A method, system and device for quickly identifying a pond tail water direct discharge behavior

Abstract

The application relates to a method, system and device for quickly identifying pond tail water direct discharge behavior, and relates to the technical field of sewage discharge monitoring and identification. The method constructs a remote sensing map tile data download, processing and sample making system, and establishes a sample set. Then, on the basis of an original YOLO11 model, a multi-dimensional dynamic snake convolution module and a bottleneck layer are seamlessly introduced, the YOLO11 model is improved, the model can dynamically adapt to the target feature of the tail water direct discharge behavior from multiple directions, and therefore the ability of capturing and feature extraction of the tail water discharge behavior is strengthened. Finally, a perception enhancement loss is introduced to enhance the stability of the model in identifying tail water of different flow directions, and finally intelligent and rapid identification of pond tail water based on high-resolution remote sensing is realized. It can be seen that the application effectively improves the identification precision and reasoning efficiency of the pond tail water discharge behavior.

Description

Technical Field

[0001] This application relates to the field of wastewater discharge monitoring and identification technology, and in particular to a rapid identification method, system and equipment for direct discharge of pond tailwater. Background Technology

[0002] Marine aquaculture, as an important industry, is mainly distributed in coastal areas. Marine aquaculture is divided into sea-based and land-based methods, with land-based marine aquaculture primarily using artificial ponds. During the aquaculture process, metabolic waste gradually accumulates in the ponds. To maintain the cleanliness of the pond water, wastewater must be discharged regularly through outlets. However, the discharged wastewater contains abundant nitrogen, phosphorus, and other substances. If discharged directly without filtration, it will cause eutrophication of surrounding natural water bodies, leading to serious ecological pollution. Therefore, the rapid and large-scale identification of direct discharge of pond wastewater is of great significance for environmental protection, pollution control, and the sustainable development of the marine economy.

[0003] Currently, with the development of remote sensing technology, land cover identification technology based on remote sensing imagery is gradually becoming the mainstream method for studying human surface activities. Remote sensing data has advantages such as wide coverage, high monitoring frequency, and high-resolution capture of land cover features. It can be used for large-scale and rapid identification and location of direct wastewater discharge behavior, effectively making up for the shortcomings of traditional monitoring and identification technologies, such as limited identification range and poor timeliness, and providing important clues for the scientific treatment of pond wastewater and environmental supervision.

[0004] Deep learning technology, with its powerful automatic feature extraction capabilities and nonlinear mapping advantages, can accurately mine subtle features from massive remote sensing data, significantly improving the accuracy and generalization ability of target detection. Existing work has introduced deep learning technology into the field of remote sensing ground object detection, achieving preliminary results. The YOLO (you only look once) model is widely used in target detection tasks, becoming a key model in the field due to its excellent recognition accuracy and fast inference speed.

[0005] However, in the monitoring and identification of direct discharge of pond wastewater, this behavior presents unique surface characteristics. This is primarily because the flow direction of discharged wastewater extends from the pond into natural water bodies, influenced by factors such as the Coriolis effect, resulting in randomness and diversity. Faced with this unique characteristic, existing general target detection models cannot accurately and efficiently identify illegal pond wastewater discharge, and its identification still largely relies on manual intervention. Summary of the Invention

[0006] This application provides a rapid identification method, system, and device for direct discharge of pond tailwater, aiming to address the technical problems of low accuracy, low efficiency, and high reliance on manual intervention when applying existing pollution monitoring technologies to illegal pond tailwater discharge. By constructing a plug-and-play multidimensional dynamic serpentine convolution module, improving the YOLO11 model architecture, and introducing a perception-enhanced loss function to improve the model's stability in identifying tailwater flowing in different directions, intelligent and rapid identification of pond tailwater is ultimately achieved. This application can effectively improve the identification accuracy and inference efficiency of pond tailwater discharge behavior, providing technical support for pond pollution prevention monitoring and environmental supervision.

[0007] Firstly, this application provides a rapid method for identifying direct discharge of pond tailwater, including:

[0008] Based on the remote sensing image samples of the acquired sample area, feature processing is performed to determine vector point positioning data, and combined with the acquired map tiles, a sample dataset is established, which includes a training set and a validation set.

[0009] Data augmentation processing is performed based on the training set, and a perception enhancement mechanism is established;

[0010] A YOLO11 model is constructed, and an improved multidimensional fusion feature convolution module is established based on the target features of pre-identified tailwater direct discharge behavior. An improved model with enhanced feature extraction is constructed, and the multidimensional fusion feature convolution module includes a dynamic serpentine convolution layer.

[0011] The improved model is trained based on the enhanced training set, and the multi-dimensional fusion feature convolution module dynamically adapts to the target from different directions, fits and extracts the depth features of the target from different perspectives, and outputs the prediction results.

[0012] Based on the perception enhancement mechanism, loss analysis is performed on the prediction results and the validation set by identifying the enhancement loss function, and the parameters of the improved model are updated to obtain the trained pond tailwater identification model;

[0013] The pond tailwater identification model is used to quickly identify the direct discharge behavior of pond tailwater from the acquired target remote sensing data, and the identification results are obtained.

[0014] Optionally, feature processing is performed on the remote sensing image samples of the acquired sample area to determine vector point positioning data, and combined with the acquired map tiles to establish a sample dataset, including:

[0015] Remote sensing image samples are collected from the sample area, and the remote sensing image samples are preprocessed and updated.

[0016] Obtain and interpret location information;

[0017] Based on the interpreted positioning information, the sewage discharge in each of the remote sensing image samples is located, the sample points are delineated, and vector point positioning data of each remote sensing image sample is obtained. The vector point positioning data carries visual features that appear as traces of pond sewage discharge.

[0018] For the vector point positioning data, map tiles corresponding to the vector point positioning data are obtained from the map tile source based on a preset map downloader;

[0019] Target samples are collected on the map tiles, and targets are selected based on the visual features to establish a sample dataset;

[0020] In the sample dataset, the selected target of each sample is surrounded by a box as a label.

[0021] Optionally, data augmentation processing is performed based on the training set, and a perceptual augmentation mechanism is established, including:

[0022] Data augmentation techniques are used to augment each image sample and annotation in the training set, thereby updating the training set.

[0023] Based on the adaptive histogram equalization method, each sample in the training set is divided into blocks for histogram equalization to limit the contrast enhancement amplitude and update the training set.

[0024] Based on the analysis of the impact of short side orientation and shape deviation on model training during bounding box regression, an identification enhancement loss function is established.

[0025] A model is introduced to identify the stability of tailwater with different flow directions, and a perception enhancement mechanism is established by combining the identification enhancement loss function.

[0026] Optionally, a YOLO11 model is constructed, and an improved multi-dimensional fusion feature convolutional module is established based on the target features of the pre-identified tailwater discharge behavior to build an improved model with enhanced feature extraction, including:

[0027] Construct the original YOLO11 model including the C3k2 feature extraction module. The C3k2 feature extraction module includes the CBS module and the C3k module for feature convolution. The C3k module performs feature extraction based on the bottleneck layer Bottleneck.

[0028] Based on the acquired manually interpreted information, the characteristics of the target objects in the direct discharge behavior of tailwater are analyzed;

[0029] Based on the target object features, a multi-dimensional fusion feature convolution module is constructed using a dynamic serpentine convolutional layer;

[0030] In the original YOLO11 model, the CBS module of the bottleneck layer Bottleneck in the C3k module is updated using the multidimensional fusion feature convolution module to construct the Bottleneck_DSC module.

[0031] Based on the bottleneck layer Bottleneck and the Bottleneck_DSC module, the C3k2 feature extraction module in the original YOLO11 model is updated to obtain an improved model.

[0032] Optionally, a multi-dimensional fusion feature convolutional module can be constructed based on dynamic serpentine convolutional layers, including:

[0033] A parallel structure is adopted to establish serpentine dynamic convolutions and vertical CBS layers in different directions of input feature input, as dynamic serpentine convolution layers;

[0034] Based on the dynamic snake-shaped convolutional layer, a multi-dimensional fusion feature convolutional module is established by combining a fusion layer with multi-view depth features.

[0035] Optionally, the improved model is trained based on the enhanced training set, dynamically adapting to the target from different directions through the multi-dimensional fusion feature convolution module, fitting and extracting depth features of the target object from different viewpoints, and outputting prediction results, including:

[0036] The training set is input into the improved model for feature extraction to obtain initial features;

[0037] Dimensionality reduction is performed based on the initial features to obtain dimensionality-reduced features;

[0038] The dimensionality reduction features are input into the multidimensional fusion feature convolution module, which dynamically adapts to the target from different directions to extract the depth features of the direct discharge behavior of tailwater.

[0039] Based on the depth features, feature fusion is performed from different perspectives to obtain fused features;

[0040] Based on the fusion features, a prediction is made and the prediction result is output.

[0041] Optionally, based on the perceptual enhancement mechanism, loss analysis is performed on the prediction results and the validation set by identifying the enhancement loss function, including:

[0042] Based on the prediction results, prediction boxes are extracted, and labeled boxes are extracted based on the validation set;

[0043] Based on the perception enhancement mechanism, the intersection-union ratio of the predicted bounding box and the labeled bounding box is calculated to obtain the target loss;

[0044] Analyze the shape of the annotation box;

[0045] Based on the analyzed annotation shape, the center point offset is weighted by direction to calculate the shape constraint loss; and based on the aspect ratio of the annotation shape, the shape weighted distance loss is calculated.

[0046] Based on the recognition enhancement loss function, the final loss is calculated using the target loss, the shape constraint loss, and the shape-weighted distance loss.

[0047] Optionally, after obtaining the recognition results, the following may also be included:

[0048] Obtain verification samples, which include remote sensing images for verification, and each image in the verification samples carries a bounding box;

[0049] Extract the target bounding box from the recognition results, and determine the localization target and confidence level of the target bounding box;

[0050] For the localization target, the target boxes are sorted based on the confidence level, and a non-maximum suppression method is introduced to traverse and remove invalid target boxes, thus suppressing the target boxes;

[0051] Based on the target bounding box and the labeled bounding box of the validation sample, the prediction precision is determined by evaluating the proportion of actual positive samples to predicted positive samples, and the prediction recall is determined by evaluating the proportion of actual positive samples to the total actual positive samples.

[0052] Based on the prediction precision and the prediction recall, the model performance is evaluated using the mean precision method.

[0053] Secondly, this application provides a rapid identification system for direct discharge of pond tailwater, comprising:

[0054] The sample acquisition module is used to perform feature processing on the remote sensing image samples of the acquired sample area, determine the vector point positioning data, and combine the acquired map tiles to establish a sample dataset, which includes a training set and a validation set.

[0055] An enhancement module is used to perform data augmentation processing based on the training set and to establish a perceptual enhancement mechanism;

[0056] The model building module is used to build a YOLO11 model. Based on the target features of the pre-identified tailwater direct discharge behavior, an improved multi-dimensional fusion feature convolution module is established to build an improved model with enhanced feature extraction. The multi-dimensional fusion feature convolution module includes a dynamic serpentine convolution layer.

[0057] The model training module is used to train the improved model based on the enhanced training set, dynamically adapt to the target from different directions through the multi-dimensional fusion feature convolution module, fit and extract the depth features of the target from different perspectives, and output the prediction results.

[0058] The loss optimization module is used to perform loss analysis on the prediction results and the validation set by identifying the enhanced loss function based on the perception enhancement mechanism, update the parameters of the improved model, and obtain the trained pond tailwater identification model.

[0059] The behavior recognition module is used to quickly identify the direct discharge behavior of pond tailwater from the acquired target remote sensing data through the pond tailwater recognition model, and obtain the recognition result.

[0060] Thirdly, this application provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0061] Memory, used to store computer programs;

[0062] A processor, when executing a program stored in memory, implements the steps of a method for rapidly identifying direct discharge of pond tailwater as described in any of the first aspects.

[0063] In summary, this application first collects remote sensing images to analyze and locate the visual features of sewage effluent, delineates samples, and obtains corresponding map tiles to form a sample dataset. The sample dataset is then augmented to expand the sample size, and a perception enhancement mechanism is constructed to improve the loss function and establish a recognition enhancement loss function. Subsequently, an original YOLO11 model is constructed. Based on the target features of the effluent discharge, a dynamic serpentine convolutional layer is used to construct a plug-and-play multidimensional fusion feature convolutional module, and the bottleneck layer is improved. The improved multidimensional fusion feature convolutional module and bottleneck layer are used to improve the feature extraction module in the YOLO11 model, resulting in a feature-enhanced improved YOLO11 model. Next, the improved model is trained using the dataset. In the improved model, the multidimensional fusion feature convolutional module dynamically adapts to the target from different directions, extracting fusion features of the direct discharge behavior of effluent. Based on the fusion features, prediction results are output. The established perception enhancement mechanism compares the prediction results with the validation samples, analyzes the loss, and updates the parameters of the improved model until training is complete. The trained model is then used as the pond effluent recognition model. Finally, the pond tailwater identification model can intelligently and rapidly identify pond tailwater based on high-resolution remote sensing imagery. It is evident that the embodiments of this application effectively improve the identification accuracy and inference efficiency of pond tailwater discharge behavior, providing technical support for pond pollution prevention monitoring and environmental supervision, and addressing the shortcomings of existing pollution monitoring technologies in identifying pond tailwater discharge behavior. Attached Figure Description

[0064] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0065] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 A flowchart illustrating a method for rapidly identifying direct discharge behavior of pond tailwater provided in an embodiment of this application;

[0067] Figure 2 This is a flowchart illustrating the steps of a method for rapidly identifying direct discharge of pond tailwater according to an optional embodiment of this application.

[0068] Figure 3 This is a flowchart of a technology for identifying direct discharge of pond tailwater, provided as an example in this application.

[0069] Figure 4 This is a sample annotation example provided in this application;

[0070] Figure 5 This is a schematic diagram of the C3k2 module structure in an existing YOLO11 model provided as an example in this application;

[0071] Figure 6 This is a diagram of different convolutional neural network layer structures provided as an example in this application;

[0072] Figure 7 This is a schematic diagram of a multidimensional dynamic serpentine convolution module provided as an example in this application;

[0073] Figure 8 This is a schematic diagram of an improved bottleneck layer provided as an example in this application;

[0074] Figure 9 This is a flowchart of Shape-IoU calculation provided as an example in this application;

[0075] Figure 10 This application provides a roadmap for whole-view remote sensing image recognition technology, which is an example of the technology.

[0076] Figure 11 This is an example of the target recognition result provided in this application;

[0077] Figure 12A structural block diagram of a rapid identification system for direct discharge of pond tailwater provided in an embodiment of this application;

[0078] Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0080] To facilitate understanding of the embodiments of this application, further explanations and descriptions will be provided below in conjunction with the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of this application.

[0081] Figure 1 A flowchart illustrating a method for rapidly identifying direct discharge of pond tailwater according to an embodiment of this application is shown. The method may specifically include the following steps:

[0082] Step 110: Perform feature processing on the remote sensing image samples of the acquired sample area to determine vector point positioning data, and combine with the acquired map tiles to establish a sample dataset.

[0083] The sample dataset includes a training set and a validation set.

[0084] In this embodiment, each map tile in the established sample dataset is a labeled sample. The labels are mainly defined by boxes (such as squares, rectangles, etc.) to delineate the pond tailwater discharge area.

[0085] This embodiment first selects a sample area and obtains remote sensing image samples of that area through a corresponding data source. Preferably, high-resolution remote sensing images with a resolution of less than 1 meter are selected as samples. After a series of preprocessing steps, each sample reflects the optical reflectance characteristics of the ground features. Through feature analysis, the sewage effluent is located, and the vector point positioning data of the samples is initially obtained.

[0086] Then, using the vector point positioning data as a reference, map tiles are obtained from the corresponding data source. The target (i.e., the area where the pond's tailwater is discharged) is selected on the map tiles to form a sample dataset. The samples in the sample dataset are randomly shuffled and divided into training and validation sets according to a certain ratio.

[0087] Step 120: Perform data augmentation processing based on the training set and establish a perception enhancement mechanism.

[0088] In practice, to improve the model's generalization ability and adaptability to different scenarios, data augmentation is performed on the training set data to expand the training set. Augmentation processes include, but are not limited to, random cropping, horizontal flipping, vertical flipping, and random rotation. Data augmentation increases the diversity of training samples and can also simulate image changes in real-world scenarios to some extent.

[0089] To enhance the object perception capability and stability of subsequent YOLO11 models, this embodiment further improves the loss function used in existing YOLO11 models by introducing a recognition-enhanced loss function and establishing a perception enhancement mechanism. Preferably, the loss function used in this embodiment for the perception enhancement mechanism is the Shape-IoU loss function.

[0090] Step 130: Construct a YOLO11 model. Based on the target features of the pre-identified tailwater discharge behavior, establish an improved multi-dimensional fusion feature convolution module to construct an improved model with enhanced feature extraction.

[0091] The multidimensional fusion feature convolution module includes a dynamic serpentine convolutional layer.

[0092] Step 140: Train the improved model based on the enhanced training set, dynamically adapt to the target from different directions through the multi-dimensional fusion feature convolution module, fit and extract the depth features of the target from different perspectives, and output the prediction result.

[0093] Step 150: Based on the perception enhancement mechanism, loss analysis is performed on the prediction results and the validation set through the recognition enhancement loss function to update the parameters of the improved model and obtain the trained pond tailwater recognition model.

[0094] Steps 130-150 are described uniformly as follows:

[0095] In its implementation, the feature extraction capability of the existing YOLO11 model (or the original YOLO11 model) is provided by the original feature extraction module. However, the module responsible for feature extraction convolution in this module uses a fixed receptive field. Limited by this fixed receptive field and standard convolutional layers, the existing YOLO11 model struggles to effectively extract features from the complex and varied pond tailwater. Therefore, this embodiment improves the model structure based on the existing YOLO11 model.

[0096] Specifically, this embodiment establishes a YOLO11 model and, combined with human visual interpretation, analyzes the target features of direct wastewater discharge behavior to determine the direction for model improvement based on these features. Since wastewater discharge behavior is typically complex and elongated, this embodiment introduces a dynamic serpentine convolutional structure to better capture target features. The model is improved by modifying the convolutional layer sampling structure, using a dynamic serpentine convolutional layer to replace the original feature extraction module, forming a multi-dimensional fusion feature convolutional module. The improved dynamic serpentine convolutional structure can dynamically adapt to the target from multiple directions simultaneously, allowing the model to focus on the serpentine structure of the wastewater flow, thereby enhancing its ability to capture and extract features from wastewater discharge behavior. Through this improved model structure, this embodiment obtains an enhanced YOLO11 model with improved feature extraction, referred to as the improved model.

[0097] Next, the training set is input into the improved YOLO11 model to train the model. Specifically, the multi-dimensional fusion feature convolution module in the model extracts features from the input training set. These features are processed by a serpentine dynamic convolutional layer input from different directions. The model focuses on the serpentine structure exhibited by the tailwater flow and, through feature enhancement, fits the depth features of the target object. Prediction is then performed using the depth features, and the prediction result is output.

[0098] During model training, a perceptual enhancement mechanism is introduced to perform loss analysis on the model's prediction results. First, predicted bounding boxes are extracted from the prediction results, and labeled bounding boxes are extracted from the corresponding validation set. The Shape-IoU loss function is used to calculate multidimensional loss based on the predicted bounding boxes and labeled bounding boxes. The model parameters are optimized based on the loss until the model training is complete. The trained and improved model is then used as the pond tailwater identification model.

[0099] Step 160: The pond tailwater identification model is used to quickly identify the direct discharge behavior of pond tailwater from the acquired target remote sensing data, and the identification result is obtained.

[0100] In practice, after training, the model can use remote sensing images with a resolution better than 1 meter as target remote sensing data. The model can then be used to quickly identify the direct discharge of tailwater from coastal ponds and obtain the corresponding identification results.

[0101] As can be seen, this application proposes a plug-and-play dedicated multidimensional dynamic serpentine convolution module for surface targets with unique characteristics, such as direct discharge of pond tailwater, and improves the YOLO11 model, enabling the model to achieve fast and accurate target detection on a lightweight basis, providing technical support for the rapid identification of illegal discharge of pond tailwater.

[0102] Reference Figure 2This illustration shows a flowchart of a method for rapidly identifying direct discharge of pond tailwater according to an optional embodiment of this application. The method may specifically include the following steps:

[0103] Step 210: Based on the remote sensing image samples of the acquired sample area, perform feature processing to determine vector point positioning data, and combine with the acquired map tiles to establish a sample dataset.

[0104] The sample dataset includes a training set and a validation set.

[0105] In practical implementation, refer to Figure 3 The technical roadmap shown in this embodiment illustrates the rapid identification method for direct discharge behavior of pond tailwater based on the improved YOLO11 model, which can be applied to the field of wastewater discharge monitoring. The following is a detailed description of each step and embodiment.

[0106] In an optional embodiment, the above-mentioned feature processing based on the remote sensing image samples of the acquired sample area to determine vector point positioning data, and combined with the acquired map tiles, to establish a sample dataset, may specifically include: collecting remote sensing image samples from the sample area and preprocessing the remote sensing image samples to update the remote sensing image samples; obtaining interpreted positioning information; locating the sewage discharge in each of the remote sensing image samples based on the interpreted positioning information, delineating sample points, and obtaining vector point positioning data for each remote sensing image sample, wherein the vector point positioning data carries visual features that appear as traces of pond sewage discharge; for the vector point positioning data, obtaining the map tiles corresponding to the vector point positioning data from the map tile source based on a preset map downloader; collecting target samples on the map tiles, selecting targets based on the visual features, and establishing a sample dataset; wherein, the selected targets of each sample in the sample dataset are surrounded by boxes as annotations.

[0107] In practical implementation, high-resolution remote sensing images better than 1 meter can be collected for the sample area (e.g., the coastal area of ​​Zhanjiang City, Province D) as remote sensing image samples. Each sample undergoes preprocessing, including radiometric correction and topographic correction, to ensure it unbiasedly reflects the optical reflectance characteristics of ground features. Subsequently, the wastewater discharge for each sample is located. This can be done through manual visual interpretation to determine visual characteristics, which serve as the interpreted location information. Based on the interpreted location information, sample points are delineated, and preliminary vector point location data for each sample, i.e., the location data of the sample area, is obtained. (Refer to...) Figure 4 As shown, the main visual features are traces formed by the tailwater discharged from the pond eroding the coast or riverbank.

[0108] Next, map tiles (a type of image, also known as map tile image data) are obtained based on vector point positioning data. These map tiles are used to create the sample dataset. The main data sources for map tiles are Google Maps, Bing, Tianditu, and ESRI. When obtaining map tiles, firstly, a map downloader is built using Python, independently exploring the APIs (Application Programming Interfaces) of map service websites. Then, based on the vector point positioning data, the manually labeled outlet sample locations are determined, and map tiles corresponding to the geographic area and its eight surrounding adjacent areas are automatically searched and stitched together to a 768×768 pixel size. The `requests` package can be used to send an HTTP request to a specified URL, returning a response object. Calling the object's `content` method retrieves the response content, i.e., the map tile corresponding to the vector point positioning data. At this point, the map tile is of type Bytes. After conversion using the `BytesIO` function, it can be read using the `Image` library and saved as a PNG format, thus obtaining the map tile. This embodiment uses a map tile downloader to collect 18 levels of map tiles for the corresponding location and eight surrounding connected areas, achieving a spatial resolution of approximately 0.6 meters, while simultaneously meeting the operational requirements for monitoring pond tailwater discharge and the technical requirements for sample set production.

[0109] Then, target samples are collected on the corresponding map tile image. Targets are manually selected on the image, i.e., outlined on the image. Each target in the image can be completely enclosed with a bounding rectangle (e.g., ...). Figure 4 As shown in the figure, the final sample dataset is obtained.

[0110] The obtained sample dataset can be divided into a training set and a validation set in an 8:2 ratio.

[0111] Step 220: Perform data augmentation processing based on the training set and establish a perception enhancement mechanism.

[0112] Optionally, data augmentation processing based on the training set and the establishment of a perception enhancement mechanism may include: using data augmentation techniques to expand each image sample and annotation in the training set and updating the training set; dividing each sample in the training set into blocks for histogram equalization based on an adaptive histogram equalization method to limit the contrast enhancement magnitude and updating the training set; establishing a recognition enhancement loss function based on the analysis of the impact of short side direction and shape deviation on model training during bounding box regression; and introducing the stability of the model in recognizing tailwaters with different flow directions, combined with the recognition enhancement loss function, to establish a perception enhancement mechanism.

[0113] To improve the model's generalization ability and adaptability to different scenarios, data augmentation is performed on the training set, including but not limited to random cropping, horizontal flipping, vertical flipping, and random rotation. This increases the diversity of training samples and can simulate image changes in real-world scenarios to some extent. The data augmentation process is as follows:

[0114] Random cropping: Select images from the training set and randomly crop out sub-images of fixed size; Random flipping: Similarly, select images from the training set and flip them horizontally or vertically with a 50% probability; Random stitching: Randomly select images from the training set in groups of four. Establish four quadrants for the stitched image, and place the four images of each group into the four quadrants of the stitched image. The size of these four images will be adjusted according to the size of the stitched image. During the stitching process, the images may be randomly cropped and scaled to generate new combined images with different sizes and contents. The target labels (such as bounding boxes) in each original image will also be adjusted accordingly to fit the position and size of the new stitched image; Blur processing: Use a kernel of random size or a median filter kernel to blur the input image. Preferably, the parameters are set to p=0.01, blur limit=(3, 7); 5. Convert to grayscale: Convert the input RGB image to a grayscale image. Preferably, the parameters are set to p=0.01, method='weightedaverage', num_output channels=3.

[0115] In practical implementation, to address potential large-area contrast differences in images, the Contrast Limited Adaptive Histogram Equalization (CLAHE) method is introduced. This method primarily enhances the image by dividing it into small blocks, performing histogram equalization on each block, and limiting the magnitude of contrast enhancement during the equalization process. This avoids over-enhancing noise or details while improving the overall visual quality of the image. The histogram equalization parameters can be set to p=0.01, clip_limit=(1.0, 4.0), and tile_grid_size=(8, 8).

[0116] In existing technologies, most object detection models use the box-loss function, a key loss function in object detection used to measure the difference between the predicted and ground truth bounding boxes, thereby optimizing model parameters. Typically, the most commonly used metric to measure the matching degree between the labeled and predicted boxes is the Intersection over Union (IoU). The ratio of the intersection to the union of two bounding boxes is calculated using the following formula: .

[0117] in, It is the predicted bounding box, that is, the bounding box extracted from the model's prediction results; These are the labeled bounding boxes, such as the labels on the samples in the validation set. When the predicted box and the bounding box completely overlap... It equals 1 at this time. The loss function is calculated as follows: .

[0118] However, through An inherent drawback of bounding box regression is that it cannot accurately describe the positional relationship between the predicted and labeled boxes when the intersection of two boxes is 0. Therefore, the existing YOLOv11 introduces... A shape loss term was added to minimize the normalized distance between the center points of the predicted and labeled boxes, thereby enhancing stability and improving detection accuracy. However, The influence of the bounding box's shape and size on the regression process is still ignored. When the predicted box is not square, there are long and short sides, and the different directions of the short sides lead to... The values ​​also differ. The flow direction of the wastewater discharged from the pond mainly extends from the pond to the natural water body, and the predicted boxes generated by the model are almost all rectangular. At the same time, due to the influence of the Coriolis force and other factors, the flow direction has randomness and diversity, resulting in different short side directions of the predicted boxes, which will affect the stability of the bounding box regression process during model training.

[0119] To address the shortcomings of the loss functions used in the existing models, this embodiment improves box-loss by introducing the Shape-IoU loss function as a recognition enhancement loss function. It also comprehensively considers the impact of deviations in the direction and shape of the short side during bounding box regression on model training, and enhances the stability of the model in recognizing tailwaters with different flow directions, thus forming a perception enhancement mechanism.

[0120] Step 230: Construct the original YOLO11 model including the C3k2 feature extraction module. The C3k2 feature extraction module includes the CBS module and the C3k module for feature convolution. The C3k module performs feature extraction based on the bottleneck layer Bottleneck.

[0121] Step 240: Based on the acquired manually interpreted information, analyze the target characteristics of the direct discharge behavior of tailwater.

[0122] Step 250: Based on the target object features, construct a multi-dimensional fusion feature convolution module using a dynamic serpentine convolutional layer.

[0123] A unified explanation of steps 230-250 is provided below:

[0124] Among related technologies, YOLO11 was released in September 2024, introducing new components such as C3k2 and C2PSA, which help improve feature extraction and computation speed. While YOLO11's C3k2 feature extraction module is generally applicable to various vision tasks, its performance often lags behind dedicated models when dealing with targets with special features.

[0125] This embodiment mainly improves the C3k2 structure responsible for feature extraction in the YOLO11 model architecture, namely the C3k2 feature extraction module, such as... Figure 5 As shown, the C3k2 structure mainly includes the CBS module and the C3k module.

[0126] CBS follows a standard convolutional module design and can perform standardization and nonlinear transformations on input features. The processing procedure is as follows: .

[0127] The formula for the activation function is: It has smooth and non-monotonic properties, which helps to alleviate the gradient vanishing problem; It is a two-dimensional convolutional neural network responsible for feature extraction. The CBS module is widely used in the backbone network of YOLO11.

[0128] The feature extraction function of the C3k module is mainly completed by the CBS module in the bottleneck layer. By reducing the number of neurons, information is compressed and the quality of feature representation is improved. The feature extraction capability of YOLO11 is mainly implemented by the CBS module in the Bottleneck structure.

[0129] However, limited by the fixed receptive field of standard convolutional layers in the CBS module, YOLO11 struggles to effectively extract features from complex and variable pond wastewater. To address this issue, this embodiment improves upon the existing YOLO11 model. First, manually interpreted information is obtained to analyze the target features of direct wastewater discharge. The manually interpreted information reveals the following characteristics: ① tubular structures resembling narrow rivers; ② a darker color due to the polluted water; ③ direct wastewater discharge is mostly located on coastlines or riverbanks, exhibiting a significant contrast with the surrounding environment. Based on these characteristics, this embodiment improves the model by modifying the convolutional layer sampling structure.

[0130] Reference Figure 6 As shown, the convolutional layer sampling structure mainly consists of: a standard convolutional structure, which fixes the relationship between the surrounding 8 connected pixels; a dilated convolutional structure, which introduces "holes" in the sampling to effectively expand the receptive field, but its sampling structure remains fixed; a variable convolutional structure, which adds a certain offset to each sampling point to better adapt to the shape changes of the target object; and a dynamic serpentine convolutional structure, which adds directional constraints to the offset to better identify slender objects. It is evident that convolutional layers with fixed sampling structures work better for regularly shaped objects; however, wastewater discharge behavior is usually characterized by complex deformations and slender shapes. Therefore, this embodiment introduces a dynamic serpentine convolutional structure to enable the model to better capture target features.

[0131] A single dynamic snake convolutional layer can only capture features in one direction and cannot adapt to complex target structures. To overcome this problem, this embodiment establishes a multidimensional dynamic snake convolutional module, DySnakeConv, which adopts a plug-and-play approach and can be effectively applied to the YOLO11 model.

[0132] In a specific implementation, the above-mentioned multi-dimensional fusion feature convolution module based on dynamic serpentine convolution layers can specifically include: using a parallel structure to establish serpentine dynamic convolutions in different directions of input feature input and a vertical CBS layer as dynamic serpentine convolution layers; and based on the dynamic serpentine convolution layers, combined with a fusion layer of multi-view depth features, to establish a multi-dimensional fusion feature convolution module.

[0133] In practical implementation, the multidimensional dynamic snake convolution module DySnakeConv adopts a parallel structure and can be composed of snake dynamic convolution layers in at least two directions (i.e., horizontal and vertical) and a vertical CBS layer. (See reference...) Figure 7 As shown, the x-direction (horizontal) includes a dynamic snake-like convolution DSConv_x, and the y-direction (vertical) includes a dynamic snake-like convolution DSConv_y and a CBS layer. Input features can be input in different directions. Finally, the multi-view depth features are fused through a Concat operation. The complete structure is as follows:

[0134] .

[0135] It is evident that the improved dynamic serpentine convolutional structure can dynamically adapt to the target from multiple directions simultaneously, thereby enhancing the ability to capture and extract features from wastewater discharge behavior.

[0136] Step 260: In the original YOLO11 model, update the CBS module of the bottleneck layer Bottleneck in the C3k module using the multidimensional fusion feature convolution module, and construct the Bottleneck_DSC module.

[0137] Step 270: Based on the bottleneck layer Bottleneck and the Bottleneck_DSC module, update the C3k2 feature extraction module in the original YOLO11 model to obtain the improved model.

[0138] A unified explanation of steps 260-270 is provided below:

[0139] In this embodiment, DySnakeConv is used to replace the CBS module of the bottleneck layer Bottleneck in C3k2, forming the Bottleneck_DSC module, as shown in the following formula:

[0140] .

[0141] The improved bottleneck layer structure is as follows: Figure 8 As shown.

[0142] Finally, the original bottleneck layer and the improved bottleneck layer Bottleneck_DSC are used to replace the C3k module in C3k2 to obtain the final YOLO11 model with enhanced feature extraction, namely the YOLO11_DSC model (hereinafter referred to as YOLO11_DSC to represent the improved YOLO11 model).

[0143] In summary, in this embodiment, the CBS module with a 1×1 convolution kernel size in the Bottleneck_DSC module reduces the feature dimension, which can reduce the model complexity to a certain extent. At the same time, by limiting the information flow, the model can focus more on the important feature information of the target and improve the generalization performance of the model.

[0144] In the CBS and DySnakeConv modules, the CBS module has a 3×3 kernel size, and the DySnakeConv module's CBS module and the deformable DSConv convolutions in different directions (X and Y directions) all have a 3×3 kernel size, serving as the core structure for feature extraction. In summary, Bottleneck_DSC employs a hierarchical feature extraction approach, enabling the model to extract deeply abstract target features layer by layer.

[0145] As can be seen, in this embodiment, the improved YOLO11_DSC abandons the redundant design of the C3k module in the original YOLO11. Its main and neck C3k2 modules are based on the improved bottleneck layer Bottleneck_DSC module and its DySnakeConv as the main structure for feature extraction. While ensuring the model's recognition accuracy, it effectively reduces the number of parameters and achieves a lightweight improvement.

[0146] Step 280: Train the improved model based on the enhanced training set, dynamically adapt to the target from different directions through the multi-dimensional fusion feature convolution module, fit and extract the depth features of the target from different perspectives, and output the prediction result.

[0147] In one optional embodiment, the improved model is trained based on the enhanced training set, and the multi-dimensional fusion feature convolution module dynamically adapts to the target from different directions to fit and extract depth features of the target from different perspectives, and outputs a prediction result. Specifically, this may include: inputting the training set into the improved model for feature extraction to obtain initial features; performing dimensionality reduction processing based on the initial features to obtain dimensionality-reduced features; inputting the dimensionality-reduced features into the multi-dimensional fusion feature convolution module, which dynamically adapts to the target from different directions to extract depth features of direct wastewater discharge behavior; fusing features from different perspectives based on the depth features to obtain fused features; and making a prediction based on the fused features to output a prediction result.

[0148] In the specific implementation, the training set of the YOLO11_DSC model is input, and features are extracted through the basic convolutional layer, namely the CBS module, to obtain preliminary features (see formula ③). Then, the features are input to the Bottleneck_DSC module, which reduces the feature dimensionality through the CBS module, achieving dimensionality reduction. The dimensionality-reduced features are then sequentially passed to the CBS module and the DySnakeConv module. Within the DySnakeConv module, the dimensionality-reduced features are input to the CBS module and deformable convolutions of DSConv in different directions (X and Y directions) to extract deep core features, obtaining deep features of the direct discharge behavior of wastewater. The three branches' deep features are concatenated through a Concat operation, outputting the fused multi-view deep features (see formulas ④ and ⑤), i.e., the fused features. Finally, prediction is performed using the multi-view deep features, and the prediction result is output.

[0149] In actual testing, refer to Table 1 below:

[0150] Table 1 Comparison of Model Parameters

[0151] .

[0152] As can be seen, the improved YOLO11_DSC model has fewer parameters, indicating that the model achieves lightweighting while improving recognition accuracy; at the same time, the GFLOPs are smaller than those of the original YOLO11, indicating that the improved model is faster in forward propagation (inference).

[0153] Step 290: Based on the perception enhancement mechanism, loss analysis is performed on the prediction results and the validation set through the recognition enhancement loss function to update the parameters of the improved model and obtain the trained pond tailwater recognition model.

[0154] Optionally, the loss analysis of the prediction result and the validation set based on the aforementioned perceptual enhancement mechanism by identifying the enhancement loss function may include the following sub-steps:

[0155] Sub-step 2901: Based on the prediction results, extract the prediction bounding boxes, and based on the validation set, extract the labeled bounding boxes.

[0156] Sub-step 2902: Based on the perception enhancement mechanism, calculate the intersection-union ratio of the predicted bounding box and the labeled bounding box to obtain the target loss.

[0157] Sub-step 2903: Analyze the shape of the annotation box.

[0158] Sub-step 2904: Based on the analyzed annotation shape of the annotation box, perform directional weighting processing on the center point offset to calculate the shape constraint loss; and perform aspect ratio processing on the annotation shape to calculate the shape weighted distance loss.

[0159] Sub-step 2905: Based on the recognition enhancement loss function, calculate the final loss using the target loss, the shape constraint loss, and the shape-weighted distance loss.

[0160] A unified explanation is provided for sub-steps 2901-2905:

[0161] In practical implementation, the Shape-IoU loss function is introduced to optimize the parameters of the YOLO11_DSC model. (Refer to...) Figure 9 As shown, firstly, based on the predicted bounding boxes and the corresponding labeled bounding boxes in the validation set, the intersection-union ratio (IU) between the predicted and labeled bounding boxes is calculated, and then the loss function is further calculated. (See Formula ①). Then, the shape of the bounding box is analyzed. Based on the shape, the shape constraint loss is calculated, and the center point offset is weighted by direction, thus imposing a greater penalty on offsets in the short side direction, causing the predicted box center point to converge in the correct direction. Based on the bounding box shape, the shape-weighted distance loss (width-height difference) is calculated, making the aspect ratio of the predicted box closer to the true target. Finally, the results are combined and summed. Shape constraint loss and shape-weighted distance loss are used to calculate the Shape-IoU loss using the following formula: .

[0162] in, This is the Shape-IoU loss. Shape-weighted distance loss, This represents the shape constraint loss.

[0163] For targets with elongated or flat bounding boxes, the Shape-IoU shape-weighted mechanism can better adapt to their geometric characteristics and reduce shape distortion. Because it has a more reasonable model of the geometric properties of the bounding box, the loss function can provide more accurate gradient signals, which helps the training process converge more stably.

[0164] Step 300: The pond tailwater identification model is used to quickly identify the direct discharge behavior of pond tailwater from the acquired target remote sensing data, and the identification result is obtained.

[0165] In practical testing, this embodiment can apply the improved model to RGB optical remote sensing imagery with a spatial resolution of 0.5 meters in the coastal area of ​​Zhanjiang City, Province D. The technical approach is as follows: Figure 10 As shown, specifically, because the entire remote sensing image has too many pixels to be directly input into the neural network, the remote sensing image is first cropped into 768×768 slices from left to right and top to bottom. The slice cropping step size is set to 192 to retain some overlapping areas between slices, ensuring the integrity of the target during the slicing process. Then, the slices are sequentially input into the model as target remote sensing data, and the detection results are output. To preserve geographic information, the result detection boxes need to be converted from pixel coordinates to geographic coordinate vector patches. Finally, all slices are stitched together, and non-maximum suppression (NMS) is introduced to remove redundancy from the vector patches of all target boxes, thus obtaining the final recognition result. Figure 11 .

[0166] Optionally, after obtaining the recognition result, the method further includes: acquiring verification samples, wherein the verification samples include verification remote sensing images, and each image in the verification samples carries a bounding box; extracting target boxes from the recognition result and determining the location target and confidence level of the target boxes; for the location target, sorting the target boxes based on the confidence level, and introducing a non-maximum suppression method to traverse and remove invalid target boxes, thereby suppressing target boxes; based on the target boxes and the bounding boxes of the verification samples, determining the prediction precision by evaluating the proportion of actual positive samples to predicted positive samples, and determining the prediction recall by evaluating the proportion of actual positive samples to the total actual positive samples; and evaluating the model performance based on the prediction precision and the prediction recall by using the mean precision method.

[0167] In practical implementation, the obtained recognition results can be further verified visually using remote sensing images. Multi-dimensional accuracy verification can be performed using current mainstream evaluation index systems to analyze and evaluate model performance. The evaluation index system includes: Precision, Recall, mAP50, and mAP50-95. These are explained in detail below:

[0168] First, after the model predicts and identifies the target, many bounding boxes in the results will repeatedly locate the same target. These candidate bounding boxes have different aspect ratios, and each candidate bounding box is assigned a confidence score. This embodiment introduces the Non-Maximum Suppression (NMS) method to remove redundant bounding boxes (i.e., remove invalid predicted boxes). The basic steps are as follows: Sort all predicted boxes according to their scores (i.e., the probability of the target). Select the predicted box M with the highest score, and suppress other predicted boxes with high overlap with M (IoU exceeding a threshold). Iterate the above process until all predicted boxes have been traversed.

[0169] Next, we evaluate precision and recall:

[0170] Specifically, precision is calculated as the proportion of truly positive samples out of all samples predicted as positive by the model. It measures the percentage of accurately predicted samples and is used as a standard to evaluate the accuracy of predicting positive examples; hence, it's called prediction precision. The formula is: .

[0171] Among them, TP (True Positives) are true positives, which are predicted to be positive and are actually positive; FP (False Positives) are false positives, which are predicted to be positive but are actually negative.

[0172] Recall, on the other hand, is based on actual samples. It measures the proportion of correctly predicted positive examples out of all actual positive examples, assessing the coverage percentage of all actual positive examples that were predicted. The recall formula is:

[0173] .

[0174] FN (false negatives) are false negatives, which are predicted to be negative but are actually positive.

[0175] Next, the mean average precision (mAP) is introduced. In object detection tasks, mAP is a commonly used performance evaluation metric to measure the average precision of the model across different categories and IoU thresholds. This embodiment can use both mAP50 and mAP50-95 for model performance evaluation.

[0176] mAP50 calculates the average AP (Average Precision) across all classes, with AP calculated at an IoU threshold of 0.5. Unlike mAP50, mAP50-95 considers a broader range of IoU, allowing for the evaluation of model performance at different levels of overlap. The AP calculated by mAP50-95 is at 0.05 steps from the IoU threshold of 0.5 to 0.95, meaning that mAP50-95 provides a more comprehensive evaluation of model performance, taking into account model behavior at different IoU levels.

[0177] For example, the model training environment is as follows: Graphics Processing Unit (GPU) is NVIDIA RTX 3060; Central Processing Unit (CPU) is Intel Core i9-12900K. The hyperparameters during training are set as follows: 500 training epochs; Batch size is set to 8; Stochastic Gradient Descent (SGD) is used as the optimizer, with a momentum parameter of 0.937 and weight decay of 0.0005; Input image size is 768×768.

[0178] After model training, the weights obtained from the optimal accuracy round are used to perform recognition and detection on the validation set. Table 2 below shows a comparison of the accuracy of the improved YOLO11_DSC model with the original YOLO11 and YOLOv8 models:

[0179] Table 2 Comparison of Model Accuracy

[0180] .

[0181] As shown in Table 2, the improved YOLO11_DSC model achieves a precision of 79.9%, indicating high accuracy in target detection, meaning a high success rate in identifying targets. However, its recall is slightly lower than the original YOLO11, suggesting that the model may miss targets in certain situations. Furthermore, the improved model achieves better results in both mAP50 and mAP50-95, demonstrating that it can generate more realistic target detection boxes, meeting the high-precision positioning requirements of large-scale remote sensing target detection tasks.

[0182] In summary, this application implements a rapid identification method for direct discharge of pond tailwater based on an improved YOLO11 model, aiming to provide effective technical support for environmental supervision and scientific management of coastal aquaculture. This embodiment utilizes remote sensing satellites to acquire high-resolution images, obtaining multi-feature information such as the spectrum, texture, and shape of pond tailwater. Then, the YOLO11 model architecture is optimized by introducing a dynamic serpentine convolutional network at the bottleneck layer, enhancing the model's ability to perceive local pond tailwater morphology and extract global features. Finally, a pond tailwater identification model based on the improved YOLO11 model is constructed, achieving generalized prediction of tailwater across the entire scene. Non-maximum suppression processing is applied to the identification results to effectively reduce the impact of repeated bounding boxes on the results during whole-scene prediction, providing technical support for the rapid identification of illegal pond tailwater discharge and addressing the shortcomings of existing sewage monitoring technologies in identifying pond tailwater discharge behavior.

[0183] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of actions. However, those skilled in the art should know that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps may be performed in other orders or simultaneously.

[0184] like Figure 12 As shown in the figure, this application embodiment also provides a rapid identification system 1200 for direct discharge of pond tailwater, including:

[0185] The sample acquisition module 1210 is used to perform feature processing on the remote sensing image samples of the acquired sample area, determine the vector point positioning data, and combine the acquired map tiles to establish a sample dataset, which includes a training set and a validation set.

[0186] Enhancement module 1220 is used to perform data augmentation processing based on the training set and establish a perception enhancement mechanism;

[0187] The model building module 1230 is used to build a YOLO11 model. Based on the target features of the pre-identified tailwater direct discharge behavior, an improved multi-dimensional fusion feature convolution module is established to build an improved model with enhanced feature extraction. The multi-dimensional fusion feature convolution module includes a dynamic serpentine convolution layer.

[0188] The model training module 1240 is used to train the improved model based on the enhanced training set, dynamically adapt to the target from different directions through the multi-dimensional fusion feature convolution module, fit and extract the depth features of the target from different perspectives, and output the prediction results.

[0189] The loss optimization module 1250 is used to perform loss analysis on the prediction results and the validation set by identifying the enhanced loss function based on the perception enhancement mechanism, update the parameters of the improved model, and obtain a trained pond tailwater identification model.

[0190] The behavior recognition module 1260 is used to quickly identify the direct discharge behavior of pond tailwater from the acquired target remote sensing data through the pond tailwater recognition model, and obtain the recognition result.

[0191] It should be noted that the system provided in this application embodiment can execute the rapid identification method for direct discharge of pond tailwater provided in any embodiment of this application, and has the corresponding functions and beneficial effects of the execution method.

[0192] In practical implementation, the aforementioned rapid identification system for direct discharge of pond tailwater can be integrated into the device. This device can identify illegal discharge of pond tailwater based on input remote sensing images using an improved YOLO11 model. As an electronic device, it improves the identification accuracy and inference efficiency of pond tailwater discharge behavior, providing technical support for pond pollution prevention monitoring and environmental supervision. This electronic device can consist of two or more physical entities, or it can consist of a single physical entity. For example, the electronic device can be a personal computer (PC), a computer, a server, etc. This application embodiment does not impose specific limitations in this regard.

[0193] like Figure 13As shown, this application embodiment provides an electronic device, including a processor 111, a communication interface 112, a memory 113, and a communication bus 114. The processor 111, the communication interface 112, and the memory 113 communicate with each other through the communication bus 114. The memory 113 is used to store computer programs. When the processor 111 executes the program stored in the memory 113, it implements the steps of the rapid identification method for direct discharge of pond tailwater provided in any of the aforementioned method embodiments. For example, the method may include the following steps: performing feature processing on remote sensing image samples of the acquired sample area to determine vector point positioning data, and combining the acquired map tiles to establish a sample dataset, the sample dataset including a training set and a validation set; performing data augmentation processing on the training set and establishing a perception enhancement mechanism; constructing a YOLO11 model, establishing an improved multi-dimensional fusion feature convolution module based on the pre-identified target features of direct wastewater discharge behavior, and constructing an improved model with enhanced feature extraction, the multi-dimensional fusion feature convolution module including a dynamic serpentine convolution layer; training the improved model according to the enhanced training set, dynamically adapting to the target from different directions through the multi-dimensional fusion feature convolution module, fitting and extracting the target depth features from different perspectives, and outputting prediction results; performing loss analysis on the prediction results and the validation set through the recognition enhancement loss function according to the perception enhancement mechanism, updating the parameters of the improved model, and obtaining a trained pond wastewater recognition model; and using the pond wastewater recognition model to quickly identify the direct discharge behavior of pond wastewater from the acquired target remote sensing data, and obtaining the recognition result.

[0194] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the rapid identification method for direct discharge of pond tailwater as provided in any of the foregoing method embodiments.

[0195] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0196] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A rapid identification method for direct discharge of pond tailwater, characterized in that, include: Based on the remote sensing image samples of the acquired sample area, feature processing is performed to determine vector point positioning data, and combined with the acquired map tiles, a sample dataset is established, which includes a training set and a validation set. Data augmentation processing is performed based on the training set, and a perception enhancement mechanism is established; A YOLO11 model is constructed, and an improved multidimensional fusion feature convolution module is established based on the target features of pre-identified tailwater direct discharge behavior. An improved model with enhanced feature extraction is constructed, and the multidimensional fusion feature convolution module includes a dynamic serpentine convolution layer. The improved model is trained based on the enhanced training set, and the multi-dimensional fusion feature convolution module dynamically adapts to the target from different directions, fits and extracts the depth features of the target from different perspectives, and outputs the prediction results. Based on the perception enhancement mechanism, loss analysis is performed on the prediction results and the validation set by identifying the enhancement loss function, and the parameters of the improved model are updated to obtain the trained pond tailwater identification model; The pond tailwater identification model is used to quickly identify the direct discharge behavior of pond tailwater from the acquired target remote sensing data, and the identification results are obtained. The data augmentation process based on the training set and the establishment of a perception enhancement mechanism include: using data augmentation techniques to expand each image sample and annotation in the training set and updating the training set; using an adaptive histogram equalization method to divide each sample in the training set into blocks for histogram equalization, limiting the contrast enhancement amplitude, and updating the training set; establishing a recognition enhancement loss function based on the analysis of the impact of short side direction and shape deviation on model training during bounding box regression; and introducing the stability of the model in recognizing tailwaters with different flow directions, combined with the recognition enhancement loss function, to establish a perception enhancement mechanism. A YOLO11 model is constructed, and an improved multidimensional fusion feature convolutional module is established based on the pre-identified target features of direct tailwater discharge behavior. This results in an improved model with enhanced feature extraction, including: constructing an original YOLO11 model including a C3k2 feature extraction module, where the C3k2 module comprises a CBS module and a C3k module for feature convolution, with the C3k module performing feature extraction based on the bottleneck layer Bottleneck; analyzing the target features of direct tailwater discharge behavior based on acquired manually interpreted information; constructing a multidimensional fusion feature convolutional module based on the target features and a dynamic serpentine convolutional layer; updating the CBS module of the bottleneck layer Bottleneck in the C3k module using the multidimensional fusion feature convolutional module in the original YOLO11 model to construct the Bottleneck_DSC module; and updating the C3k2 feature extraction module in the original YOLO11 model based on the bottleneck layer Bottleneck and the Bottleneck_DSC module to obtain the improved model. Based on the perception enhancement mechanism, loss analysis is performed on the prediction results and the validation set using an identification enhancement loss function, including: extracting prediction boxes based on the prediction results, and extracting labeled boxes based on the validation set; calculating the intersection-union ratio (IUU) of the prediction boxes and the labeled boxes based on the perception enhancement mechanism to obtain the target loss; analyzing the labeled shape of the labeled boxes; performing direction-weighted processing on the center point offset based on the analyzed labeled shape of the labeled boxes to calculate the shape constraint loss, and performing aspect ratio processing on the labeled shape to calculate the shape-weighted distance loss; and calculating the final loss using the target loss, the shape constraint loss, and the shape-weighted distance loss based on the identification enhancement loss function.

2. The method according to claim 1, characterized in that, Based on the acquired remote sensing image samples of the sample area, feature processing is performed to determine vector point positioning data. Combined with the acquired map tiles, a sample dataset is established, including: Remote sensing image samples are collected from the sample area, and the remote sensing image samples are preprocessed and updated. Obtain and interpret location information; Based on the interpreted positioning information, the sewage discharge in each of the remote sensing image samples is located, the sample points are delineated, and vector point positioning data of each remote sensing image sample is obtained. The vector point positioning data carries visual features that appear as traces of pond sewage discharge. For the vector point positioning data, map tiles corresponding to the vector point positioning data are obtained from the map tile source based on a preset map downloader; Target samples are collected on the map tiles, and targets are selected based on the visual features to establish a sample dataset; In the sample dataset, the selected target of each sample is surrounded by a box as a label.

3. The method according to claim 1, characterized in that, Based on dynamic serpentine convolutional layers, a multi-dimensional fusion feature convolutional module is constructed, including: A parallel structure is adopted to establish serpentine dynamic convolutions and vertical CBS layers in different directions of input feature input, as dynamic serpentine convolution layers; Based on the dynamic snake-shaped convolutional layer, a multi-dimensional fusion feature convolutional module is established by combining a fusion layer with multi-view depth features.

4. The method according to claim 1, characterized in that, The improved model is trained based on the enhanced training set, and dynamically adapts to the target from different directions through the multi-dimensional fusion feature convolution module, fitting and extracting depth features of the target object from different viewpoints, and outputting prediction results, including: The training set is input into the improved model for feature extraction to obtain initial features; Dimensionality reduction is performed based on the initial features to obtain dimensionality-reduced features; The dimensionality reduction features are input into the multidimensional fusion feature convolution module, which dynamically adapts to the target from different directions to extract the depth features of the direct discharge behavior of tailwater. Based on the depth features, feature fusion is performed from different perspectives to obtain fused features; Based on the fusion features, a prediction is made and the prediction result is output.

5. The method according to any one of claims 1-4, characterized in that, After obtaining the recognition results, it also includes: Obtain verification samples, which include remote sensing images for verification, and each image in the verification samples carries a bounding box; Extract the target bounding box from the recognition results, and determine the localization target and confidence level of the target bounding box; For the localization target, the target boxes are sorted based on the confidence level, and a non-maximum suppression method is introduced to traverse and remove invalid target boxes, thus suppressing the target boxes; Based on the target bounding box and the labeled bounding box of the validation sample, the prediction precision is determined by evaluating the proportion of actual positive samples to predicted positive samples, and the prediction recall is determined by evaluating the proportion of actual positive samples to the total actual positive samples. Based on the prediction precision and the prediction recall, the model performance is evaluated using the mean precision method.

6. A rapid identification system for direct discharge of pond tailwater, characterized in that, include: The sample acquisition module is used to perform feature processing on the remote sensing image samples of the acquired sample area, determine the vector point positioning data, and combine the acquired map tiles to establish a sample dataset, which includes a training set and a validation set. An enhancement module is used to perform data augmentation processing based on the training set and to establish a perceptual enhancement mechanism; The model building module is used to build a YOLO11 model. Based on the target features of the pre-identified tailwater direct discharge behavior, an improved multi-dimensional fusion feature convolution module is established to build an improved model with enhanced feature extraction. The multi-dimensional fusion feature convolution module includes a dynamic serpentine convolution layer. The model training module is used to train the improved model based on the enhanced training set, dynamically adapt to the target from different directions through the multi-dimensional fusion feature convolution module, fit and extract the depth features of the target from different perspectives, and output the prediction results. The loss optimization module is used to perform loss analysis on the prediction results and the validation set by identifying the enhanced loss function based on the perception enhancement mechanism, update the parameters of the improved model, and obtain the trained pond tailwater identification model. The behavior recognition module is used to quickly identify the direct discharge behavior of pond tailwater from the acquired target remote sensing data through the pond tailwater recognition model, and obtain the recognition result. The data augmentation process based on the training set and the establishment of a perception enhancement mechanism include: using data augmentation techniques to expand each image sample and annotation in the training set and updating the training set; using an adaptive histogram equalization method to divide each sample in the training set into blocks for histogram equalization, limiting the contrast enhancement amplitude, and updating the training set; establishing a recognition enhancement loss function based on the analysis of the impact of short side direction and shape deviation on model training during bounding box regression; and introducing the stability of the model in recognizing tailwaters with different flow directions, combined with the recognition enhancement loss function, to establish a perception enhancement mechanism. A YOLO11 model is constructed, and an improved multidimensional fusion feature convolutional module is established based on the pre-identified target features of direct tailwater discharge behavior. This results in an improved model with enhanced feature extraction, including: constructing an original YOLO11 model including a C3k2 feature extraction module, where the C3k2 module comprises a CBS module and a C3k module for feature convolution, with the C3k module performing feature extraction based on the bottleneck layer Bottleneck; analyzing the target features of direct tailwater discharge behavior based on acquired manually interpreted information; constructing a multidimensional fusion feature convolutional module based on the target features and a dynamic serpentine convolutional layer; updating the CBS module of the bottleneck layer Bottleneck in the C3k module using the multidimensional fusion feature convolutional module in the original YOLO11 model to construct the Bottleneck_DSC module; and updating the C3k2 feature extraction module in the original YOLO11 model based on the bottleneck layer Bottleneck and the Bottleneck_DSC module to obtain the improved model. Based on the perception enhancement mechanism, loss analysis is performed on the prediction results and the validation set using an identification enhancement loss function, including: extracting prediction boxes based on the prediction results, and extracting labeled boxes based on the validation set; calculating the intersection-union ratio (IUU) of the prediction boxes and the labeled boxes based on the perception enhancement mechanism to obtain the target loss; analyzing the labeled shape of the labeled boxes; performing direction-weighted processing on the center point offset based on the analyzed labeled shape of the labeled boxes to calculate the shape constraint loss, and performing aspect ratio processing on the labeled shape to calculate the shape-weighted distance loss; and calculating the final loss using the target loss, the shape constraint loss, and the shape-weighted distance loss based on the identification enhancement loss function.

7. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the rapid identification method for direct discharge of pond tailwater as described in any one of claims 1-5.

Images