Urban floodwater depth detection method based on multi-scale target detection network

Abstract

The application discloses a city floodwater depth detection method based on a multi-scale target detection network, and belongs to the field of image recognition and deep learning; the method comprises the following steps: selecting a reference object for city floodwater depth detection; constructing a city floodwater area reference object target detection network; training a reference object target detection model; obtaining a reference object target detection result and calculating a reference object pixel height; and quantitatively calculating the floodwater depth according to the mapping relationship between the reference object pixel height change and the actual height. The method fully gives play to the advantages of deep learning in the field of image recognition and detection, designs a target detection network comprising a multi-branch stacking module and a feature pyramid structure, and solves the problems of weak multi-scale target feature extraction capability and poor feature fusion effect of a traditional target detection algorithm. The model detection precision is improved under the condition of keeping the model calculation complexity low, the floodwater depth calculation precision is improved, and technical possibility is provided for non-contact floodwater depth detection based on vision.

Description

Technical Field

[0001] This invention relates to the fields of target detection and deep learning technology, and in particular to a method for detecting the depth of urban floodwater based on a multi-scale target detection network. Background Technology

[0002] With global warming and accelerated urbanization, urban flooding caused by extreme rainfall events is becoming increasingly frequent and severe. Road flooding depth directly impacts traffic safety, drainage system efficiency, and the accuracy of disaster emergency response, making it a key monitoring indicator for urban safety management. In urban scenarios with complex infrastructure and spatial structures, the dynamic changes in water depth in flood-prone areas such as low-lying roads, bridge culverts, and underground spaces urgently require efficient, real-time, and automated detection technologies. In recent years, the densely deployed surveillance camera networks in urban areas have provided massive amounts of video image data, laying the data foundation for the development of computer vision-based non-contact water depth detection technology. Intelligent analysis of surveillance video images to extract characteristic changes in flooded scenes and convert them into actual water depth values ​​has become an important technological direction for addressing the challenges of urban flooding monitoring.

[0003] There are three main traditional methods for detecting urban flooding: manual on-site observation, fixed-point water level sensors, and visual detection methods based on traditional image processing. However, these methods have significant technical bottlenecks and limitations in practical applications: manual measurement relies on on-site observation by inspectors and is characterized by intuitive operation and low cost, but it is limited by high manpower input, low monitoring frequency, and high operational risks in extreme weather, making it difficult to achieve real-time and comprehensive monitoring of water depth; contact water level sensors can provide high-precision single-point measurement data, but they face challenges such as high equipment deployment costs, susceptibility to damage from water flow, and high maintenance difficulty, especially in large urban areas or complex terrain where it is difficult to achieve grid coverage and cannot capture the spatial correlation of water distribution; early visual detection methods based on traditional image processing attempted to achieve non-contact detection by analyzing features such as water surface reflection and edge contours, but because they rely on manually designed fixed features and thresholds, the detection accuracy drops significantly in complex scenarios such as changes in lighting, obstruction by debris, and water surface fluctuations, and it is difficult to establish a quantitative relationship between the degree of flooding and the actual water depth, severely restricting the practicality of the technology.

[0004] In recent years, the development of object detection technology has opened up new avenues for detecting urban flood depth: by detecting and locating reference objects in images, calculating their pixel heights, and establishing a mapping relationship between pixel heights and actual physical heights, changes in flood depth can be detected. Compared to traditional image processing visual detection methods that rely on manually designed features and thresholds, object detection technology, based on deep learning, possesses autonomous feature learning and extraction capabilities. It can autonomously learn multi-level visual features from a large number of urban scene images without human intervention, and uses a convolution-pooling-activation structure to fit complex functional relationships between features, autonomously completing the detection of reference objects in urban flood scenes. However, traditional object detection networks are limited by single-scale feature extraction and relatively simple feature fusion mechanisms, resulting in insufficient feature representation capabilities for multi-scale targets (reference objects with different visual sizes due to varying distances) in complex scenes. This leads to limited object detection accuracy and scale adaptability, and the inability to accurately fit the detection box to the reference object, thus affecting the accuracy of flood depth calculation. Summary of the Invention

[0005] Therefore, the purpose of this invention is to propose a method for detecting urban floodwater depth based on a multi-scale target detection network, in order to solve the problems of high detection cost, low efficiency, and low accuracy of existing methods.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for detecting urban floodwater depth based on a multi-scale target detection network includes the following specific steps:

[0008] Step 1: Selection of reference points for urban floodwater depth;

[0009] Within the floodwater detection area, to meet the multi-dimensional detection needs, two types of objects (traffic cones and shared bicycles) perpendicular to the ground and with relatively fixed heights were selected as reference objects for depth detection, and the actual height H of the reference objects was recorded. r The traffic cones selected are 70cm orange traffic cones that conform to the national standard GB / T 24725-2009; the shared bicycles selected are standard shared bicycles with the top of the handlebar stem as the height reference point and the vertical height of the handlebar center as 105cm.

[0010] Step 2: Construct a target detection network for reference objects in urban flood-prone areas;

[0011] This target detection network consists of a backbone feature extraction network, an enhanced feature extraction network, and a detection head.

[0012] Furthermore, the backbone feature extraction network consists of three shallow feature extraction modules and three deep feature extraction modules, used to perform preliminary feature extraction on the input image. The shallow feature extraction module consists of a 3×3 convolution, a batch normalization layer, and a SiLu activation function layer. During the shallow feature extraction process, the height and width of the feature map are continuously compressed and the number of channels is continuously expanded through layer-by-layer convolution operations, converting the shallow spatial information of the feature map into deep semantic information. The deep feature extraction module consists of a multi-branch stacked structure and a SiLu activation function layer. The multi-branch stacked structure contains three branches, which perform one, three, and five 3×3 convolution operations on the input feature map, respectively, and concatenate the convolution results according to the channel dimension to obtain a feature map that integrates multi-receptor field feature information.

[0013] Furthermore, the feature extraction network is strengthened by taking the output feature maps of the three deep feature modules in the backbone feature extraction network as input and performing deep feature extraction on the preliminary features generated by them. To address the weaknesses of traditional methods in multi-scale target feature extraction and feature fusion, a feature pyramid structure and a transition module are introduced into the enhanced feature extraction network. This architecture, composed of the feature pyramid structure and the transition module, achieves multi-scale feature extraction and fusion, outputting three feature maps containing deeper semantic information. The feature pyramid structure performs pooling downsampling on the input feature map at sizes of 5×5, 9×9, and 13×13, respectively. The multi-scale pooling results are then concatenated along the channel dimension to expand the receptive field of the detection network. The concatenated results are then subjected to 3×3 convolution and batch normalization. Finally, residual connections are used to concatenate the input feature map to alleviate network degradation. The transition module comprises two branches: pooling downsampling and convolutional downsampling. Pooling downsampling consists of a 2×2 max-pooling layer and a 3×3 convolutional layer, while convolutional downsampling consists of two 3×3 convolutional layers. The output feature maps from the two downsampling branches are then stacked to achieve feature fusion. The design of the feature pyramid and transition module expands the receptive field through multi-scale pooling, preserves the original feature information through residual connections, and enhances feature interaction through bi-branch downsampling. This enables cross-layer fusion of shallow high-resolution detail features and deep semantic abstract features, thus constructing a hierarchical and robust feature representation space.

[0014] Furthermore, the detection head consists of 3×3 convolutions and 1×1 convolutions, and its main function is to generate target detection boxes based on the feature maps output by the feature extraction network. Specifically, the three feature maps of different scales obtained after processing by the enhanced feature extraction network are input into three detection heads with the same structure but different parameter configurations. Each detection head first performs local feature aggregation on the input feature map through 3×3 convolutions, preserving the spatial location and contextual information of the target, and then uses 1×1 convolutions to adjust the feature channels, realizing the conversion of image features into detection parameters. The final output includes the detection result containing the target category probability and the coordinate information of the detection box.

[0015] Step 3: Train the reference target detection network model;

[0016] The training process of the reference target detection model in flooded areas mainly includes four stages: model pre-training, multi-source dataset construction, model training optimization, and performance testing.

[0017] Furthermore, the model pre-training steps are as follows: The backbone feature extraction network is pre-trained using the ImageNet large-scale general dataset to learn general visual features such as edges and textures, thereby improving the convergence efficiency of subsequent training. Further, the dataset construction steps include: collecting urban scene images containing traffic cones and shared bicycles through web crawling and on-site photography; constructing the dataset after data augmentation processing such as rotation, scaling, and cropping; finally, normalizing the images in the dataset and dividing them into training, validation, and test sets in a 9:1:1 ratio. Further, model training... During the process, CIoULoss was used as the localization loss function between the ground truth bounding box and the predicted bounding box, and CrossEntrophyLoss was used as the loss function for object category prediction. The network parameters were updated using the stochastic gradient descent algorithm, with an initial learning rate of 0.01, a momentum of 0.937 to accelerate convergence, a weight decay of 0.0005, and a minimum learning rate of 0.00001 to prevent overfitting. Furthermore, in the model testing phase, the model's detection accuracy, recall, and other performance metrics were evaluated using an independent test set, and the model's hyperparameters were fine-tuned based on the test results.

[0018] Step 4: Obtain the target detection results for the reference object and calculate the pixel height of the reference object;

[0019] Image sequences of reference objects in both dry and wet states within the same monitoring area are input into a trained reference object detection model. The model outputs a set of detection result images containing detection boxes. The optimized model possesses high-precision detection capabilities, achieving pixel-level alignment between the detection boxes and the edges of the reference objects. The pixel heights of the detection boxes are calculated separately: the height of the reference object in dry state is denoted as H1, and the height of the reference object in wet state is denoted as H2.

[0020] Step 5: Quantitative calculation of water depth;

[0021] Based on the data obtained from the above steps: the actual height H of the reference object. r Given the reference pixel height H1 under waterless conditions and the reference pixel height H2 under water conditions, the water depth can be calculated. The water depth D can be expressed as:

[0022]

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] This method abandons traditional contact-based water level sensor detection and adopts video image-based visual detection technology to quantify water depth by analyzing pixel height changes of fixed reference objects in the water. It effectively solves the problems of high equipment maintenance costs, poor deployment flexibility, and low real-time monitoring efficiency in traditional detection methods, significantly reducing infrastructure investment and labor costs for urban flood monitoring.

[0025] A feature pyramid structure and transition module are introduced into the object detection network architecture to construct a multi-level feature fusion mechanism to enhance object detection capabilities in complex scenes. By fusing shallow high-resolution features with deep semantic features, the network can accurately capture the visual feature changes of objects of different scales in flooded scenes, significantly improving the detection accuracy of targets with varying scales in complex scenes. It achieves pixel-level alignment between the detection box and the edge of the target reference object, thereby improving the detection accuracy of water depth and providing a reliable data foundation for accurately detecting the pixel height of the reference object. Attached Figure Description

[0026] Figure 1 A schematic diagram of the overall process of the urban floodwater depth detection method based on a multi-scale target detection network provided by the present invention;

[0027] Figure 2 This is a schematic diagram of the reference object detection network structure constructed in the urban flooding scenario according to the present invention;

[0028] Figure 3 This is a schematic diagram of the water depth calculation process of the present invention. Detailed Implementation

[0029] This invention primarily achieves urban floodwater depth detection based on multi-scale target detection. The specific implementation method of this invention will be described in detail below with reference to the accompanying drawings.

[0030] Specifically, the flowchart of the urban floodwater depth detection method based on multi-scale target detection is as follows: Figure 1As shown, the process includes the following four steps: S1: Selecting reference objects for urban floodwater depth detection. S2: Constructing a reference object target detection network for urban flood-prone areas. S3: Training the reference object target detection model. S4: Obtaining the reference object target detection results and calculating the reference object pixel height. S5: Quantitatively calculating the water depth based on the mapping relationship between the reference object pixel height change and the actual height.

[0031] For S1: Select a reference object for detecting the depth of urban floodwaters;

[0032] In urban floodwater depth measurement tasks, reference objects serve as benchmarks and scales for measuring water depth. Therefore, the selection of reference objects should meet the characteristics of fixed height, relatively fixed position, and perpendicularity to the ground, and should be objects relatively common in urban environments. In view of this, this invention selects traffic cones and shared bicycles as reference objects, both with fixed heights and common in cities. The actual height is denoted as H. r The height configurations for both are as follows: the traffic cones are 70cm orange traffic cones conforming to the national standard GB / T24725-2009; the shared bicycles are standard shared bicycles with a vertical height of 105cm from the top of the handlebar stem as the height reference point.

[0033] For S2: Construct a network model for detecting reference targets in urban flood-prone areas;

[0034] In this invention, the network architecture for detecting reference targets in urban flood-prone areas is as follows: Figure 2 As shown, the system can be broadly divided into three parts: the backbone feature extraction network, the enhanced feature extraction network, and the detection head. More specifically, the backbone feature extraction network comprises three shallow feature extraction modules and three deep feature extraction modules. The enhanced feature extraction network consists of three feature pyramid structures and transition modules, corresponding to the three deep feature extraction modules in the backbone network.

[0035] Shallow Feature Extraction Module: The image input to the reference target detection network is RGB three-channel data with a resolution of 640×640. As the input layer of the network, the shallow feature extraction module consists of a standard processing unit composed of a 3×3 convolution, a batch normalization layer, and a SiLu activation function layer. The three shallow modules use convolution kernel strides of 1, 2, and 2 respectively, with output channel numbers of 32, 64, and 64. The first module maintains spatial resolution through convolution with a stride of 1, achieving initial feature mapping; the latter two modules compress the image spatial size through downsampling operations with a stride of 2 (the feature map size is halved sequentially), simultaneously encoding basic visual features such as edges and textures into the channel dimension, providing basic feature input for subsequent deep feature extraction.

[0036] Deep Feature Extraction Modules: The three deep feature extraction modules have the same structure and parameter configuration, differing only in the shape of the input and output feature maps. The following explanation uses the first deep feature extraction module as an example to illustrate its structure and function: Figure 2 As shown in the lower left, this module takes the 160×160 resolution, 64-channel feature map output by the third shallow feature extraction module as input. It consists of a multi-branch stacked module, a batch normalization (BN) layer, and a SiLu activation function layer cascaded together. The multi-branch stacked module contains three parallel convolutional branches: the first branch downsamples the input feature map through one 3×3 convolution with a stride of 2, compressing the spatial resolution to 80×80 and extracting basic geometric features such as edges and corners; the second branch integrates mid-level visual features such as texture and shape through three consecutive 3×3 convolutions with a stride of 2 in a progressive downsampling process, outputting an 80×80 composite feature map; the third branch expands the receptive field while maintaining the original 160×160 resolution through five consecutive 3×3 convolutions with a stride of 1, capturing refined texture and local contextual features. The feature maps output by the three branches are then aligned in size and concatenated along the channel dimension to form a multi-scale feature representation that integrates the details of the original resolution and the semantic information of the downsampled data. Batch normalization and SiLu activation are then applied to enhance the robustness of the model. Three deep feature extraction modules, through progressively increasing convolutional complexity and downsampling stride, produce output feature map sizes of 80×80, 40×40, and 20×20, respectively, with the number of channels correspondingly expanded to 64, 256, and 512. This design constructs a feature processing approach of "detail preservation - mid-level abstraction - semantic enhancement" through a multi-branch parallel extraction mechanism. It not only preserves the subtle contour information of the reference object but also extracts the overall structural features of the reference object. Its channel splicing operation significantly increases feature diversity, providing rich multi-dimensional features for the subsequent cross-scale fusion of feature pyramids. From the underlying architecture, it improves the robustness and localization accuracy of multi-scale target detection in complex urban flood scenarios.

[0037] Characteristic pyramid structure: such as Figure 2 As shown in the middle, the output feature maps of the three deep feature extraction modules are respectively input into three feature pyramid modules with identical structures, with the first feature pyramid module (… Figure 2Taking the upper middle section as an example, the input feature map is first enhanced by a 3×3 convolution with a stride of 2. Then, it passes through a parallel three-branch pooling layer, performing 5×5, 9×9, and 13×13 pooling operations on the input feature map to extract features from multiple receptive fields. The feature maps from different receptive fields are concatenated along the channel dimension, and then fused and standardized by a 3×3 convolution and batch normalization. Finally, this module superimposes the input feature map and the processed output feature map through residual connections, enhancing multi-scale feature representation capabilities while preserving the original spatial information. On the one hand, cross-scale pooling operations allow a single feature location to simultaneously contain contextual information from different receptive fields, effectively solving the problem of insufficient multi-target feature extraction in traditional networks. On the other hand, the residual connection mechanism alleviates the degradation of deep networks, ensuring the information integrity of multi-scale features during cross-layer fusion.

[0038] Transition Module: The transition module takes the output feature map of the corresponding feature pyramid module as input and constructs a dual-branch differential downsampling architecture to enhance multi-scale feature fusion capabilities. The pooling downsampling branch uses a cascaded 2×2 max-pooling layer and a 3×3 convolutional layer. First, max-pooling is used to compress the spatial resolution, and then convolutional operations are used to extract abstract semantic information. The convolutional downsampling branch uses two consecutive 3×3 convolutional layers to gradually reduce the resolution while expanding the receptive field through stacked convolutional kernels, preserving detailed features such as edges and textures. The feature maps output by the two branches are stacked along the channel dimension after size alignment to form a fused feature that includes the semantic features of the pooling branch and the detailed features of the convolutional branch. This design, through a differential downsampling strategy, makes the transition module the core module for multi-scale feature fusion, effectively improving the detection accuracy and feature robustness of target references of different sizes in complex scenes, and solving the technical problem that traditional single downsampling methods cannot simultaneously achieve semantic abstraction and detail preservation.

[0039] Detection Heads: The three detection heads form the final layer of the entire reference target detection network, responsible for multi-scale feature integration and bounding box generation. Each detection head consists of a 3×3 convolution with a stride of 1 and a 1×1 convolution. The 3×3 convolution performs local feature aggregation on the input feature map, preserving the spatial location and contextual information of the target. Subsequently, the 1×1 convolution adjusts the feature channels, resulting in an output feature map containing prior bounding box information. This map is divided into seven channels, recording the target presence confidence, the probability that the target type is a traffic cone, the probability that the target type is a shared bicycle, the center coordinates of the prior bounding box, and the height and width of the prior bounding box. The generated prior bounding boxes undergo non-maximum suppression (NMS). An intersection-over-union (IoU) threshold is used to select the predicted bounding box with the highest confidence at each scale for each detection head, suppressing duplicate detections. The three detection heads are input with feature maps of different resolutions (80×80, 40×40, and 20×20) to form a multi-scale detection system: the high-resolution branch focuses on the localization of smaller and more distant targets, the medium and low-resolution branch is adapted to the localization and recognition of medium and large targets, and finally outputs a predicted bounding box containing target category, location and scale information.

[0040] For S3: Train the reference target detection model;

[0041] The core process of training a reference object detection model includes: pre-training the model backbone (backbone feature extraction network) on a publicly available large-scale dataset; collecting reference object image data in flood scenarios and performing data preprocessing and enhancement; training the model on the training set; and finally evaluating the model accuracy using an independent test set.

[0042] Model pre-training: To address the overfitting problem that may result from the large parameter size of the target detection network and the limited sample size of the flood scene reference dataset, the backbone feature extraction network of the model is pre-trained using the public dataset ImageNet. During the training process, the enhanced feature extraction part and the detection head of the model are frozen, and only the parameters of the backbone network are updated, so that it learns the low-level visual features common to cross-domains such as image edges, textures, and basic geometric shapes, forming initial network parameters with general feature extraction capabilities.

[0043] Data Acquisition: A combination of web crawling and on-site photography was used to acquire diverse image data covering varying light intensities, water depths, and distances to reference points in urban scenes. To enhance data diversity and model generalization ability, multi-dimensional data augmentation was performed on the original images, including geometric transformations such as random rotation, scaling, and center cropping; photometric transformations such as brightness and contrast adjustments and Gaussian noise addition; and the overlay of flood scene interference factors such as water surface reflection and raindrop blurring. Subsequently, the augmented images were manually annotated using the LabelImg tool. Minimum bounding rectangle detection boxes were defined for traffic cones (based on their top tips) and shared bicycles (based on the top of the handlebar stem), and converted to XML format, recording the coordinates and dimensions of the detection boxes.

[0044] Model Training: The labeled data was divided into training, validation, and test sets in a 9:1:1 ratio. The training set images were fed into the reference object detection model for training. During training, CIoULoss was used as the localization loss function between the ground truth bounding boxes and the predicted bounding boxes, and CrossEntrophyLoss was used as the loss function for object category prediction. The SGD stochastic gradient descent algorithm was used to update the network parameters, with an initial learning rate of 0.01, momentum of 0.937 to accelerate convergence, weight decay of 0.0005, and a minimum learning rate of 0.00001 to prevent overfitting. A freeze-thaw training strategy was adopted to speed up training and reduce the resources consumed during training. The training was frozen for 50 epochs, during which the parameters of the backbone feature extraction network were frozen, and only the parameters of the enhancement feature extraction network were updated. The training was unfrozen for 150 epochs, during which the parameters of the entire model network were updated.

[0045] Model accuracy evaluation: After obtaining the trained model, end-to-end performance verification of the trained model is performed using independent test set data. The model is evaluated by calculating the intersection-union ratio of the predicted bounding boxes and the ground truth bounding boxes, as well as the detection precision and recall. Based on the test results, the model hyperparameters and training strategies are further fine-tuned to achieve higher detection accuracy.

[0046] For S4: Obtain the target detection results of the reference object and calculate the pixel height of the reference object;

[0047] like Figure 3 As shown in the upper part, the monitoring video stream of the urban flood-prone area containing the reference object is divided into frames, and data is collected every 60 frames to obtain image data before and during the flood. The waterless image before the flood is denoted as t0, and the watery image during the flood is denoted as tn. The acquired images are input into the reference object target detection model to obtain the target detection results and calculate the pixel height of the reference object detection box at each time: H1Pixels when there is no water and H2Pixels when there is a flood.

[0048] For S5: Quantitatively calculate the water depth based on the mapping relationship between the pixel height change of the reference object and the actual height;

[0049] Based on the data obtained from the above steps: the actual height H of the reference object. r Given the reference pixel height H1 under waterless conditions and the reference pixel height H2 under water conditions, the water depth D is calculated using the following formula:

[0050]

[0051] The above specific embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all similar technical solutions obtained by equivalent substitution or equivalent transformation are within the protection scope of the present invention.

Claims

1. A method for detecting urban floodwater depth based on a multi-scale target detection network, characterized in that: This method includes the following specific steps: Step 1: Selection of reference object for depth detection; Within the floodwater detection area, to meet the multi-dimensional detection needs, traffic cones and shared bicycles, two objects perpendicular to the ground with relatively fixed heights, were selected as reference objects for depth detection, and their actual heights were recorded. ; Step 2: Construct a reference object target detection network; The urban flood-affected area reference target detection network consists of a backbone feature extraction network, an enhanced feature extraction network, and a detection head. The enhanced feature extraction network takes the output feature maps of three deep feature modules in the backbone feature extraction network as input. Through an architecture composed of a feature pyramid structure and transition modules, it achieves deeper feature extraction and outputs three feature maps containing deeper semantic information. Specifically, the feature pyramid structure performs pooling downsampling on the input feature maps at sizes of 5×5, 9×9, and 13×13, respectively. The multi-scale pooling results are then concatenated along the channel dimension to expand the receptive field of the detection network. Finally, 3×3 convolution and batch normalization operations are performed on the concatenated results. Finally, residual connections are used to concatenate the input feature maps to alleviate network degradation. The transition module includes two branches: pooling downsampling and convolutional downsampling. Pooling downsampling consists of a 2×2 max pooling layer and a 3×3 convolutional layer, while convolutional downsampling consists of two 3×3 convolutional layers. The output feature maps of the two downsampling branches are stacked to achieve feature fusion. The design of the feature pyramid and transition module expands the receptive field through multi-scale pooling, preserves the original feature information through residual connections, and enhances feature interaction through dual-branch downsampling. This achieves cross-layer fusion between shallow high-resolution detailed features and deep semantic abstract features, constructing a hierarchical and highly robust feature representation space. Step 3: Train the reference target detection network model; The training process for the reference target detection model in flood-prone areas includes four stages: model pre-training, multi-source data acquisition, model training optimization, and performance testing. Step 4: Obtain the target detection results and calculate the reference object pixel height; Image sequences of reference objects in both dry and flooded states within the same monitoring area are input into a trained flood-affected area reference object detection model. The model outputs a set of detection result images containing bounding boxes. The improved and trained model possesses high-precision detection capabilities, achieving pixel-level alignment between the bounding boxes and the edges of the reference objects. The pixel height of the bounding boxes is calculated: the height of the reference object in the dry state is denoted as... The height of the reference point when there is water is recorded as ; Step 5: Quantitative calculation of water depth; Based on the data obtained in step 4: the actual height of the reference object Reference object pixel height under waterless conditions Reference pixel height under water conditions The water depth D is represented as: 。 2. The urban floodwater depth detection method based on a multi-scale target detection network according to claim 1, characterized in that: In step 2, the backbone feature extraction network is used to perform preliminary feature extraction on the input image. It consists of three shallow feature extraction modules and three deep feature extraction modules. The shallow feature extraction module consists of a 3×3 convolution, a batch normalization layer, and a SiLu activation function layer. During the shallow feature extraction process, the height and width of the feature map are continuously compressed and the number of channels is continuously expanded through layer-by-layer convolution operations, thereby converting the shallow spatial information of the feature map into deep semantic information. The deep feature extraction module consists of a multi-branch stacked structure and a SiLu activation function layer. The multi-branch stacked structure contains three branches, which perform one, three, and five 3×3 convolution operations on the input feature map, respectively, and concatenate the convolution results according to the channel dimension to obtain a feature map that integrates multi-scale feature information. The detection head consists of 3×3 convolutions and 1×1 convolutions, generating target detection boxes based on the feature maps output by the feature extraction network. The three feature maps of different scales obtained after processing by the enhanced feature extraction network are input into three detection heads with identical structures. Each detection head first performs local feature aggregation on the input feature map through 3×3 convolutions to preserve the spatial location and contextual information of the target. Then, it uses 1×1 convolutions to adjust the feature channels, realizing the dimensional transformation of the detection parameters. The final output includes the detection result containing key information such as the target category probability and the center coordinates of the detection box.

3. The urban floodwater depth detection method based on a multi-scale target detection network according to claim 1, characterized in that: In step 3, the backbone feature extraction network is first pre-trained using the ImageNet large-scale general dataset to learn general visual features of edges and textures, thereby improving the convergence efficiency of subsequent training. During the data acquisition phase, urban scene images containing traffic cones and shared bicycles are collected through web crawling and on-site photography. After data augmentation processing such as rotation, scaling, and cropping, a dataset is constructed. During the model training phase, the data is divided into training, validation, and test sets in a 9:1:1 ratio. Finally, the detection accuracy and recall performance metrics of the model are evaluated using an independent test set, and the model hyperparameters are fine-tuned based on the test results.

4. The urban floodwater depth detection method based on a multi-scale target detection network according to claim 1, characterized in that: The backbone feature extraction network of the object detection model was pre-trained using the large-scale ImageNet dataset to generate pre-trained weights with basic image representation capabilities.

5. The urban floodwater depth detection method based on a multi-scale target detection network according to claim 1, characterized in that: The model training phase configuration is as follows: CIoULoss is used as the localization loss function between the ground truth bounding box and the predicted bounding box, and CrossEntrophyLoss is used as the loss function for object category prediction; the network parameters are updated using the stochastic gradient descent algorithm, with an initial learning rate of 0.01, momentum of 0.937 to accelerate convergence, weight decay of 0.0005, and a minimum learning rate of 0.00001 to prevent model overfitting; the training process adopts a freeze-thaw training strategy to accelerate the training speed. During the freeze training period, the parameters of the backbone feature extraction network are frozen, and only the parameters of the enhancement feature extraction network are updated; during the thaw training period, the parameters of the entire model network are updated.

6. The urban floodwater depth detection method based on a multi-scale target detection network according to claim 1, characterized in that: A feature pyramid structure is introduced to expand the model's receptive field and extract multi-scale features, enabling the model to detect targets of different sizes in complex urban backgrounds. At the same time, a transition module is introduced to achieve multi-scale feature fusion. Through bidirectional feature interaction, the semantic consistency of features at different resolutions is optimized, thereby improving the model's accuracy in detecting multi-scale targets and achieving pixel-level alignment between the detection box and the reference object.

7. The urban floodwater depth detection method based on a multi-scale target detection network according to claim 1, characterized in that: After obtaining the target detection results, the actual height of the reference object under normal waterless conditions and the corresponding original pixel height are recorded. The pixel height change of the same reference object during floods is extracted. By establishing the proportional relationship between the pixel height change of the reference object and the original pixel height, and combining it with the actual physical height of the reference object, the water depth is calculated proportionally.

Images