A method, system, device and medium for urban storm flood risk assessment

Abstract

The application provides a city rainstorm flood risk assessment method, system, device and medium, and belongs to the technical field of flood disaster emergency management. The method comprises the following steps: obtaining real-time social media data of a target city rainstorm flood event, and constructing different types of data sets; based on the constructed data sets, using a trained ViT to screen out rainstorm flood picture real-time data; using a trained DeepLab V3+ to extract water body objects; using a trained YOLO V8 to extract human and vehicle targets; detecting whether the water body objects and the human and vehicle targets overlap, and if they overlap, marking them as risky pictures; synchronizing the risky positions corresponding to all risky pictures to a map, and generating a heat map; and evaluating a city rainstorm flood risk area through the heat map. The application can accurately evaluate a city rainstorm flood risk area.

Description

Technical Field

[0001] This invention belongs to the field of flood disaster emergency management technology, specifically relating to a method, system, equipment and medium for urban rainstorm and flood risk assessment. Background Technology

[0002] Floods, as a natural disaster, cause severe damage to human life, agricultural production, and the ecological environment. In recent years, with the increase in global climate change and extreme weather events, urban flooding has become more frequent and its destructive power has continued to intensify. Therefore, timely and accurate monitoring of severely flooded urban areas and accurate assessment of urban rainstorm and flood risks are of great significance for flood prevention and disaster reduction, disaster emergency response, and post-disaster reconstruction.

[0003] Currently, the main method for assessing urban rainstorm and flood risk is to use natural language processing technology to analyze the acquired disaster information, mine flood-related events, and assess urban rainstorm and flood risk based on the mined information. However, when using this method to analyze disaster information, the lack of accurate spatial positioning and scene description makes it impossible to effectively represent the real-time dynamic changes of floods. Therefore, it cannot accurately assess urban rainstorm and flood risk. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies in accurately assessing urban rainstorm and flood risks, this invention provides a method for assessing urban rainstorm and flood risks, comprising the following steps:

[0005] Acquire real-time social media data on rainstorm and flood events in the target city, and construct a real-time dataset of positive samples containing flood scene images, a real-time water body dataset of flood images, a real-time target dataset of flood images containing people and vehicles, and a real-time dataset of negative samples containing non-flood scene images using the real-time social media data.

[0006] Data from a real-time dataset containing positive samples of flood scene images and a real-time dataset containing negative samples of non-flood scene images are input into a pre-trained ViT dataset to filter out real-time rainstorm and flood image data, and then the address information of the real-time rainstorm and flood image data is extracted. Data from a real-time water body dataset of flood images and a real-time dataset containing negative samples of non-flood scene images are input into a pre-trained DeepLab V3+ dataset to extract water body objects. Data from a real-time target dataset of flood images containing people and vehicles and a real-time dataset containing negative samples of non-flood scene images are input into a pre-trained YOLO V8 dataset to extract people and vehicle targets. The overlap between water body objects and people and vehicle targets is detected. If there is an overlap, the image is marked as a dangerous image.

[0007] By combining the address information from real-time rainstorm and flood images, the locations of all potentially hazardous images are synchronized to the map and a heat map is generated; the heat map is then used to assess the risk areas of urban rainstorm and flooding.

[0008] Preferably, the step of extracting the address information of real-time rainstorm and flood image data includes the following steps:

[0009] The ChatGLM3-6B large language model was used to extract place name and address information from real-time data of rainstorm and flood images, and only the place name and address information was retained.

[0010] Use Baidu's coordinate picking system to convert place name and address information into latitude and longitude coordinate data.

[0011] Preferably, the detection of whether the water object overlaps with the people and vehicles is performed. If they overlap, the image is marked as a potentially dangerous image. Specifically, the detection process involves checking whether there are water pixels within the bounding box containing the people and vehicles. If the water pixels extracted by DeepLab V3+ do not appear inside the rectangular bounding box, the people and vehicles are located on the roadside, not submerged, and the image containing them is not a potentially dangerous image. If the water pixels extracted by DeepLab V3+ appear inside the rectangular bounding box, the rectangular bounding box is expanded by 1 / 2 and 1 times, respectively. The percentage of water pixels within the three different sized rectangular bounding boxes is counted. When the percentage reaches 10% or more, the image is marked as a potentially dangerous image.

[0012] Preferably, the pre-training process for ViT, DeepLab V3+, and YOLO V8 includes the following steps:

[0013] Acquire historical social media data related to urban rainstorms and floods, and construct a positive sample historical dataset containing flood scene images, a historical water body dataset containing flood images, a historical target dataset containing people and vehicles containing flood images, and a negative sample historical dataset containing non-flood scene images based on the historical social media data related to urban rainstorms and floods.

[0014] The deep learning model ViT was trained using a historical dataset of positive samples containing flood scene images and a negative sample dataset containing non-flood scene images; DeepLab V3+ was trained using a historical dataset of flood scene images of water bodies and a negative sample dataset containing non-flood scene images; and YOLO V8 was trained using a historical dataset of flood scene images of targets containing people and vehicles and a historical dataset containing negative sample images of non-flood scene images.

[0015] Preferably, before training DeepLab V3+ using the historical water body dataset of flood images and the historical negative sample dataset containing non-flood scene images, and before training YOLO V8 using the historical target dataset of flood images containing people and vehicles and the historical negative sample dataset containing non-flood scene images, the method further includes: using the Labelme tool to select water areas in the images of the historical water body dataset of flood images and saving the water sample data; using the Labelme tool to select river, lake, and pond water areas in the historical negative sample dataset of non-flood scene images and saving them as water body negative sample data; using the Labelme tool to select human and vehicle targets in the historical target dataset of flood images containing people and vehicles and saving the human and vehicle target sample data; and using the Labelme tool to select non-human and non-vehicle targets in the historical negative sample dataset of non-flood scene images and saving them as negative sample data.

[0016] Preferably, training the deep learning model ViT using a historical dataset of positive samples containing flood scene images and a historical dataset of negative samples containing non-flood scene images includes the following steps:

[0017] The historical dataset of positive samples of flood scene images and the historical dataset of negative samples containing non-flood scene images were randomly shuffled in a 1:1 ratio and divided into training set and validation set.

[0018] The data in the training set is input into the ViT model. The ViT model is used to divide the input image into image patches, and the relationship between the image patches is captured through the self-attention operation. The global features of the image patches are extracted, and the image is classified according to the global features.

[0019] At the end of each training cycle, the ViT model is evaluated using a validation set, and the classification performance is evaluated using cross-entropy loss and classification accuracy. The trained model parameters are then saved to obtain the trained ViT.

[0020] Preferably, the step of training DeepLab V3+ using a historical dataset of flood images and a historical dataset of negative samples containing images of non-flood scenes includes the following steps:

[0021] The historical water body dataset of flood images selected by the Labelme tool and the historical negative sample dataset containing non-flood scene images were randomly shuffled in a 1:1 ratio and divided into training set and validation set.

[0022] Input the data from the training set into DeepLab V3+ for training and learning;

[0023] At the end of each training cycle, DeepLab V3+ is evaluated using a validation set. The model classification performance is evaluated using cross-entropy loss and pixel accuracy metrics. The trained model parameters are saved to obtain the trained DeepLab V3+.

[0024] The process of training YOLO V8 using a historical target dataset of flood images containing people and vehicles and a negative sample historical dataset containing non-flood scene images includes the following steps:

[0025] The historical target dataset of flood images containing people and vehicles (selected by the Labelme tool) and the historical negative sample dataset containing non-flood scene images were randomly shuffled in a 1:1 ratio and divided into training set and validation set.

[0026] Input the data from the training set into YOLO V8 for training and learning;

[0027] At the end of each training cycle, the YOLO V8 model is evaluated using the validation set, and the model is evaluated using cross-entropy loss and classification accuracy. The trained model parameters are saved to obtain the trained YOLO V8 model.

[0028] This invention also provides an urban rainstorm and flood risk assessment system, comprising:

[0029] The dataset construction module is used to acquire real-time social media data of rainstorm and flood events in the target city. It constructs a real-time dataset of positive samples containing flood scene images, a real-time water body dataset of flood images, a real-time target dataset of flood images containing people and vehicles, and a real-time dataset of negative samples containing non-flood scene images using real-time social media data.

[0030] The target extraction module is used to input data from a real-time dataset containing positive samples of flood scene images and a real-time dataset containing negative samples of non-flood scene images into a pre-trained ViT dataset to filter out real-time rainstorm and flood image data, and then extract the address information of the real-time rainstorm and flood image data; input data from a real-time water body dataset of flood images and a real-time dataset containing negative samples of non-flood scene images into a pre-trained DeepLab V3+ dataset to extract water body objects; input data from a real-time target dataset of flood images containing people and vehicles and a real-time dataset containing negative samples of non-flood scene images into a pre-trained YOLO V8 dataset to extract people and vehicle targets; detect whether water body objects overlap with people and vehicle targets, and if they overlap, mark them as potentially dangerous images;

[0031] The risk assessment module combines address information from real-time rainstorm and flood images to synchronize the locations of all at-risk images to the map and generate a heat map; the heat map is then used to assess the risk areas of urban rainstorms and floods.

[0032] The present invention also provides a computer device, characterized in that it includes a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the urban rainstorm and flood risk assessment method.

[0033] The present invention also provides a computer-readable storage medium storing a computer program adapted for loading by a processor to execute the urban rainstorm and flood risk assessment method.

[0034] The urban rainstorm and flood risk assessment method, system, equipment, and medium provided by this invention have the following beneficial effects:

[0035] This invention acquires real-time social media data on rainstorm and flood events in target cities, providing more intuitive and objective disaster information for accurate flood risk assessment. It constructs a real-time dataset containing positive samples of flood scene images, a real-time water body dataset of flood images, a real-time target dataset of flood images containing people and vehicles, and a real-time negative sample dataset containing non-flood scene images. This process distinguishes disaster information into different data types, facilitating accurate assessment of urban rainstorm and flood risks. By processing the input data using pre-trained ViT, it can filter out real-time rainstorm and flood image data and extract address information, enabling more accurate location of flooding. By using pre-trained DeepLab V3+ to extract water body objects, it can accurately segment water bodies. Furthermore, by using pre-trained YOLO... V8 extracts human and vehicle targets, enabling effective detection of these targets. By detecting whether water objects overlap with human and vehicle targets, it can achieve multi-target identification and correlation analysis of water objects, human and vehicle targets, thus more accurately determining the location of potential danger. It synchronizes the location of potential danger corresponding to all potential danger images to the map and generates a heat map, which can accurately assess the risk areas of urban rainstorms and floods. Attached Figure Description

[0036] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of the urban rainstorm and flood risk assessment method according to an embodiment of the present invention;

[0038] Figure 2It is a framework for classifying, segmenting, and detecting objects in social media images using three deep learning models. Detailed Implementation

[0039] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0040] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0041] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of this invention, it should be noted that, unless otherwise explicitly specified or limited, the terms "connected" or "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances. In the description of this invention, unless otherwise stated, "a plurality of" means two or more, which will not be elaborated further here.

[0042] Example

[0043] This invention provides a method for assessing urban rainstorm and flood risk, specifically as follows: Figure 1 As shown, it includes the following steps:

[0044] Step 1: Obtain historical social media data related to urban rainstorms and floods, and use this data to pre-train deep learning models ViT (Vision Transformer), DeepLab V3+, and YOLO V8. This includes the following steps:

[0045] (1) Download historical social media data related to urban rainstorms and floods from social media platforms, and construct a positive sample historical dataset containing flood scene images, a historical water body dataset containing flood images, a historical target dataset containing people and vehicles, and a negative sample historical dataset containing non-flood scene images based on the historical social media data related to urban rainstorms and floods.

[0046] The social media platforms used were primarily from a certain website. The main method for downloading historical social media data related to urban rainstorms and floods was to enter the keywords "rainstorm" and "waterlogging" on the website and download the relevant historical social media data from the posted microblogs. Following a 1:1:1:1 image ratio, historical datasets were saved as follows: a positive sample dataset containing flood scene images, a historical water body dataset containing flood images, a historical target dataset containing flood images of people and vehicles, and a negative sample dataset containing natural rivers, lakes, or other irrelevant information.

[0047] (2) Pre-train the deep learning model ViT.

[0048] The historical dataset of positive samples of flood scene images and the historical dataset of negative samples containing non-flood scene images were randomly shuffled in a 1:1 ratio, and then divided into training set, validation set and test set in a 6:2:2 ratio.

[0049] The data in the training set is input into the ViT model. The ViT model is used to divide the input images into image patches, and the relationship between image patches is captured through self-attention operation. The global features of the image patches are extracted, and the images are classified according to the global features to filter out historical data of rainstorm and flood images.

[0050] At the end of each training cycle, the ViT model is evaluated using a validation set, and the classification performance is evaluated using cross-entropy loss and classification accuracy. The trained model parameters are then saved to obtain the trained ViT.

[0051] (9) Use the image annotation tool Labelme to annotate the data.

[0052] Use the Labelme tool to select water areas in historical flood image datasets and save the water sample data. Use the Labelme tool to select river, lake, and pond water areas in non-flood images and save them as negative water sample data. Use the Labelme tool to select human and vehicle targets in historical flood image target datasets that contain people and vehicles and save the human and vehicle target sample data. Use the Labelme tool to select non-human and non-vehicle targets in non-flood images and save them as negative sample data.

[0053] Samples are created using the Labelme tool for water bodies, people, and vehicles, respectively. These samples can then be input into DeepLab V3+ and YOLO V8 for learning and training.

[0054] (10) Pre-train DeepLab V3+.

[0055] The historical water body dataset of flood images selected by the Labelme tool and the historical negative sample dataset containing non-flood scene images were randomly shuffled in a 1:1 ratio and divided into training set, validation set and test set in a 6:2:2 ratio.

[0056] Input the data from the training set into DeepLab V3+ for training and learning;

[0057] At the end of each training cycle, DeepLab V3+ is evaluated using a validation set. The model classification performance is evaluated using cross-entropy loss and pixel accuracy metrics. The trained model parameters are saved to obtain the trained DeepLab V3+.

[0058] (5) Pre-train YOLO V8.

[0059] The historical target dataset of flood images containing people and vehicles and the historical negative sample dataset containing non-flood scene images, selected by the Labelme tool, were randomly shuffled in a 1:1 ratio and divided into training set, validation set and test set in a 6:2:2 ratio.

[0060] Input the data from the training set into YOLO V8 for training and learning;

[0061] At the end of each training cycle, the YOLO V8 model is evaluated using the validation set, and the model is evaluated using cross-entropy loss and classification accuracy. The trained model parameters are saved to obtain the trained YOLO V8 model.

[0062] Step 2: Acquire real-time social media data on rainstorm and flood events in the target city. Construct a real-time dataset containing positive samples of flood scene images, a real-time water body dataset of flood images, a real-time target dataset of flood images containing people and vehicles, and a real-time negative sample dataset containing non-flood scene images. Input the data from the positive and negative sample datasets into a pre-trained ViT dataset to filter out real-time rainstorm and flood image data; then extract the address information from the real-time rainstorm and flood image data. Input the data from the real-time water body dataset and the negative sample dataset into a pre-trained DeepLab V3+ dataset to extract water body objects. Input the data from the real-time target dataset of flood images containing people and vehicles and the negative sample dataset into a pre-trained YOLO V8 dataset to extract people and vehicle targets. Detect whether water body objects overlap with people and vehicle targets; if they overlap, mark them as potentially dangerous images. (Details follow...) Figure 2 As shown, it includes the following steps:

[0063] (11) Use ViT to filter out real-time data of rainstorm and flood images.

[0064] Download real-time social media data on rainstorm and flood events in the target city based on keywords, and use a trained ViT to filter images related to rainstorms and floods.

[0065] First, the keywords used are city-related. For example, if a city experiences a rainstorm or flood, the system downloads real-time text and images from Weibo, using keywords like "XX rainstorm," where "XX" is the city name. Based on web crawlers and the Weibo API, only Weibo messages containing both text and images are downloaded.

[0066] Secondly, real-time social media data on rainstorm and flood events in the target city are input into the trained ViT. For a piece of social media data with an image, if an image is detected as related to rainstorm and flood, the image and the Weibo text will be retained; otherwise, the Weibo text and the corresponding image will be discarded.

[0067] (2) Use the large language model ChatGLM3-6B to extract place name and address information from real-time rainstorm and flood images.

[0068] The ChatGLM3-6B large language model was used to detect address information from real-time rainstorm and flood image data. Only images with corresponding text containing address information were retained. Then, the Baidu coordinate picking system was used to convert the place name address information into latitude and longitude coordinates.

[0069] The URLs of the Weibo posts corresponding to the preserved flood-related images are parsed. The task instruction submitted to ChatGLM3-6B is: "Please identify and extract all specific place name and address information from the following Weibo post text." The large language model ChatGLM3-6B will return the specific place name and address information in JSON format. If no place name or address information can be detected from the text, the Weibo post text and its corresponding image are discarded.

[0070] This invention uses the Baidu Coordinate Picker system to convert place name and address information into latitude and longitude coordinates. The Baidu Coordinate Picker system provides two functions: converting addresses to latitude and longitude coordinates and vice versa. This invention inputs the parsed place name and address information into the Baidu Coordinate Picker system, which automatically extracts the converted latitude and longitude coordinate data.

[0071] (3) Use the trained DeepLab V3+ to extract water objects from real-time rainstorm and flood image data; use the trained YOLO V8 to extract people and vehicle targets from real-time rainstorm and flood image data.

[0072] Input real-time data of rainstorm and flood images into the trained DeepLab V3+, and DeepLab V3+ will automatically identify the pixels where the water bodies are located in the real-time data of rainstorm and flood images.

[0073] By inputting real-time data of rainstorm and flood images into the trained YOLO V8, YOLO V8 can use rectangular bounding boxes to identify people and vehicles.

[0074] (4) Detect whether the water object overlaps with the people and vehicles. If they overlap, mark it as a potentially dangerous image.

[0075] The process detects whether the outlines of people and vehicles overlap with water objects. If they do, the image is marked as potentially dangerous. The location of the danger is then confirmed using address information stored in the corresponding text. This process primarily checks if the people and vehicles are in water. If they are not, the image is not considered dangerous. The main criterion is whether there are water pixels within the outline of the people and vehicles. If the water pixels extracted by DeepLab V3+ do not appear inside the outline, it indicates that the people and vehicles are likely located on the roadside, not submerged, and therefore the image is not considered dangerous. Similarly, if the Weibo post and all its corresponding images are not marked as dangerous, the post is discarded.

[0076] YOLO V8 uses a rectangular frame to mark the vehicle target detection results. This invention further expands this rectangular frame by 1 / 2 and 1 times its original size. For the three different sizes of rectangular frames, the proportion of water pixels in the rectangular frame to the total pixels of the rectangular frame is counted. A proportion of more than 10% is marked as a vehicle involved in danger.

[0077] Step 3: Synchronize the locations of all the hazardous images to the map and generate a heat map; use the heat map to assess the risk areas of urban rainstorms and floods.

[0078] The system detects all images of potential urban flooding hazards and displays them on a map. It also uses heat maps to assess urban flooding hotspots and evaluate the risk of urban flooding.

[0079] The location information (latitude and longitude) of place names and addresses extracted from the corresponding Weibo posts for all affected images was displayed on a map. Then, a heat map was automatically generated using ArcGIS heatmap analysis. Hotspot locations on the heat map were marked as areas with higher flood risk and were key areas of focus for emergency relief and disaster relief. This heat map is a dynamic process; over time, some previously inactive locations will become hotspots, and vice versa, thus illustrating the dynamic migration of flood risk.

[0080] This invention also provides an urban rainstorm and flood risk assessment system, comprising:

[0081] The dataset construction module is used to acquire real-time social media data of rainstorm and flood events in the target city. It constructs a real-time dataset of positive samples containing flood scene images, a real-time water body dataset of flood images, a real-time target dataset of flood images containing people and vehicles, and a real-time dataset of negative samples containing non-flood scene images using real-time social media data.

[0082] The target extraction module is used to input data from a real-time dataset containing positive samples of flood scene images and a real-time dataset containing negative samples of non-flood scene images into a pre-trained ViT dataset to filter out real-time rainstorm and flood image data, and then extract the address information of the real-time rainstorm and flood image data; input data from a real-time water body dataset of flood images and a real-time dataset containing negative samples of non-flood scene images into a pre-trained DeepLab V3+ dataset to extract water body objects; input data from a real-time target dataset of flood images containing people and vehicles and a real-time dataset containing negative samples of non-flood scene images into a pre-trained YOLO V8 dataset to extract people and vehicle targets; detect whether water body objects overlap with people and vehicle targets, and if they overlap, mark them as potentially dangerous images;

[0083] The risk assessment module combines address information from real-time rainstorm and flood images to synchronize the locations of all at-risk images to the map and generate a heat map; the heat map is then used to assess the risk areas of urban rainstorms and floods.

[0084] The present invention also provides a computer device, characterized in that it includes a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to perform an urban rainstorm and flood risk assessment method.

[0085] The present invention also provides a computer-readable storage medium storing a computer program adapted for loading by a processor to execute an urban rainstorm and flood risk assessment method.

[0086] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited thereto. Any simple changes or equivalent substitutions of the technical solutions that can be obviously obtained by those skilled in the art within the scope of the technology disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims

1. A method for assessing urban rainstorm and flood risk, characterized in that, Includes the following steps: Acquire real-time social media data on rainstorm and flood events in the target city, and construct a real-time dataset of positive samples containing flood scene images, a real-time water body dataset of flood images, a real-time target dataset of flood images containing people and vehicles, and a real-time dataset of negative samples containing non-flood scene images using the real-time social media data. The data from the real-time dataset containing positive samples of flood scene images and the real-time dataset containing negative samples of non-flood scene images are input into the pre-trained ViT. The real-time data of rainstorm and flood images are then filtered out, and the address information of the real-time data of rainstorm and flood images is extracted. Data from a real-time flood image dataset and a real-time negative sample dataset containing non-flood scene images are input into a pre-trained DeepLab V3+ to extract water objects. Data from a real-time flood image dataset containing people and vehicles and a real-time negative sample dataset containing non-flood scene images are input into a pre-trained YOLO V8 to extract people and vehicle targets. The system detects whether water objects overlap with people and vehicle targets; if they overlap, the images are marked as potentially dangerous. By combining the address information from real-time rainstorm and flood images, the locations of all potential hazards corresponding to the images are synchronized to the map, and a heat map is generated. Heat maps are used to assess areas at risk of urban rainstorms and flooding.

2. The urban rainstorm and flood risk assessment method according to claim 1, characterized in that, The process of extracting the address information of real-time rainstorm and flood image data includes the following steps: The ChatGLM3-6B large language model was used to extract place name and address information from real-time data of rainstorm and flood images, and only the place name and address information was retained. Use Baidu's coordinate picking system to convert place name and address information into latitude and longitude coordinate data.

3. The urban rainstorm and flood risk assessment method according to claim 1, characterized in that, The detection method checks whether the water body object overlaps with the people and vehicles. If they overlap, the image is marked as potentially dangerous. Specifically, it checks whether there are water pixels within the outer frame of the people and vehicles. If the water pixels extracted by DeepLab V3+ do not appear inside the rectangular frame, the people and vehicles are located on the roadside, not submerged, and the image containing them is not a potentially dangerous photo. If the water pixels extracted by DeepLab V3+ appear inside the rectangular frame, the rectangular frame is expanded by 1 / 2 and 1 times, respectively. The percentage of water pixels within the three different sizes of the rectangular frame is counted. When the percentage reaches 10% or more, the image is marked as potentially dangerous.

4. The urban rainstorm and flood risk assessment method according to claim 1, characterized in that, The pre-training process for ViT, DeepLabV3+, and YOLO V8 includes the following steps: Acquire historical social media data related to urban rainstorms and floods, and construct a positive sample historical dataset containing flood scene images, a historical water body dataset containing flood images, a historical target dataset containing people and vehicles containing flood images, and a negative sample historical dataset containing non-flood scene images based on the historical social media data related to urban rainstorms and floods. The deep learning model ViT was trained using a historical dataset of positive samples containing flood scene images and a negative sample dataset containing non-flood scene images; DeepLab V3+ was trained using a historical dataset of flood scene images of water bodies and a negative sample dataset containing non-flood scene images; and YOLO V8 was trained using a historical dataset of flood scene images of targets containing people and vehicles and a historical dataset containing negative sample images of non-flood scene images.

5. The urban rainstorm and flood risk assessment method according to claim 4, characterized in that, Before training DeepLab V3+ using a historical water body dataset of flood images and a historical negative sample dataset containing non-flood scene images, and before training YOLO V8 using a historical target dataset of flood images containing people and vehicles and a historical negative sample dataset containing non-flood scene images, the process includes: using the Labelme tool to select water areas in the images of the historical water body dataset of flood images and saving the water sample data; using the Labelme tool to select river, lake, and pond water areas in the historical negative sample dataset of non-flood scene images and saving them as water body negative sample data; using the Labelme tool to select human and vehicle targets in the historical target dataset of flood images containing people and vehicles and saving the human and vehicle target sample data; and using the Labelme tool to select non-human and non-vehicle targets in the historical negative sample dataset of non-flood scene images and saving them as negative sample data.

6. The urban rainstorm and flood risk assessment method according to claim 4, characterized in that, The process of training the deep learning model ViT using a historical dataset containing positive samples of flood scene images and a historical dataset containing negative samples of non-flood scene images includes the following steps: The historical dataset of positive samples of flood scene images and the historical dataset of negative samples containing non-flood scene images were randomly shuffled in a 1:1 ratio and divided into training set and validation set. The data in the training set is input into the ViT model. The ViT model is used to divide the input image into image patches, and the relationship between the image patches is captured through the self-attention operation. The global features of the image patches are extracted, and the image is classified according to the global features. At the end of each training cycle, the ViT model is evaluated using a validation set, and the classification performance is evaluated using cross-entropy loss and classification accuracy. The trained model parameters are then saved to obtain the trained ViT.

7. The urban rainstorm and flood risk assessment method according to claim 5, characterized in that, The process of training DeepLab V3+ using a historical dataset of flood images and a historical dataset of negative samples containing images of non-flood scenes includes the following steps: The historical water body dataset of flood images selected by the Labelme tool and the historical negative sample dataset containing non-flood scene images were randomly shuffled in a 1:1 ratio and divided into training set and validation set. Input the data from the training set into DeepLab V3+ for training and learning; At the end of each training cycle, DeepLab V3+ is evaluated using a validation set. The model classification performance is evaluated using cross-entropy loss and pixel accuracy metrics. The trained model parameters are saved to obtain the trained DeepLab V3+. The process of training YOLO V8 using a historical target dataset of flood images containing people and vehicles and a negative sample historical dataset containing non-flood scene images includes the following steps: The historical target dataset of flood images containing people and vehicles (selected by the Labelme tool) and the historical negative sample dataset containing non-flood scene images were randomly shuffled in a 1:1 ratio and divided into training set and validation set. Input the data from the training set into YOLO V8 for training and learning; At the end of each training cycle, the YOLO V8 model is evaluated using the validation set, and the model is evaluated using cross-entropy loss and classification accuracy. The trained model parameters are saved to obtain the trained YOLO V8 model.

8. A city rainstorm and flood risk assessment system, characterized in that, include: The dataset construction module is used to acquire real-time social media data of rainstorm and flood events in the target city. It constructs a real-time dataset of positive samples containing flood scene images, a real-time water body dataset of flood images, a real-time target dataset of flood images containing people and vehicles, and a real-time dataset of negative samples containing non-flood scene images using real-time social media data. The target extraction module is used to input data from the real-time dataset containing positive samples of flood scene images and the real-time dataset containing negative samples of non-flood scene images into the pre-trained ViT, filter out the real-time data of rainstorm and flood images, and then extract the address information of the real-time data of rainstorm and flood images. Data from a real-time flood image dataset and a real-time negative sample dataset containing non-flood scene images are input into a pre-trained DeepLab V3+ to extract water objects. Data from a real-time flood image dataset containing people and vehicles and a real-time negative sample dataset containing non-flood scene images are input into a pre-trained YOLO V8 to extract people and vehicle targets. The system detects whether water objects overlap with people and vehicle targets; if they overlap, the images are marked as potentially dangerous. The risk assessment module combines the address information from real-time rainstorm and flood images to synchronize the locations of all at-risk images to the map and generate a heat map. Heat maps are used to assess areas at risk of urban rainstorms and flooding.

9. A computer device, characterized in that, It includes a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to perform the urban rainstorm and flood risk assessment method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted for loading by a processor to execute the urban rainstorm and flood risk assessment method according to any one of claims 1-7.

Images