in the preset random-driven dataset, numerical calculations are performed using the low-resolution flood numerical model and the high-resolution flood numerical model to obtain the preset flood flow field sequence, which satisfies the preset time resolution.

3. The method according to claim 2, characterized in that, Before constructing the high-resolution and low-resolution flood numerical models based on the preset fused DEM data, the following steps are also included: Based on the preset requirements for urban flood numerical simulation, raster datasets and spatial geographic datasets are obtained; Based on the raster dataset and the spatial geographic dataset, and using the preset DEM-LR as a reference, the preset DEM-HR is fused and aligned to the preset DEM-LR to obtain the preset fused DEM data.

4. The method according to claim 3, characterized in that, The process of fusing and aligning a preset DEM-HR to the preset DEM-LR, based on a preset DEM-LR, to obtain the fused DEM data, includes: Based on the preset DEM-HR measurement boundary line, determine the pixel mean offset between the preset DEM-HR and the preset DEM-LR along the measurement boundary line. Based on the preset measurement boundary line of the DEM-HR and the first preset multiple, a buffer is generated according to the preset DEM-HR cell size; Based on the boundaries of the buffer, the preset DEM-HR is mosaicked into the preset DEM-LR, and based on the preset buffer smoothing strategy, the preset fused DEM data is obtained according to the mosaicking result.

5. The method according to any one of claims 1-4, characterized in that, The preset LSR-UNet-L2 model is: ； in, To balance the weighting coefficients of flood flow field variable prediction errors, To address the mean square error in the prediction of flood flow field variables, This refers to the binarized cross-entropy error for predicting the extent of flooding.

6. A multi-scale progressive short-term flood intelligent early warning device, characterized in that, include: The acquisition module is used to acquire rainfall forecast data for the target area within a preset time period; The first prediction module is used to drive a low-resolution numerical model to perform numerical calculations based on rainfall forecast data within the preset time period to obtain the first-resolution flood flow field prediction results. The second prediction module is used to perform partitioned parallel computation based on the first resolution flood flow field prediction results and using a preset LSR-UNet-L1 model to obtain the second resolution flood flow field prediction results. The third prediction module is used to enhance the local flood field resolution based on the second resolution flood field prediction results using a preset LSR-UNet-L2 model to obtain the third resolution flood field prediction results. The early warning module is used to determine the pixel range based on the maximum water depth information in the third-resolution flood flow field prediction results, and based on the pixel range, determine the current early warning level of the target area according to the maximum water depth information and the maximum flow velocity information in the third-resolution flood flow field prediction results, so as to issue a flood early warning according to the current early warning level. Before performing partitioned parallel computation using the preset LSR-UNet-L1 model to obtain the second-resolution flood flow field prediction results, the region to be simulated is determined, decomposed into multiple sub-regions, and constraints for each sub-region are defined. Based on the constraints of each sub-region, rainfall-flood flow field datasets are extracted from the preset flood flow field sequence for each sub-region. Based on a preset loss function, the preset LSR-UNet-L1 model is trained using the rainfall-flood flow field datasets of each sub-region. The preset LSR-UNet-L1 model takes 300-meter resolution flow field variables, 30-meter elevation, and rainfall distribution as inputs, outputs 30-meter resolution flow field variables, and is divided according to a first preset ratio to form training datasets, validation datasets, and test datasets. Training the preset LSR-UNet-L1 model includes: All sub-regions jointly construct a single parameter. intelligent model The intelligent model traverses the sub-regions during training to minimize the loss function. , wherein For intelligent calculation value and numerical simulation values Mean square error of the mask for pixel-by-pixel comparison; Freeze the smart model within each sub-region. The remaining parameters, excluding the last output layer, are fine-tuned using samples from the corresponding sub-regions to ultimately form... The parameters are intelligent model ; Furthermore, based on the overlapping region sample traversal connection loss function and , ； in, For the first loss function, For the second loss function, sub-region Mean square error of element simulation values ​​in overlapping regions sub-region Mean square error of the simulated values ​​of the elements in the overlapping region.

7. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-scale progressive short-term flood intelligent early warning method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the multi-scale progressive short-term flood intelligent early warning method as described in any one of claims 1-5.

Images