Urban inland inundation prediction method based on visual identification

Abstract

The invention discloses an urban inland inundation prediction method based on visual identification, belongs to the technical field of urban inland inundation prediction, and can solve the problems that an existing prediction method is low in prediction accuracy and limited in prediction space and time. The method comprises the following steps: S1, determining the real-time ponding depth of each ponding position according to ponding images of a plurality of ponding positions in a target area; s2, according to the meteorological data, the historical ponding depth and the real-time ponding depth of the target area, determining a predicted water depth time sequence of each ponding position in a prediction time period; and S3, according to the predicted water depth time sequence of all the ponding positions in the prediction time period, determining predicted waterlogging information of the target area in the prediction time period. The method is used for predicting urban inland inundation information.

Description

Technical Field

[0001] The invention relates to an urban waterlogging prediction method based on visual recognition, belonging to the technical field of urban waterlogging prediction. Background Art

[0002] Urban waterlogging prediction is a method of predicting possible urban waterlogging disasters in the future by monitoring the waterlogging conditions in urban areas. It is of great significance to ensure the safety of urban functions and residents' travel.

[0003] At present, urban waterlogging prediction methods usually simulate waterlogging data of the target area based on meteorological data, and then make waterlogging predictions based on the simulated waterlogging data. Since there is a deviation between the simulated waterlogging data and the real waterlogging data, the prediction accuracy of the existing prediction methods is low. At the same time, the existing prediction methods can only predict the waterlogging conditions of several specific waterlogging locations in the target area at multiple discrete prediction moments, and can neither predict the overall waterlogging state of the target area nor the waterlogging state of the target area in the future continuous period. Therefore, the prediction space and time of the existing prediction methods are limited. Summary of the invention

[0004] The present invention provides an urban waterlogging prediction method based on visual recognition, which can solve the problems of low prediction accuracy and limited prediction space and time of existing prediction methods.

[0005] The present invention provides a method for predicting urban waterlogging based on visual recognition, the method comprising:

[0006] S1. Determine the real-time water depth of each water accumulation location according to water accumulation images of multiple water accumulation locations in the target area;

[0007] S2. Determine the predicted water depth time series of each waterlogging location in the predicted period according to the meteorological data, the historical waterlogging depth and the real-time waterlogging depth of the target area;

[0008] S3. Determine predicted waterlogging information of the target area in the prediction period according to the predicted water depth time series of all waterlogging locations in the prediction period.

[0009] Optionally, the S1 is specifically:

[0010] S11, determining a real-time flooding level of each waterlogging location according to waterlogging images of multiple waterlogging locations in the target area;

[0011] S12: Determine the real-time water depth of each waterlogged location according to the real-time flooding level.

[0012] Optionally, the S11 is specifically:

[0013] S111, identifying a preset reference object from a water accumulation image of each water accumulation position in the target area, and determining real-time boundary information of the preset reference object in the water accumulation image;

[0014] S112: Determine the real-time flooding level of each waterlogging location according to the real-time boundary information.

[0015] Optionally, the S112 is specifically:

[0016] Constructing a first correspondence between the flooding level and the boundary information of the preset reference object in the waterlogging image;

[0017] The real-time flooding level of each waterlogging location is determined according to the first corresponding relationship and the real-time boundary information.

[0018] Optionally, the S12 is specifically:

[0019] Constructing a second correspondence between the flooding level and the water depth;

[0020] The real-time water accumulation depth of each water accumulation location is determined according to the second corresponding relationship and the real-time flooding level.

[0021] Optionally, the meteorological data includes historical meteorological data and short-term meteorological data; S2 is specifically:

[0022] S21, constructing a water depth time series prediction model using a spatiotemporal graph neural network based on historical meteorological data and historical waterlogging depth of the target area;

[0023] S22. According to the short-term meteorological data of the target area and the real-time waterlogging depth, the predicted water depth time series of each waterlogging location in the prediction period is determined using the water depth time series prediction model.

[0024] Optionally, the water depth time series prediction model includes a prediction unit and a correction unit; S22 is specifically:

[0025] Determine the predicted water depth of each waterlogged location at the current prediction time within the prediction period by using the prediction unit according to the short-term meteorological data of the target area and the real-time waterlogging depth;

[0026] The correction unit is used to correct the predicted water depth at the current prediction moment, and the corrected predicted water depth is used as the input of the prediction unit to perform water depth prediction at the next prediction moment, until the predicted water depths at all prediction moments in the prediction period are determined, and the predicted water depth time series of each waterlogging location in the prediction period is obtained.

[0027] Optionally, S3 is specifically:

[0028] Determine the predicted water depth at any time within the predicted period at any location in the target area according to the predicted water depth time series of all waterlogged locations;

[0029] The predicted waterlogging information of the target area in the prediction period is determined according to the predicted water depth at each position in the target area corresponding to each moment in the prediction period.

[0030] Optionally, before S1, the method further includes:

[0031] Based on the image information of multiple monitoring points in the target area, a large language model is used to screen out multiple waterlogging locations from multiple monitoring points.

[0032] Optionally, the S111 is specifically:

[0033] A semantic segmentation model is used to identify a preset reference object from a water accumulation image at each water accumulation position in a target area, and to determine real-time boundary information of the preset reference object in the water accumulation image.

[0034] The beneficial effects that the present invention can produce include:

[0035] The present invention identifies the real-time water depth of each waterlogging location based on the waterlogging images of multiple waterlogging locations, and then determines the predicted water depth time series of multiple waterlogging locations in the prediction period in combination with the meteorological data and historical waterlogging depth of the target area, and then predicts the waterlogging information of the entire target area in the prediction period based on the predicted water depth time series of multiple waterlogging locations. Since the real-time waterlogging depth based on which the present invention predicts is obtained based on the recognition of waterlogging images, it has authenticity and real-time performance, thus improving the prediction accuracy of waterlogging information. At the same time, the present invention can predict the predicted water depth of each waterlogging location at any moment in the prediction period based on the predicted water depth time series of each waterlogging location in the prediction period, and can predict the predicted water depth of any location in the target area based on the predicted water depths of multiple waterlogging locations in the target area, that is, the present invention can predict the waterlogging status of a continuous period based on the waterlogging status at multiple discrete moments, and can also predict the overall waterlogging status of the target area based on the waterlogging status at multiple dispersed locations, thereby expanding the prediction space and prediction time, and being able to more comprehensively predict urban waterlogging information. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flowchart of a method for predicting urban waterlogging based on visual recognition provided by an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of determining the real-time water depth provided by an embodiment of the present invention;

[0038] Figure 3 A technical roadmap for the prediction method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The present invention is described in detail below in conjunction with embodiments, but the present invention is not limited to these embodiments.

[0040] The embodiment of the present invention provides a method for predicting urban waterlogging based on visual recognition. Figure 1 As shown, the method includes:

[0041] S1. Determine the real-time water depth of each water accumulation location according to water accumulation images of multiple water accumulation locations in the target area;

[0042] S2. Determine the predicted water depth time series of each waterlogging location during the prediction period based on the meteorological data, historical waterlogging depth and real-time waterlogging depth of the target area;

[0043] S3. Determine the predicted waterlogging information of the target area during the prediction period based on the predicted water depth time series of all waterlogging locations during the prediction period.

[0044] Before S1, the method may further include:

[0045] Based on the image information of multiple monitoring points in the target area, a large language model is used to screen out multiple waterlogging locations from multiple monitoring points.

[0046] Specifically, this embodiment sets up multiple monitoring points in the target area, uses a fixed camera for static monitoring at each monitoring point, and uses a mobile camera for dynamic inspection, so as to obtain video data of each monitoring point. The video data is input into the large language model for visual recognition, and multiple monitoring points where waterlogging occurs can be screened out. The location of the monitoring point where waterlogging occurs is the location of waterlogging. The large language model used in this embodiment is the existing open source large model LargeLanguage Model MetaAI 3 (LLaMA 3 for short).

[0047] Specifically, after screening out multiple water accumulation locations, this embodiment performs data preprocessing on the video data of the multiple water accumulation locations. The data preprocessing includes:

[0048] 1. Data annotation: Extract key frames of video data to obtain pictures containing waterlogging information, and then annotate the images in detail. The annotation content includes the boundaries of the waterlogging range, water depth values, and severity levels, so as to obtain labeled data.

[0049] 2. Data enhancement: In order to obtain high-quality labeled data, the labeled data can be enhanced by methods such as image rotation, scaling, translation, and color transformation to improve the generalization ability and robustness of the subsequent semantic segmentation model.

[0050] By performing data preprocessing on the video data of the water accumulation location, the labeled picture obtained is the water accumulation image of the water accumulation location.

[0051] S1 can specifically be:

[0052] S11, determining a real-time flooding level of each waterlogging location according to waterlogging images of multiple waterlogging locations in the target area;

[0053] S12. Determine the real-time water depth of each waterlogged location according to the real-time flooding level.

[0054] In this embodiment, S11 may specifically be:

[0055] S111, identifying a preset reference object from a water accumulation image of each water accumulation position in the target area, and determining real-time boundary information of the preset reference object in the water accumulation image.

[0056] Specifically, the preset reference object can be a common reference object such as a human body or a vehicle in the water, which is easy to estimate the height. The boundary information of the preset reference object in the water image refers to the outline of the part that is not submerged by the water. According to the real-time boundary information of the preset reference object, the current degree of submergence of the water can be determined.

[0057] Furthermore, if Figure 2 As shown, S111 may specifically be:

[0058] The semantic segmentation model is used to identify the preset reference objects from the water accumulation image of each water accumulation location in the target area, and determine the real-time boundary information of the preset reference objects in the water accumulation image.

[0059] Semantic segmentation is an important task in the field of computer vision. It aims to make fine semantic divisions for each pixel in an image, thereby generating a more fine-grained understanding, and is usually used for scene segmentation and object contour recognition. Semantic segmentation models include Fully Convolutional Networks (FCN), the first model to achieve end-to-end semantic segmentation, U-Net, which has excellent performance in medical image processing with encoder-decoder architecture and jump connections, DeepLab series that expands the receptive field and improves resolution through hollow convolution, and MaskR-CNN, which can achieve both target detection and instance segmentation and is widely used in tasks that require precise contours. These semantic segmentation models have promoted the widespread application of semantic segmentation technology in fields such as autonomous driving, remote sensing analysis, and image editing.

[0060] This embodiment uses a semantic segmentation model to assign pixel-by-pixel semantic categories to the water accumulation image of each water accumulation location, so as to extract rich semantic information from the image and achieve fine-grained understanding and analysis of the image, thereby accurately identifying the position and boundary information of preset reference objects in the water accumulation image and performing detailed analysis of the boundary information.

[0061] S112. Determine the real-time flooding level of each waterlogging location based on the real-time boundary information.

[0062] Furthermore, S112 may specifically be:

[0063] Constructing a first correspondence between the flooding level and the boundary information of the preset reference object in the waterlogging image;

[0064] The real-time flooding level of each waterlogging location is determined based on the first corresponding relationship and the real-time boundary information.

[0065] Specifically, in this embodiment, the submergence degree may be divided into a plurality of submergence levels according to the boundary information of the preset reference object at different submergence levels, so that the boundary information at different submergence levels has a corresponding submergence level, forming a first corresponding relationship.

[0066] In this embodiment, S12 may specifically be:

[0067] Constructing a second correspondence between the flooding level and the water depth;

[0068] The real-time water depth of each waterlogging location is determined according to the second corresponding relationship and the real-time flooding level.

[0069] In practice, the second correspondence between the flooding level and the water depth can be flexibly determined according to actual conditions. For example, the second correspondence can be as follows: Figure 2 As shown, taking the truck as the preset reference object, when the inundation level is level 0, the water depth is about 0 cm, when the inundation level is level 1, the water depth is about 80 cm, when the inundation level is level 2, the water depth is about 200 cm; when the inundation level is level 3, the water depth is about 300 cm; taking the human body as the preset reference object, when the inundation level is level 0, the water depth is about 0 cm, when the inundation level is level 1, the water depth is about 40 cm, when the inundation level is level 2, the water depth is about 80 cm; when the inundation level is level 3, the water depth is about 120 cm, and when the inundation level is level 4, the water depth is about 150 cm.

[0070] Specifically, meteorological data may include historical meteorological data and short-term meteorological data. The historical meteorological data may include historical rainfall, historical cumulative rainfall, etc.; the short-term meteorological data may include short-term rainfall, short-term cumulative rainfall, etc.

[0071] S2 can specifically be:

[0072] S21. Based on the historical meteorological data and historical water depth of the target area, a water depth time series prediction model is constructed using a spatiotemporal graph neural network;

[0073] S22. According to the short-term meteorological data of the target area and the real-time water depth of each waterlogging location, a water depth time series prediction model is used to determine the predicted water depth time series of each waterlogging location in the prediction period.

[0074] Specifically, the historical waterlogging depth can be obtained through monitoring data from waterlogging monitoring stations.

[0075] Since the water depth at the waterlogging location is a typical time series data, that is, the rainfall conditions and water depth at the previous few moments will have a greater impact on the water depth at the next few moments, the water depth prediction problem at the waterlogging location can be converted into a supervised learning problem. For example, the rainfall, cumulative rainfall, and water depth at the waterlogging location in the previous few hours can be used to predict the water depth in the next hour.

[0076] Specifically, this embodiment uses the historical meteorological data and historical water depth of the target area to train the spatiotemporal graph neural network and construct a water depth time series prediction model; then the short-term meteorological data of the target area and the real-time water depth of each waterlogging location are used to make actual predictions of the predicted water depth at each prediction moment in the prediction period, and the predicted water depth time series of each waterlogging location in the prediction period is obtained.

[0077] Specifically, before training or prediction, this embodiment needs to extract data such as rainfall, water accumulation depth, and time labels in combination with meteorological data, and then clean and fuse the extracted data. Then, multivariate time series data is generated through methods such as data normalization and reconstruction into supervised data. The multivariate time series data is used as input for subsequent water depth time series prediction model training or prediction.

[0078] Specifically, the spatiotemporal graph neural network used in this embodiment is a multivariate time series graph neural network (MTGNN). MTGNN is a graph neural network model designed specifically for multivariate time series prediction. MTGNN can automatically learn the potential associations between variables by constructing a dynamic adaptive graph structure. It combines graph convolution and time series convolution to make full use of the spatial information of the graph structure and the dynamic characteristics of the time series, thereby improving the prediction accuracy.

[0079] MTGNN is mainly composed of an adaptive graph learning module, a temporal convolutional network module, and a graph convolutional neural network module:

[0080] 1. Adaptive graph learning module

[0081] The goal of this module is to automatically learn the spatial dependencies between nodes.

[0082] Assume that the input multivariate time series data is Where N is the number of nodes, T is the time step, and F is the feature dimension.

[0083] The adaptive graph learning module constructs the graph adjacency matrix through the following formula:

[0084] A ij =softmax(f(x i , x j )) (1)

[0085] In formula (1), A ij represents the association weight between node i and node j, x i represents the feature of node i, x j represents the feature of node j, f(x i , x j ) indicates that based on x i and x j A learnable similarity function.

[0086] 2. Temporal Convolutional Network Module

[0087] This module is responsible for modeling in the time dimension and uses dilated convolution to capture long-term dependencies. The specific formula is:

[0088] H (l+1) =ReLU(W*H (l) +b (l) ) (2)

[0089] In formula (2), H (l+1) represents the hidden state of the l+1th layer, H (l) represents the hidden state of the lth layer, W represents the convolution kernel, * represents the convolution operation, b (l) Represents the bias term of the l+1th layer.

[0090] 3. Graph Convolutional Neural Network Module

[0091] This module performs graph convolution operations on the adaptive graph structure. The specific formula is:

[0092] H (l+1) =σ(AH (l) W (l) +b (l) ) (3)

[0093] In formula (3), H (l+1)represents the hidden state of the l+1th layer, A is the learned graph adjacency matrix, H (l) represents the hidden state of the lth layer, W (l) represents the learnable weight matrix, b (l) represents the bias term of the l+1th layer, and σ represents the activation function.

[0094] MTGNN uses an error-based loss function for training. The formula of the loss function is:

[0095]

[0096] In formula (4), L represents the loss function, represents the predicted value of the model, Y t represents the true label and T is the time step.

[0097] Through back-propagation and stochastic gradient descent algorithms (such as the Adam optimizer), the model's parameters are continuously updated during the training process, ultimately achieving accurate time series prediction.

[0098] In the water depth time series prediction task implemented in this paper, MTGNN can regard the water depth of different water accumulation monitoring stations as nodes in the graph, and construct a dynamic graph structure by implicitly learning the correlation between water accumulation monitoring stations. At the same time, the changes in the time dimension are modeled through the time series convolutional network module. This joint modeling method can effectively capture the spatiotemporal characteristics of water depth changes, thereby achieving better water depth time series prediction results.

[0099] To further improve the accuracy of the prediction, this embodiment introduces a correction unit in the water depth time series prediction model. The correction unit can correct the predicted water depth of the current prediction according to the real-time water depth, and use the corrected predicted water depth as the input for the next prediction, so as to correct the accuracy of the water depth prediction model for multiple future time steps, thereby enhancing the reliability of the prediction.

[0100] Specifically, the water depth time series prediction model includes a prediction unit and a correction unit. In this embodiment, the prediction unit is constructed using a spatiotemporal graph neural network, and the correction unit is constructed using a temporal convolutional graph neural network.

[0101] After the correction unit is introduced, S22 can be specifically:

[0102] According to the short-term meteorological data of the target area and the real-time water depth of each waterlogging location, the prediction unit is used to determine the predicted water depth of each waterlogging location at the current prediction time within the prediction period;

[0103] The predicted water depth at the current prediction moment is corrected by the correction unit, and the corrected predicted water depth is used as the input of the prediction unit to predict the water depth at the next prediction moment, until the predicted water depths at all prediction moments in the prediction period are determined, and the predicted water depth time series of each waterlogging location in the prediction period is obtained.

[0104] Specifically, the prediction period includes a plurality of prediction moments. The prediction unit can predict the water depth at a future prediction moment based on the current real time moment, and the correction unit can receive the real-time water depth at the real time moment determined by S1 in real time.

[0105] For example, if the real time is m and the water depth is to be predicted n times later, the predicted time is (m+n); if the real time is (m+1), the predicted time is (m+n+1). Assuming that the predicted time (m+n) is the first predicted time of the prediction period, the prediction and correction process of this embodiment is as follows:

[0106] First, the water depth at the prediction time (m+n) is predicted based on the real time m. At this time, the prediction unit predicts the water depth at the first prediction time (m+n) based on the short-term meteorological data at the real time m and the real-time water depth, and obtains the predicted water depth at the first prediction time (m+n).

[0107] Then, the water depth at the prediction time (m+n+1) is predicted based on the real time (m+1). At this time, the correction unit compares the predicted water depth at the prediction time (m+n) with the real-time water depth at the real time (m+1), and corrects the predicted water depth at the prediction time (m+n) according to the comparison result to obtain the corrected predicted water depth; the corrected predicted water depth is used as the input of the prediction unit, and the prediction unit predicts the water depth this time according to the corrected predicted water depth and the short-term meteorological data at the real time (m+1), and obtains the predicted water depth at the prediction time (m+n+1).

[0108] Subsequently, the water depth at the prediction time (m+n+2) is predicted based on the real time (m+2). At this time, the correction unit compares the predicted water depth at the prediction time (m+n+1) with the real-time water depth at the real time (m+2), and corrects the predicted water depth at the prediction time (m+n+1) according to the comparison result to obtain the corrected predicted water depth; the corrected predicted water depth is used as the input of the prediction unit, and the prediction unit predicts the water depth this time based on the corrected predicted water depth and the short-term meteorological data at the real time (m+2), and obtains the predicted water depth at the prediction time (m+n+2).

[0109] The subsequent prediction and correction process is as above, until the predicted water depths at all prediction moments in the prediction period are obtained, thereby obtaining the predicted water depth time series of the prediction period.

[0110] By introducing the correction module, this embodiment realizes dynamic feedback and precise correction of the prediction results, which is beneficial to improving the accuracy and reliability of the prediction.

[0111] After obtaining the predicted water depth time series of each waterlogging location in the prediction period, the water depth data of each location in the target area can be supplemented, and the water depth data between adjacent moments in the prediction period can be supplemented at the same time to achieve a comprehensive prediction of waterlogging information in the target area.

[0112] S3 can be specifically:

[0113] S31, determining the predicted water depth at any time within the prediction period corresponding to any position in the target area according to the predicted water depth time series of all waterlogged locations;

[0114] S32. Determine predicted waterlogging information of the target area during the prediction period according to the predicted water depth at each position in the target area at each moment during the prediction period.

[0115] Specifically, for a certain prediction moment, this embodiment can supplement the predicted water depth of any other position in the target area according to the predicted water depth of the above-mentioned multiple dispersed water accumulation positions; for a certain water accumulation position, this embodiment can supplement the predicted water depth of any other time in the prediction period according to the predicted water depth of multiple discrete prediction moments in the prediction period. For example, if the predicted water depth time series of the water accumulation position only includes the predicted water depths at prediction moments such as 3:00, 4:00, and 5:00, this embodiment can determine the water depth data at any time between adjacent prediction times such as 3:01, 3:05, 3:10, 4:45, and 4:56 according to the predicted water depths at these prediction moments. Similarly, if the predicted water depth time series of the water accumulation position only includes the predicted water depths at prediction moments such as 6:00, 7:00, and 8:00, this embodiment can determine the water depth data at any time between adjacent prediction times such as 5:05, 7:20, and 8:50 according to the predicted water depths at these prediction moments.

[0116] Further, S31 may specifically be:

[0117] According to the predicted water depth time series of all waterlogged locations, an interpolation algorithm is used to determine the predicted water depth at any location in the target area at any moment within the prediction period.

[0118] Specifically, this embodiment uses the Inverse Distance Weighting (IDW) and Kriging interpolation methods to determine the predicted water depth at any location in the target area; and uses the Autoregressive Integrated Moving Average (ARIMA) to determine the predicted water depth at any moment in the prediction period.

[0119] In practice, the above three interpolation methods have their own limitations. For example, IDW and Kriging only use spatial information but not time series information, while ARIMA mainly uses time series information but ignores spatial distribution information. Therefore, this embodiment needs to use an integrated learning method to combine the above three interpolation methods, which can not only make up for the shortcomings of the three interpolation methods, but also combine the advantages of the three interpolation methods to complete the water depth data more accurately, so as to achieve accurate prediction of waterlogging information in the target area.

[0120] According to the above method, the technical roadmap of the prediction method of the present invention is as follows Figure 3 As shown. The present invention identifies the real-time water depth of each waterlogging location based on the waterlogging images of multiple waterlogging locations, and then determines the predicted water depth time series of multiple waterlogging locations in the prediction period in combination with the meteorological data and historical waterlogging depth of the target area, and then predicts the waterlogging information of the entire target area in the prediction period based on the predicted water depth time series of multiple waterlogging locations. Since the real-time waterlogging depth based on which the present invention predicts is obtained based on waterlogging image recognition, it has authenticity and real-time performance, thus improving the prediction accuracy of waterlogging information. At the same time, the present invention can predict the predicted water depth of each waterlogging location at any moment in the prediction period based on the predicted water depth time series of each waterlogging location in the prediction period, and can predict the predicted water depth of any location in the target area based on the predicted water depths of multiple waterlogging locations in the target area, that is, the present invention can predict the waterlogging conditions in a continuous period based on the waterlogging conditions at multiple discrete moments, and can also predict the overall waterlogging state of the target area based on the waterlogging conditions at multiple dispersed locations, thereby expanding the prediction space and prediction time, and being able to more comprehensively predict urban waterlogging information.

[0121] The above are only a few embodiments of the present application and do not constitute any form of limitation to the present application. Although the present application is disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the profession, without departing from the scope of the technical solution of the present application, using the technical content disclosed above to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A method for predicting urban flooding based on visual recognition, characterized in that: The method comprises: S1. Determine the real-time water depth of each water accumulation location according to water accumulation images of multiple water accumulation locations in the target area; S2. Determine the predicted water depth time series of each waterlogging location in the predicted period according to the meteorological data, the historical waterlogging depth and the real-time waterlogging depth of the target area; S3. Determine predicted waterlogging information of the target area in the prediction period according to the predicted water depth time series of all waterlogging locations in the prediction period.

2. The method according to claim 1, characterized in that The S1 is specifically: S11, determining a real-time flooding level of each waterlogging location according to waterlogging images of multiple waterlogging locations in the target area; S12: Determine the real-time water depth of each waterlogged location according to the real-time flooding level.

3. The method according to claim 2, characterized in that The S11 is specifically: S111, identifying a preset reference object from a water accumulation image of each water accumulation position in the target area, and determining real-time boundary information of the preset reference object in the water accumulation image; S112: Determine the real-time flooding level of each waterlogging location according to the real-time boundary information.

4. The method according to claim 3, characterized in that The S112 is specifically: Constructing a first correspondence between the flooding level and the boundary information of the preset reference object in the waterlogging image; The real-time flooding level of each waterlogging location is determined according to the first corresponding relationship and the real-time boundary information.

5. The method according to claim 2, characterized in that: The S12 is specifically: Constructing a second correspondence between the flooding level and the water depth; The real-time water accumulation depth of each water accumulation location is determined according to the second corresponding relationship and the real-time flooding level.

6. The method according to claim 1, characterized in that The meteorological data includes historical meteorological data and short-term meteorological data; S2 is specifically: S21, constructing a water depth time series prediction model using a spatiotemporal graph neural network based on historical meteorological data and historical waterlogging depth of the target area; S22. According to the short-term meteorological data of the target area and the real-time waterlogging depth, the predicted water depth time series of each waterlogging location in the prediction period is determined using the water depth time series prediction model.

7. The method according to claim 6, characterized in that The water depth time series prediction model includes a prediction unit and a correction unit; S22 is specifically: Determine the predicted water depth of each waterlogged location at the current prediction time within the prediction period by using the prediction unit according to the short-term meteorological data of the target area and the real-time waterlogging depth; The correction unit is used to correct the predicted water depth at the current prediction moment, and the corrected predicted water depth is used as the input of the prediction unit to perform water depth prediction at the next prediction moment, until the predicted water depths at all prediction moments in the prediction period are determined, and the predicted water depth time series of each waterlogging location in the prediction period is obtained.

8. The method according to claim 1, characterized in that The S3 is specifically: Determine the predicted water depth at any time within the predicted period at any location in the target area according to the predicted water depth time series of all waterlogged locations; The predicted waterlogging information of the target area in the prediction period is determined according to the predicted water depth at each position in the target area corresponding to each moment in the prediction period.

9. The method according to claim 1, characterized in that: Before S1, the method further includes: Based on the image information of multiple monitoring points in the target area, a large language model is used to screen out multiple waterlogging locations from multiple monitoring points.

10. The method according to claim 3, characterized in that The S111 is specifically: A semantic segmentation model is used to identify a preset reference object from a water accumulation image at each water accumulation position in a target area, and to determine real-time boundary information of the preset reference object in the water accumulation image.

Images