Intelligent risk assessment method and system for water body plastic garbage pollution based on multi-index fusion, terminal and readable storage medium

Abstract

The application discloses a kind of based on multi-index fusion's water plastic garbage pollution intelligent risk assessment method and system, terminal and readable storage medium, the assessment method includes: image or video is collected, input optimized YOLO26 target detection model, identify and output the total number of water surface garbage and the number of plastic garbage, combined with the area of the water body area to be evaluated, obtain clean coast index and plastic abundance index, while combining the mapping relationship of preset clean coast index and pollution level, the mapping relationship of plastic abundance index and abundance level, obtain the pollution level and abundance level of the water body area to be evaluated, combined with weighted fusion rule, the pollution risk level of water plastic garbage can be quickly and accurately evaluated, with the advantages of high overall evaluation efficiency, comprehensive evaluation dimension, good reliability, strong applicability, etc., can directly support emergency response decision-making, high in use value, good application prospect.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision, environmental monitoring and risk assessment technology, and relates to an intelligent risk assessment method and system, terminal and readable storage medium for water plastic waste pollution based on multi-indicator fusion. Background Technology

[0002] Plastic waste pollution has become a global environmental crisis, particularly posing a serious threat to ecosystems and human health in coastal areas, rivers, and lakes. Therefore, effective pollution assessment is a prerequisite for developing remediation strategies and monitoring their effectiveness. However, existing technologies suffer from systemic deficiencies and cannot meet the needs for large-scale, high-frequency, objective, and quantitative intelligent assessments, as detailed below: (I) Deficiencies of Traditional Manual Visual Survey Methods Traditional manual visual survey methods rely on manual walking surveys along the shoreline segment by segment, which is extremely inefficient, with a single person covering only 0.1-0.5 km² per day. 2 The system suffers from significant subjective error; different investigators can arrive at judgments for the same area with discrepancies exceeding 20%. It also fails to quantify spatial density and exhibits severe data lag. Consequently, it has virtually no practical value in large-scale, high-frequency dynamic pollution monitoring scenarios, and is particularly incapable of supporting emergency response decisions.

[0003] (II) Technical Limitations of Traditional Image Processing Methods Traditional image processing methods, such as color thresholding, edge detection, and texture analysis, suffer from a false detection rate of 30-50% in complex real-world scenarios involving changes in lighting, water reflection, and wave interference. These methods can only perform binary classification of "present / absent," failing to subdivide waste categories (such as plastic bottles, foam boards, and fishing gear), lacking area normalization capabilities, and unable to output any quantifiable pollution indicators, thus failing to support standardized assessment systems.

[0004] (III) Common defects of YOLOv8 / YOLOv11 improvement schemes Existing technologies based on YOLOv8 / YOLOv11 improvements all employ deep learning object detection technology and possess a certain degree of automatic detection capability, but they generally suffer from the following six prominent shortcomings: (1) Only the target count is output, and area normalization is not implemented. All detection results in the existing technology are "total target N", and it is impossible to calculate the density of garbage per unit area (pieces / m²). 2 This means that the Clean Coast Index (CCI) cannot be calculated automatically, and all quantitative analyses require extensive manual post-processing.

[0005] (2) Lack of photogrammetric integration leads to large area estimation errors. The shooting height, sensor parameters and detection system are completely disconnected, requiring manual calculation of the area based on GPS trajectory; the area estimation error based on GPS is usually >15%, and the error is even greater in the UAV video scanning scenario, which cannot meet the requirements of scientific-level quantitative assessment.

[0006] (3) The assessment indicators are too simplistic and cannot distinguish between complex pollution scenarios. Relying solely on the total amount of waste or the detection results of a single target, without introducing specific indicators such as the Plastic Abundance Index (PAI), it is impossible to distinguish between two completely different pollution scenarios: "a large amount of waste but mainly non-plastic" and "a small amount of waste but extremely high plastic density". This results in highly misleading risk assessment conclusions.

[0007] (4) Lack of a multi-indicator consistency verification mechanism. Existing technologies cannot quantify the degree of synergy and internal consistency of multi-indicator evaluation results, and the system output results lack self-verification capabilities.

[0008] (5) The performance of the basic detection model has bottlenecks. On a custom 11-class water surface debris dataset, YOLOv8 / v11 has an mAP@0.5 of only 0.882-0.890 for small objects (Small_Plastic), a recall rate of 0.823-0.825, and an inference latency of 2.2-2.3 ms. The ability to classify 11 classes is insufficient, which restricts the accuracy of subsequent multi-index quantization calculations.

[0009] (6) The system architecture is fragmented and cannot achieve integrated assessment. There is no automatic data flow between the detection module and the risk assessment module. Multiple independent tools are required for data conversion and processing. The automation level of the entire "detection-area quantification-index calculation-risk assessment" chain is extremely low. In actual application, there are many operation steps and the overall efficiency is low.

[0010] The aforementioned six deficiencies prevent existing technologies from meeting the needs for large-scale, high-frequency, objective, and quantitative intelligent assessment of plastic waste pollution. Systematic breakthroughs are urgently needed in five dimensions: detection models, photogrammetric area quantification, multi-indicator fusion, consistency verification, and system integration.

[0011] For the reasons stated above, this invention is proposed. Summary of the Invention

[0012] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology by providing a smart risk assessment method and system for water plastic waste pollution based on multi-indicator fusion, a terminal and a readable storage medium, which has high overall assessment efficiency, comprehensive assessment dimensions, good reliability and strong applicability.

[0013] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A smart risk assessment method for water pollution by plastic waste based on multi-indicator fusion includes the following steps: S1. Collect images or videos of the water body area to be evaluated; S2. Input the images or videos obtained in step S1 into the optimized YOLO26 target detection model to identify and output the total number of surface debris and the number of plastic debris. S3. Based on the images or videos collected in step S1, obtain the area of ​​the water body to be evaluated. Combine the total amount of surface debris and the amount of plastic debris obtained in step S2 to obtain the Clean Coast Index and Plastic Abundance Index. At the same time, combine the preset mapping relationship between the Clean Coast Index and the pollution level, and the mapping relationship between the Plastic Abundance Index and the abundance level to obtain the pollution level and abundance level of the water body to be evaluated. S4. Based on the pollution level and abundance level, and combined with the weighted fusion rules, assess the pollution risk level of plastic waste in the water body.

[0014] The above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved in step S2 by using the image obtained in step S1 and combining it with the following formula to obtain the area of ​​the water body to be assessed: , Where A represents the area of ​​the water body to be assessed, in meters (m²). 2 ; This refers to the camera's shooting height, in meters (m). This refers to the camera's focal length, in mm. These are the width and height of the sensor in the camera, respectively, in mm.

[0015] The above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved in step S2 by using the video obtained in step S1 and combining it with the following formula to obtain the area of ​​the water body to be assessed: , Where A represents the area of ​​the water body to be assessed, in units of... The shooting height of the camera, in units of The width of the sensor in the camera, in mm; The relative flight speed of the camera, in units of t represents the camera's flight time, in units of... ; The focal length of the camera, in units of .

[0016] The aforementioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved by allowing the camera to fly in a downstream mode. Obtained by the following formula: , in, The camera's flight speed, in units of The water flow velocity in the area to be evaluated is expressed in units of... .

[0017] The aforementioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved by allowing the camera to fly against the current when in counter-current mode. Obtained by the following formula: , in, The camera's flight speed, in units of The water flow velocity in the area to be evaluated is expressed in units of... ;when Normal progress is being made, when When hovering, the process is handled.

[0018] The aforementioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved by allowing the camera to fly in a transverse flight mode. Obtained by the following formula: , in, The camera's flight speed, in units of The water flow velocity in the area to be evaluated is expressed in units of... .

[0019] The aforementioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved by allowing the camera to hover in flight mode. Obtained by the following formula:

[0020] The water flow velocity in the area to be evaluated is expressed in units of... .

[0021] In a further improvement to the above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion, the Clean Coast Index in step S3 is obtained by the following formula: CCI = (N_total / A) × K, Wherein, CCI is the Clean Coast Index, N_total is the total amount of surface debris, A is the area of ​​the water body to be assessed, and K is the standardization coefficient, K=20.

[0022] In a further improvement to the above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-index fusion, the plastic abundance index in step S3 is obtained by the following formula: PAI=[N_plastic / log 10 [(N_total)] / A×K, Wherein, PAI is the plastic abundance index, N_plastic is the amount of plastic waste, N_total is the total amount of surface debris, A is the area of ​​the water body to be evaluated, and K is the standardization coefficient, K=20.

[0023] The aforementioned intelligent risk assessment method for water pollution by plastic waste based on multi-indicator fusion is further improved in step S3 by setting a mapping relationship between the Clean Coast Index and the pollution level. This includes five pollution levels based on the Clean Coast Index: when 0 ≤ CCI < 2, the pollution level is classified as Level 1, and the corresponding pollution description is extremely low pollution; when 2 ≤ CCI < 5, the pollution level is classified as Level 2, and the corresponding pollution description is slightly polluted; when 5 ≤ CCI < 10, the pollution level is classified as Level 3, and the corresponding pollution description is moderate pollution; when 10 ≤ CCI < 20, the pollution level is classified as Level 4, and the corresponding pollution description is heavily polluted; when CCI ≥ 20, the pollution level is classified as Level 5, and the corresponding pollution description is extremely heavily polluted.

[0024] The aforementioned intelligent risk assessment method for water pollution by plastic waste based on multi-indicator fusion is further improved in step S3 by pre-setting the mapping relationship between the plastic abundance index and the abundance level. This includes classifying the plastic abundance index into 5 levels: when 0 ≤ PAI < 0.1, the abundance level is classified as level 1, corresponding to a plastic threat level of almost no plastic threat; when 0.1 ≤ PAI < 1, the abundance level is classified as level 2, corresponding to a plastic threat level of slight plastic threat; when 1 ≤ PAI < 4, the abundance level is classified as level 3, corresponding to a plastic threat level of moderate plastic threat; when 4 ≤ PAI < 8, the abundance level is classified as level 4, corresponding to a plastic threat level of severe plastic threat; and when PAI ≥ 8, the abundance level is classified as level 5, corresponding to a plastic threat level of extremely severe plastic threat.

[0025] The above-mentioned intelligent risk assessment method for water pollution by plastic waste based on multi-indicator fusion is further improved in step S4, where the weighted fusion rule is as follows: when the pollution level is ≥4 and the abundance level is ≥4, it is determined to be extremely high risk; when the pollution level is ≥3 or the abundance level is ≥4, it is determined to be high risk; when the pollution level is ≥2 or the abundance level is ≥2, it is determined to be medium risk; and in other cases, it is determined to be low risk.

[0026] The above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved in step S1 by taking pictures or videos of the water body to be assessed directly above it, and collecting images or videos of the water body to be assessed.

[0027] The above-mentioned intelligent risk assessment method for water pollution by plastic waste based on multi-indicator fusion is further improved in step S2, wherein the surface waste includes plastic waste and non-plastic waste; the plastic waste includes foam blocks, plastic buoys, fishing nets, ropes, plastic buoys, plastic bottles, foam boxes, other plastics, and foam buoys; and the non-plastic waste includes glass and metal.

[0028] The aforementioned intelligent risk assessment method for water plastic waste pollution based on multi-index fusion is further improved in step S2 by including the following processing before inputting the image or video into the optimized YOLO26 target detection model: performing reflection suppression preprocessing on the image; after completing the reflection suppression preprocessing, inputting the processed image into the optimized YOLO26 target detection model; the reflection suppression preprocessing includes the following steps: T1. Estimate the pseudo-normal field corresponding to the image based on the brightness gradient of the image to be processed; T2. Combining the preset light source direction and viewing angle direction, based on the bidirectional reflectance distribution function (BRDF) model, predict the specular component of the image according to the pseudo-normal field. The bidirectional reflectance distribution function model includes a Lambertian diffuse reflection term and a Blinn-Phong specular term. T3. Subtract the predicted highlight component from the image to be processed to obtain the image after eliminating water surface reflection interference. Then, input the image after eliminating water surface reflection interference into the optimized YOLO26 target detection model.

[0029] In step S2, before inputting the video into the optimized YOLO26 target detection model, the following processing is also included: decoding the video into a continuous sequence of video frames, performing reflection suppression preprocessing on each video frame; after completing the reflection suppression preprocessing, re-encoding / stitching the processed image frames into a video in their original temporal order; and inputting the re-encoded / stitched video into the optimized YOLO26 target detection model. The reflection suppression preprocessing steps used for each video frame are the same as those used for the processed images.

[0030] The above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved in step S2, where the construction method of the optimized YOLO26 target detection model includes the following steps: (1) Collect images of different water areas, label the types and quantities of surface debris in the images to form a dataset, and perform data augmentation on the dataset; the types of surface debris are foam blocks, plastic buoys, fishing nets, ropes, plastic buoys, plastic bottles, foam boxes, other plastics, foam buoys, glass, and metal; (2) Divide the labeled dataset into training set, validation set and test set; (3) Input the training set, validation set and test set into the YOLO26 object detection model for training, testing and evaluation until the detection accuracy is ≥0.890 and the inference delay time is ≤1.5 ms, and the optimized YOLO26 object detection model is obtained.

[0031] The above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved by the following method for constructing the optimized YOLO26 target detection model: During training the YOLO26 target detection model, a small target perception label allocation method is used to determine positive and negative samples. This small target perception label allocation method includes the following steps: U1. Determine whether the annotation box is a small target based on the ratio of the area of ​​the annotation box to the area of ​​the image. When the ratio is less than a preset small target area threshold, the annotation box is determined to be a small target. U2. Calculate the alignment score between the candidate anchor point and the bounding box. The alignment score is calculated based on the classification prediction probability of the candidate anchor point and the intersection-union ratio of the candidate anchor point and the bounding box. For the bounding box that is determined to be a small target, the alignment score of its corresponding target is increased by a preset score amplification factor. U3. Select candidate anchor points based on alignment scores. For the bounding boxes that are determined to be small targets, the number of candidate anchor points selected is increased by the preset small target candidate anchor point increment on the basis of the preset base candidate anchor point number. U4. Candidate anchor points are filtered based on the intersection-union ratio (IU) threshold to obtain positive sample anchor points. Among them, the IU threshold of the bounding box that is determined to be a small target is lower than the IU threshold of other bounding boxes.

[0032] The aforementioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved by the following method for constructing the optimized YOLO26 target detection model: During training the YOLO26 target detection model, a progressive loss weight scheduling method is used to determine the loss weights for each training round. This progressive loss weight scheduling method includes: dynamically adjusting the weights of the three losses—boundary box regression loss, classification loss, and distribution focus loss—according to a preset scheduling strategy based on the training progress of the current training round relative to the total number of training rounds. Specifically, the bounding box regression loss weight is greater in the early training phase than in the later training phase, and the classification loss weight is less in the early training phase than in the later training phase. The preset scheduling strategy includes any one of cosine annealing scheduling, linear scheduling, and step-wise scheduling.

[0033] The above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved by including, in step S4, the verification of the assessment results: S41. Based on the Clean Coast Index and Plastic Abundance Index of multiple sampling points, obtain the pollution level and abundance level. Construct a 5×5 confusion matrix with the pollution level as the row and the abundance level as the column, and obtain the precision and recall for each level. S42. Based on the precision and recall of each level, obtain the macro average F1 score and Spearman rank correlation coefficient; S43. Based on the macro-average F1 score and Spearman rank correlation coefficient, assess the degree of pollution and the consistency of indicators.

[0034] In a further improvement to the above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion, the accuracy rate in step S41 is calculated by the following formula: , in, For the first Precision rate for each category; The element in the k-th row and k-th column of the confusion matrix, where k ≤ 5; Let i be the element in the i-th row and k-th column of the confusion matrix, where i ≤ 5.

[0035] In a further improvement to the above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion, the recall rate in step S41 is calculated by the following formula: , in, For the first Recall rate for each category; The element in the k-th row and k-th column of the confusion matrix, where k ≤ 5; Let j be the element in the k-th row and j-th column of the confusion matrix, where j≤5.

[0036] The above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion is further improved in that, in step S42, the macro-average F1 score is calculated by the following formula: , , in, The macro average F1 score, For the first F1 scores for each category For the first Precision rate for each category For the first Recall rates for each category.

[0037] In a further improvement to the above-mentioned intelligent risk assessment method for water plastic waste pollution based on multi-indicator fusion, the Spearman rank correlation coefficient in step S42 is calculated using the following formula: , in, The Spearman rank correlation coefficient. ∈[ [1, 1], the closer ρ is to 1, the higher the consistency between the two indicators; d i The difference between the CCI rank and the PAI rank for each sample, where n is the total number of samples.

[0038] The above-mentioned intelligent risk assessment method for water pollution by plastic waste based on multi-indicator fusion is further improved in step S43. When using the Spearman rank correlation coefficient as the evaluation index, the standard for assessing the pollution level and the consistency of the indicators is as follows: when |ρ|≥0.7, the evaluation results are highly consistent, indicating that the evaluation results are objective and reliable; when 0.4≤|ρ|<0.7, the evaluation results are moderately consistent; when |ρ|<0.4, the two indicators need to be evaluated separately; ρ<0 indicates a negative correlation, requiring on-site verification. When using the macro-average F1 score as the evaluation index, the standard for assessing the pollution level and the consistency of the indicators is as follows: when F1... macro When F1 is ≥0.7, the evaluation results are highly consistent, indicating that the classification results of the two indicators are highly consistent; when F1 is ≤0.4, the evaluation results are highly consistent. macro When F1 < 0.7, the evaluation results are moderately consistent; when F1 macroWhen the value is less than 0.4, the consistency of the evaluation results is low, and the two indicators need to be evaluated separately. When the consistency judgment results given by the macro average F1 score and the Spearman rank correlation coefficient are inconsistent, the lower consistency level of the two shall be taken as the final judgment result, and the strict principle shall be followed. When |ρ| < 0.4 or F1 macro When <0.4, any of the following methods shall be used to assess the degree of pollution and the consistency of the indicators: Method 1: The Clean Coast Index independently determines the pollution risk level of the sampling point according to its preset pollution level classification standards; Method 2: The plastic abundance index independently determines the pollution risk level of the sampling point according to its preset abundance level classification standard; Method 3: If the risk levels determined by Method 1 and Method 2 are inconsistent, the higher risk level shall be taken as the final pollution risk level of the sampling point; Method 4: Trigger the on-site verification process for the sampling point, collect on-site measured data, and verify and correct the judgment results.

[0039] As a general technical concept, the present invention also provides an intelligent risk assessment system for water pollution caused by plastic waste, used to assess the risk of plastic waste pollution in water bodies according to the above-described assessment method; the intelligent risk assessment system includes: The acquisition module is used to acquire images or videos of the water body area to be assessed and to calculate the area of ​​the water body area to be assessed. The target detection module is used to identify images or videos and output the total amount of debris and the amount of plastic waste on the water surface; The index calculation module is used to calculate the Clean Coast Index / Plastic Abundance Index and assess the pollution risk level of plastic waste in water bodies; The confusion matrix and evaluation module is used to construct a two-index confusion matrix and calculate the macro-average F1 score and Spearman rank correlation coefficient. The graphical user interface module provides a human-computer interaction interface, enabling parameter setting, result visualization, and data export.

[0040] As a general technical concept, the present invention also provides a terminal including a processor and a memory, the memory storing a computer program, the processor calling the computer program to perform the steps of the evaluation method described above.

[0041] As a general technical concept, the present invention also provides a readable storage medium storing a computer program that is invoked by a processor to perform the steps of the evaluation method described above.

[0042] Compared with the prior art, the advantages of the present invention are as follows: (1) To address the technical shortcomings of existing assessment methods, such as numerous operational steps, low overall assessment efficiency, single assessment indicators, poor reliability, and poor adaptability, this invention creatively provides an intelligent risk assessment method for water body plastic waste pollution based on multi-indicator fusion. The optimized YOLO26 target detection model identifies images or videos of the water body area to be assessed and outputs the total number of surface debris and the amount of plastic waste. This significantly improves detection accuracy, recall, and precision, and significantly reduces inference latency. Specifically, the detection precision (mAP@0.5) is improved by 0.45%-1.36%, the recall rate is improved by 3.86%, and the inference latency can be shortened to within 1.5 mm, significantly reducing the false negative rate, especially for small targets such as Small_Plastic, and greatly shortening the detection time. Based on this, the area of ​​the water body to be assessed is obtained from the collected images or videos, and combined with the total number of surface debris and the amount of plastic waste, the Clean Coast Index and the... The plastic abundance index, combined with the pre-defined mapping relationship between the Clean Coast Index and pollution level, and the mapping relationship between the plastic abundance index and abundance level, obtains the pollution level and abundance level of the water body to be assessed. By fusing the Clean Coast Index and the plastic abundance index, it can achieve fine semantic separation of 7 types of plastics / 4 types of non-plastics, distinguishing different pollution scenarios and supporting differentiated emergency response decisions. This makes the assessment more comprehensive and completely solves the problem of misleading single indicators. At the same time, the area calculation is more accurate, with the photogrammetric area error being <0.5%, which is about 30 times more accurate than the existing GPS error >15%. In particular, it supports accurate calculation of non-repeating coverage area using four drone video flight modes. Finally, based on the pollution level and abundance level, combined with weighted fusion rules, the pollution risk level of plastic waste in water bodies can be quickly and accurately assessed. It has the advantages of high overall assessment efficiency, comprehensive assessment dimensions, good reliability, and strong applicability. It can directly support emergency response decisions, has high use value, and good application prospects.

[0043] (2) In this invention, the consistency verification of the 5×5 dual-index confusion matrix + macro average F1 score and Spearman rank correlation coefficient is pioneered, which gives the system the ability to self-verify and anomaly warning, and the verification ability is more objective, filling the gap in the field of multi-index verification in the existing technology.

[0044] (3) In this invention, the entire link of automatic linkage of photogrammetry → YOLO26 detection → dual index calculation → risk assessment → GUI visualization is adopted, which has a higher degree of system integration and reduces the operation steps by about 70% compared with the existing cross-tool solution, and greatly improves the efficiency of actual application. Attached Figure Description

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0046] Figure 1 This is a schematic diagram of the intelligent risk assessment process for water plastic waste pollution based on multi-index fusion in Embodiment 1 of the present invention.

[0047] Figure 2 This is a schematic diagram of the photogrammetric geometry principle in Embodiment 1 of the present invention.

[0048] Figure 3 This is a diagram of the 11x11 normalized confusion matrix structure of the test set in Embodiment 1 of the present invention.

[0049] Figure 4 This is a schematic diagram illustrating the division of pollution level and abundance level in Embodiment 1 of the present invention. Detailed Implementation

[0050] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0051] It should be noted that the following embodiments are only used to further illustrate the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-essential improvements and adjustments made to the present invention by those skilled in the art based on the above-described invention content are still within the scope of protection of the present invention.

[0052] The system operating environment in this example is as follows: Python 3.10, PyTorch 2.x, CUDA 11.8, GTX3060 (12GB), Windows 11 / Ubuntu 22.04 dual-platform compatibility. The YOLO26s model weights are trained using the Ultralytics framework, and the GUI is built based on PyQt6.

[0053] Example 1 A smart risk assessment method for waterborne plastic waste pollution based on multi-indicator fusion is proposed, specifically for intelligent risk assessment of coastal plastic waste pollution. The assessment process is illustrated in the diagram below. Figure 1 As shown, it includes the following steps: S1. Acquire images of the water body area to be evaluated, specifically: like Figure 2 As shown, a camera was used to photograph the water area to be evaluated (coastal zone) directly above the coastline to collect images of the coastline. The camera used was a Sony A7R IV with a focal length of f=35mm, a sensor width and height of 35.9mm×24mm, and a shooting height of H=30m.

[0054] In this embodiment, before performing step S2, the following processing is also included: performing reflection suppression preprocessing on the image, and after completing the reflection suppression preprocessing, inputting the processed image into the optimized YOLO26 target detection model.

[0055] In this embodiment, the image acquired in step S1 is preprocessed to suppress reflection. Specifically, after converting the image to grayscale, the brightness gradient of the grayscale image is calculated along the horizontal and vertical directions, and a pseudo-normal field corresponding to the image is constructed based on the brightness gradient. Combining the preset light source direction (azimuth angle 135°, elevation angle 60°) and the camera's viewing angle direction (taking the camera's optical axis direction, i.e., vertically downward), the highlight component of each pixel in the image is calculated based on the bidirectional reflectance distribution function (BRDF) model, where the ambient light reflectance coefficient is 0.25, the diffuse reflectance is the normalized original image pixel value, the specular reflectance coefficient is 0.15, and the highlight index is 32. The original image and the calculated highlight component are subtracted pixel by pixel to obtain the image after eliminating water surface reflection interference. This image is used as the input image for step S2 and input into the optimized YOLO26 target detection model.

[0056] In this embodiment, by performing reflection suppression preprocessing based on the bidirectional reflectance distribution function (BRDF) model on the image before step S2, the interference of water surface mirror reflection and wave highlights on target detection is effectively eliminated, overcoming the technical defects of existing technologies with a false detection rate of up to 30-50% in complex real-world scenarios such as lighting changes, water surface reflection, and wave interference.

[0057] S2. Input the image obtained in step S1 into the optimized YOLO26 target detection model to identify and output the total number of surface debris and the number of plastic debris.

[0058] In step S2, surface debris includes both plastic and non-plastic waste. Plastic waste is categorized into nine types: Styrofoam Piece, Plastic Buoy China, Net, Rope, Plastic Buoy, Plastic Bottle, Styrofoam Box, Other Plastics (Plastic_ETC), and Styrofoam Buoy. Non-plastic waste is categorized into two types: Glass and Metal. For example, glass could be fragments of a wine bottle, and metal could be an aluminum can.

[0059] In step S2 of this embodiment, the method for constructing the optimized YOLO26 target detection model includes the following steps: (1) Images of different water areas were collected, specifically 2377 images (2048×2048) from three types of scenes: coast, river and lake. The images were taken from drone aerial photography (20-80m altitude) and ground photography. LabelImg was used to label the type and quantity of water surface debris in each image. Water surface debris includes plastic and non-plastic debris. There are 9 types of plastic debris, specifically: Styrofoam Piece, Plastic Buoy China, and Fishing Net. The dataset was compiled from various categories of waste, including rope, plastic buoy, plastic bottle, styrofoam box, other plastics (Plastic_ETC), and non-plastic waste (glass and metal). Data augmentation was performed on the dataset using techniques such as random horizontal flipping, random cropping, color dithering (HSV enhancement), and Mosaic stitching. The results are shown in Table 1 below.

[0060] Table 1. Statistics on the number of labels in the surface debris dataset.

[0061] (2) The labeled dataset is divided into training set, validation set and test set according to training set: validation set: test set = 7:2:1.

[0062] (3) Input the training set and validation set into the YOLO26 object detection model for training and validation until the detection accuracy is ≥0.890 and the inference latency is ≤1.5 ms, to obtain the optimized YOLO26 object detection model, specifically: After 50 rounds of fine-tuning on YOLO26 (batch=16, lr=0.001), the final results on the test set were mAP@0.5=0.894, F1=0.882, recall=0.861, precision=0.904, and inference latency=1.3ms. The corresponding YOLO26 object detection model at this point is the optimized YOLO26 object detection model.

[0063] After 50 rounds of fine-tuning on YOLO26 (batch=16, lr=0.001), the final results on the test set were mAP@0.5=0.894, F1=0.882, recall=0.861, precision=0.904, and inference latency=1.3ms. The corresponding YOLO26 object detection model at this point is the optimized YOLO26 object detection model.

[0064] In this embodiment, when fine-tuning the YOLO26 object detection model, the label allocation stage uses a small object perception label allocation method to determine positive and negative sample anchor points. The small object perception label allocation method includes the following steps: U1. Determine whether the annotation box is a small target based on the ratio of the area of ​​the annotation box to the area of ​​the image. When the ratio is less than 0.01, the annotation box is determined to be a small target. U2. Calculate the alignment score between the candidate anchor point and the bounding box. The alignment score is obtained by multiplying the 0.5 power of the classification prediction probability of the candidate anchor point by the 6th power of the intersection-union ratio of the candidate anchor point and the bounding box. For the bounding box that is determined to be a small target, its corresponding alignment score is multiplied by an amplification factor of 1.5 to adjust it upward. U3. Select candidate anchor points based on alignment scores. Select 13 candidate anchor points for regular annotation boxes. For annotation boxes that are determined to be small targets, add 7 more candidate anchor points, for a total of 20 candidate anchor points. U4. Candidate anchor points are filtered according to the intersection-union ratio (IU) threshold to obtain positive sample anchor points. The IU threshold for regular bounding boxes is 0.25, and the IU threshold for bounding boxes identified as small targets is reduced to 0.15.

[0065] In this embodiment, by introducing the Small Target Aware Label Allocation (STAL) method during the model training phase, the alignment score of small target bounding boxes is increased, candidate anchor points are expanded, and the cross-union ratio threshold is reduced. This significantly improves the positive sample allocation quality of small targets (such as small pieces of plastic and foam fragments), enhances the detection recall rate of small targets, and provides a more reliable detection data foundation for the accurate calculation of the Plastic Abundance Index (PAI).

[0066] In this embodiment, when fine-tuning the YOLO26 object detection model, a progressive loss weight scheduling method is used to determine the loss weights for each training epoch. Based on the training progress of the current training epoch relative to the total number of training epochs, the weights of the bounding box regression loss, classification loss, and distribution focus loss are dynamically adjusted according to a preset scheduling strategy. The bounding box regression loss weight is greater in the early training phase than in the later training phase, and the classification loss weight is less in the early training phase than in the later training phase. The preset scheduling strategy includes any one of cosine annealing scheduling, linear scheduling, and step-wise scheduling. In this embodiment, cosine annealing scheduling is used. The warm-up epochs are 1 / 60 of the total training epochs. The bounding box regression loss weight transitions from 7.5 in the early training phase to 5.0 in the later training phase using cosine annealing. The classification loss weight transitions from 0.5 in the early training phase to 2.0 in the later training phase using cosine annealing. The distribution focus loss weight remains unchanged at 1.5.

[0067] In this embodiment, by introducing the progressive loss weight scheduling (ProgLoss) method during the model training phase, the bounding box regression loss weight and classification loss weight are dynamically transitioned with each training round, balancing the convergence speed of coordinate regression in the early stage of training with the classification refinement capability in the later stage of training, thereby further improving the overall detection performance of the model.

[0068] In addition, the training and validation sets were input into other YOLO versions of the model for training and validation. The test results are shown in Table 2.

[0069] Table 2. Performance comparison of different YOLO models on 11 types of surface debris datasets.

[0070] As shown in Table 2, YOLO26 outperforms existing YOLOv8s / v11s in all three metrics: mAP@0.5, recall, and precision. It also features reduced inference latency and a parameter count of only 7.42M, making it more suitable for edge computing deployments. The significant improvement in accuracy for small target detection such as Small_Plastic provides a reliable N-value for subsequent accurate PAI calculations. plastic The data foundation fundamentally improves the accuracy bottleneck of a single indicator system.

[0071] In addition, the evaluation of the optimized YOLO26 object detection model using a test set includes the following steps: (a) For each image in the test set, its true category is first determined by manual annotation, and then the trained model is used to predict the image to obtain the predicted category of the model.

[0072] (b) For each sample, record the correspondence between the “true category” and the “model predicted category”, count the number of samples in each category, and fill them into an 11×11 confusion matrix.

[0073] (c) Use color shades to represent the count of each cell in the matrix, with the diagonal representing the case where "True Class = Predicted Class".

[0074] Figure 3 This is a diagram of the 11×11 normalized confusion matrix structure of the test set in Embodiment 1 of the present invention. Figure 3 As shown, the model's recall rate for identifying various types of surface waste is generally higher than 0.8, with an accuracy rate of nearly 100% for specific categories (such as plastic buoys and foam buoys). This indicates that the 11 fine classification schemes provided by this invention can effectively distinguish waste of different materials and shapes, avoiding the feature confusion problem caused by coarse classification schemes.

[0075] Figure 3In the diagram, the diagonal elements represent the number of samples where the CCI and PAI ratings are consistent, reflecting the degree of consistency between the two indicators. Off-diagonal elements represent samples where the two indicators are inconsistent; the further away from the diagonal, the greater the difference between the two indicators.

[0076] In step S2, the optimized YOLO26 target detection model's detection results are: total number of surface debris N_total=254, number of plastic debris N_plastic=198, and number of non-plastic debris N_non_plastic=56.

[0077] S3. Based on the images acquired in step S1, obtain the area of ​​the water body region (coastal zone) to be evaluated. Combined with the total amount of surface debris and the amount of plastic debris obtained in step S2, obtain the Clean Coast Index and Plastic Abundance Index. Simultaneously, combining the preset mapping relationship between the Clean Coast Index and pollution level, and the mapping relationship between the Plastic Abundance Index and abundance level, obtain the pollution level and abundance level of the water body region (coastal zone) to be evaluated, as detailed below: In step S3, the area of ​​the water body (coastal zone) to be evaluated is obtained based on the image obtained in step S1, using the following formula: , Where A represents the area of ​​the water body to be assessed, in meters (m²). 2 ; This refers to the camera's shooting height, in meters (m). This refers to the camera's focal length, in mm. These are the width and height of the sensor in the camera, respectively, in mm.

[0078] The results show that A = (30×35.9 / 35) × (30×24 / 35) = 30.77×20.57 = 633.2 m 2 .

[0079] In step S3, the Clean Coast Index is obtained using the following formula: CCI = (N_total / A) × K, Wherein, CCI is the Clean Coast Index, N_total is the total amount of surface debris, A is the area of ​​the water body to be assessed, and K is the standardization coefficient, K=20.

[0080] The results show that CCI = (N_total / A) × K = (254 / 633.2) × 20 = 8.02.

[0081] In step S3, the plastic abundance index is obtained by the following formula: PAI=[N_plastic / log 10 [(N_total)] / A×K, Wherein, PAI is the plastic abundance index, N_plastic is the amount of plastic waste, N_total is the total amount of surface debris, A is the area of ​​the water body to be evaluated, and K is the standardization coefficient, K=20.

[0082] The results show that PAI = [N_plastic / log] 10 (N_total)] / A × 20 = (198 / log 10 (254)) / 633.2 × 20 = (198 / 2.405) / 633.2 × 20 = 2.60.

[0083] In step S3, the preset mapping relationship between the Clean Coast Index and the pollution level includes 5 pollution levels classified according to the Clean Coast Index, as shown in Table 3 and... Figure 4 As shown, the pollution levels are as follows: when 0 ≤ CCI < 2, the pollution level is classified as Level 1, and the corresponding pollution description is very low pollution; when 2 ≤ CCI < 5, the pollution level is classified as Level 2, and the corresponding pollution description is light pollution; when 5 ≤ CCI < 10, the pollution level is classified as Level 3, and the corresponding pollution description is moderate pollution; when 10 ≤ CCI < 20, the pollution level is classified as Level 4, and the corresponding pollution description is heavy pollution; when CCI ≥ 20, the pollution level is classified as Level 5, and the corresponding pollution description is extremely heavy pollution.

[0084] Table 3 Pollution Level Classification Standards

[0085] In step S3, the preset mapping relationship between plastic abundance index and abundance level includes the 5 abundance levels divided according to the plastic abundance index, as shown in Table 4 and Figure 4 As shown, the abundance levels are as follows: when PAI=0, the abundance level is 1, corresponding to a near-zero plastic threat; when 0.1≤PAI<1, the abundance level is 2, corresponding to a slight plastic threat; when 1≤PAI<4, the abundance level is 3, corresponding to a moderate plastic threat; when 4≤PAI<8, the abundance level is 4, corresponding to a severe plastic threat; and when PAI≥8, the abundance level is 5, corresponding to an extremely severe plastic threat.

[0086] Table 4 Abundance Classification Standards

[0087] S4. Based on the Clean Coast Index and the Plastic Abundance Index, and combined with the weighted fusion rule, assess the pollution risk level of plastic waste in the water body.

[0088] In step S4, the weighted fusion rules are as follows: when the pollution level is ≥4 and the abundance level is ≥4, it is determined to be extremely high risk; when the pollution level is ≥3 or the abundance level is ≥4, it is determined to be high risk; when the pollution level is ≥2 or the abundance level is ≥2, it is determined to be medium risk; otherwise, it is determined to be low risk.

[0089] Based on the results in step S3, in this embodiment, the pollution level is level 3 and the abundance level is level 3, which is determined to be of medium risk. It is recommended to increase the frequency of daily patrols and monitoring.

[0090] Step S4 also includes verifying the evaluation results: S41. Based on the Clean Coast Index and Plastic Abundance Index of multiple sampling points (different images taken at different sampling points) of the coastal zone, obtain the pollution level and abundance level. Construct a 5×5 confusion matrix with the pollution level as the row and the abundance level as the column, and obtain the precision and recall for each category.

[0091] In step S41, the 5×5 confusion matrix is ​​shown in Table 5.

[0092] Table 5. 5×5 Confusion Matrix of CCI×PAI for Each Sampling Point

[0093] In step S41, the accuracy is calculated using the following formula: , in, For the first Precision rate for each category; The element in the k-th row and k-th column of the confusion matrix represents the number of samples whose true class is k and which are predicted as k by the model. k is the index of the class, representing "the k-th class". k ≤ 5, and there are a total of 5 classes, with a corresponding upper limit of 5 for the summation. This represents the summation of i from 1 to 5, which is the total number of samples predicted by the model to be of class k. The element in the i-th row and k-th column of the confusion matrix represents the number of samples whose true class is i but are predicted as k, where i ≤ 5.

[0094] In step S41, the recall rate is calculated using the following formula: , in, For the first Recall rates for each category. The element in the k-th row and k-th column of the confusion matrix represents the number of samples whose true class is k and which are predicted as k by the model. k is the index of the class, representing "the k-th class", and k ≤ 5. This represents the summation of j from 1 to 5, which is the total number of samples whose true class is the k-th class. The element in the k-th row and j-th column of the confusion matrix represents the number of samples whose true class is k but are predicted as j, where j ≤ 5.

[0095] S42. For each level, calculate the precision and recall, obtain the macro-average F1 score and Spearman rank correlation coefficient.

[0096] In step S42, the macro-average F1 score is calculated using the following formula: , , in, The macro average F1 score, For the first F1 scores for each category For the first Precision rate for each category For the first Recall rates for each category.

[0097] The results showed that the macro average F1 score was F1macro=0.823.

[0098] In step S42, the Spearman rank correlation coefficient is calculated using the following formula: , in, The Spearman rank correlation coefficient. ∈[ [1, 1], the closer ρ is to 1, the higher the consistency between the two indicators; d is the difference between the CCI rank and PAI rank of each sample, and n is the total number of samples.

[0099] ρ=1: Perfectly positive correlation, the two variables are ordered exactly the same.

[0100] ρ=-1: Perfectly negative correlation, the two variables are ordered in completely opposite order.

[0101] ρ=0: Unrelated; the order of the two variables is irrelevant.

[0102] The aforementioned consistency verification mechanism fills the gap in the field of multi-index verification in existing technologies, ensures the objectivity and reliability of evaluation conclusions, and endows the system with self-verification and anomaly warning capabilities.

[0103] The results showed that the Spearman rank correlation coefficient ρ = 0.81 (p < 0.01).

[0104] S43. Based on the macro-average F1 score and Spearman rank correlation coefficient, assess the degree of pollution and the consistency of indicators.

[0105] In step S43, when using the Spearman rank correlation coefficient as the evaluation index, the standard for assessing the pollution level and the consistency of the index is as follows: when |ρ|≥0.7, the evaluation results are highly consistent, indicating that the evaluation results are objective and reliable; when 0.4≤|ρ|<0.7, the evaluation results are moderately consistent; when |ρ|<0.4, the two indicators need to be evaluated separately; when ρ<0, it indicates a negative correlation and on-site verification is required.

[0106] In step S43, when using the macro-average F1 score as the evaluation index, the standard for assessing the degree of pollution and the consistency of the index is: when F1 macro When F1 is ≥0.7, the evaluation results are highly consistent, indicating that the classification results of the two indicators are highly consistent; when F1 is ≤0.4, the evaluation results are highly consistent. macro When F1 < 0.7, the evaluation results are moderately consistent; when F1 macro When the value is less than 0.4, the consistency of the evaluation results is low, and the two indicators need to be evaluated separately.

[0107] In step S43, when the consistency judgment results given by the macro average F1 score and the Spearman rank correlation coefficient are inconsistent, the lower consistency level of the two shall be taken as the final judgment result, following the strict principle.

[0108] When |ρ| < 0.4 or F1 macro When <0.4, any of the following methods shall be used to assess the degree of pollution and the consistency of the indicators: Method 1: The Clean Coast Index independently determines the pollution risk level of the sampling point according to its preset pollution level classification standards; Method 2: The plastic abundance index independently determines the pollution risk level of the sampling point according to its preset abundance level classification standard; Method 3: If the risk levels determined by Method 1 and Method 2 are inconsistent, the higher risk level shall be taken as the final pollution risk level of the sampling point; Method 4: Trigger the on-site verification process for the sampling point, collect on-site measured data, and verify and correct the judgment results.

[0109] Based on the results in step S42, it can be seen that the CCI and PAI dual-indicator evaluation conclusions are highly consistent, the system evaluation results are objective and reliable, and can be directly used for emergency response decision-making.

[0110] A smart risk assessment system for water pollution by plastic waste is provided, used to assess the risk of plastic waste pollution in water bodies according to the assessment method of Example 1, wherein the smart risk assessment system includes: The acquisition module is used to acquire images or videos of the water body area to be assessed and to calculate the area of ​​the water body area to be assessed.

[0111] The target detection module is used to identify images or videos and output the total number of surface debris and the amount of plastic waste.

[0112] The index calculation module is used to calculate the Clean Coast Index / Plastic Abundance Index and assess the pollution risk level of plastic waste in water bodies.

[0113] The Confusion Matrix and Evaluation module is used to construct a two-index confusion matrix and calculate the macro-average F1 score and Spearman rank correlation coefficient.

[0114] The graphical user interface module provides a human-computer interaction interface for parameter setting, result visualization, and data export.

[0115] A terminal includes a processor and a memory, wherein the memory stores a computer program, and the processor invokes the computer program to perform the steps of the evaluation method of Embodiment 1.

[0116] A readable storage medium storing a computer program, wherein the computer program is invoked by a processor to perform the steps of the evaluation method of Embodiment 1.

[0117] Example 2 A smart risk assessment method for waterborne plastic waste pollution based on multi-indicator fusion, specifically for intelligent risk assessment of coastal plastic waste pollution, includes the following steps: S1. Collect video footage of the water body area to be assessed, specifically: A camera was used directly above the coastline to record video of the water area to be evaluated (coastline). The camera used was a Sony A7R IV with a focal length of f=35mm, a sensor width and height of 35.9mm×24mm, and a shooting height of H=25m.

[0118] In this step, the camera is mounted on the drone, which then performs a dynamic scan of the coastal zone. The flight parameters are: flight mode is counter-current flight, and the drone's flight speed is [not specified]. =5.0 m / s, water flow velocity =1.2 m / s, flight altitude H=25m, video frame rate 25fps, flight duration 120s.

[0119] In other embodiments, the flight mode can also be downstream flight, cross-cutting flight, or hovering flight. Through the inter-frame area accumulation algorithm of the above four flight modes, the overlapping areas between video frames are eliminated, and the accurate integration of the non-overlapping coverage area is achieved with a theoretical error of <0.5%, which is about 30 times more accurate than the existing GPS integration scheme.

[0120] S2. Input the video obtained in step S1 into the optimized YOLO26 target detection model (same as in Example 1) to identify and output the total number of surface debris and the number of plastic debris.

[0121] In step S2, before inputting the video into the optimized YOLO26 target detection model, the following processing is also included: decoding the video into a continuous video frame sequence, performing reflection suppression preprocessing on each video frame; after completing the reflection suppression preprocessing, re-encoding / stitching the processed image frames into a video in the original temporal order; and inputting the re-encoded / stitched video into the optimized YOLO26 target detection model.

[0122] In this embodiment, reflection suppression preprocessing is performed on the image (each video frame), specifically including the following steps: T1: After grayscale processing of each video frame, calculate the brightness gradient along the horizontal and vertical directions to construct the pseudo normal field corresponding to the image of that frame; T2: Combining the preset light source direction (azimuth 135°, elevation 60°) and camera view direction, based on the BRDF model (Lambert diffuse reflection term + Blinn-Phong specular reflection term), predict the specular component of the image for this frame (ambient light reflectance coefficient 0.25, diffuse albedo taken as normalized pixel value, specular reflectance coefficient 0.15, specular index 32). T3: Subtract the original frame from the highlight component pixel by pixel to obtain the image frame after eliminating water surface reflection interference. Input the processed frame sequence into the YOLO26 model for detection.

[0123] In this embodiment, firstly, the problem of incomplete method and insufficient disclosure caused by the lack of explanation of video frame splitting processing can be avoided; secondly, frame-by-frame processing ensures that the detection accuracy is consistent with that of static images, and will not introduce new errors due to video compression and inter-frame continuity.

[0124] The results showed that the total number of surface debris was N_total=8720, the number of plastic debris was N_plastic=6834, and the number of non-plastic debris was N_non_plastic=1886.

[0125] S3. Based on the images acquired in step S1, obtain the area of ​​the water body region (coastal zone) to be evaluated. Combined with the total amount of surface debris and the amount of plastic debris obtained in step S2, obtain the Clean Coast Index and Plastic Abundance Index. Simultaneously, combining the preset mapping relationship between the Clean Coast Index and pollution level, and the mapping relationship between the Plastic Abundance Index and abundance level, obtain the pollution level and abundance level of the water body region (coastal zone) to be evaluated, as detailed below: In step S3, the area of ​​the water body (coastal zone) to be evaluated is obtained based on the video obtained in step S1, using the following formula: , Where A represents the area of ​​the water body to be assessed, in units of... The shooting height of the camera, in units of The width of the sensor in the camera, in mm; The relative flight speed of the camera, in units of t represents the camera's flight time, in units of... ; The focal length of the camera, in units of ; In practical applications of step S3, when the camera's flight direction is opposite to the water flow direction (counter-current flight): = |5.0 - 1.2| = 3.8 m / s.

[0126] The results show that the area of ​​the water body to be assessed is A = 11700 m². 2 .

[0127] In other embodiments, when the camera's flight mode is downstream flight, Obtained by the following formula: , in, The camera's flight speed, in units of The water flow velocity in the area to be evaluated is expressed in units of... .

[0128] In other embodiments, when the camera's flight mode is traverse flight, Obtained by the following formula: , in, The camera's flight speed, in units of The water flow velocity in the area to be evaluated is expressed in units of... .

[0129] In other embodiments, when the camera's flight mode is hovering, Obtained by the following formula: , in, The water flow velocity in the area to be evaluated is expressed in units of... .

[0130] In step S3, the Clean Coast Index is obtained using the following formula: CCI = (N_total / A) × K, Wherein, CCI is the Clean Coast Index, N_total is the total amount of surface debris, A is the area of ​​the water body to be assessed, and K is the standardization coefficient, K=20.

[0131] The results show that CCI = (N_total / A) × K = (8720 / 11700) × 20 = 14.91.

[0132] In step S3, the plastic abundance index is obtained by the following formula: PAI=[N_plastic / log 10 [(N_total)] / A×K, Wherein, PAI is the plastic abundance index, N_plastic is the amount of plastic waste, N_total is the total amount of surface debris, A is the area of ​​the water body to be evaluated, and K is the standardization coefficient, K=20.

[0133] The results show that PAI = [N_plastic / log] 10 (N_total)] / A × 20 = (6834 / log 10 (8720)) / 11700 × 20 = (6834 / 3.941) / 11700 × 20 = 2.96.

[0134] In step S3, the preset mapping relationship between the Clean Coast Index and the pollution level includes 5 pollution levels based on the Clean Coast Index: when 0 ≤ CCI < 2, the pollution level is classified as Level 1, and the corresponding pollution description is very low pollution; when 2 ≤ CCI < 5, the pollution level is classified as Level 2, and the corresponding pollution description is light pollution; when 5 ≤ CCI < 10, the pollution level is classified as Level 3, and the corresponding pollution description is moderate pollution; when 10 ≤ CCI < 20, the pollution level is classified as Level 4, and the corresponding pollution description is heavy pollution; when CCI ≥ 20, the pollution level is classified as Level 5, and the corresponding pollution description is extremely heavy pollution.

[0135] In step S4, the preset mapping relationship between plastic abundance index and abundance level includes 5 abundance levels divided according to the plastic abundance index, including: when PAI=0, the abundance level is classified as level 1, and the corresponding plastic threat level is almost no plastic threat; when 0.1≤PAI<1, the abundance level is classified as level 2, and the corresponding plastic threat level is slight plastic threat; when 1≤PAI<4, the abundance level is classified as level 3, and the corresponding plastic threat level is moderate plastic threat; when 4≤PAI<8, the abundance level is classified as level 4, and the corresponding plastic threat level is severe plastic threat; when PAI≥8, the abundance level is classified as level 5, and the corresponding plastic threat level is extremely severe plastic threat.

[0136] S4. Based on the pollution level and abundance level, and combined with the weighted fusion rules, assess the pollution risk level of plastic waste in the water body.

[0137] In step S4, the weighted fusion rules are as follows: when the pollution level is ≥4 and the abundance level is ≥4, it is determined to be extremely high risk; when the pollution level is ≥3 or the abundance level is ≥4, it is determined to be high risk; when the pollution level is ≥2 or the abundance level is ≥2, it is determined to be medium risk; otherwise, it is determined to be low risk.

[0138] Based on the results in step S3, it can be seen that in this embodiment, the pollution level is level 4 and the abundance level is level 3, which is judged as high risk. It is recommended to arrange targeted cleaning operations as soon as possible and to continuously monitor the situation.

[0139] A smart risk assessment system for water pollution by plastic waste is provided, used to assess the risk of plastic waste pollution in water bodies according to the assessment method of Example 2, wherein the smart risk assessment system includes: The acquisition module is used to acquire images or videos of the water body area to be assessed and to calculate the area of ​​the water body area to be assessed.

[0140] The target detection module is used to identify images or videos and output the total number of surface debris and the amount of plastic waste.

[0141] The index calculation module is used to calculate the Clean Coast Index / Plastic Abundance Index and assess the pollution risk level of plastic waste in water bodies.

[0142] The Confusion Matrix and Evaluation module is used to construct a two-index confusion matrix and calculate the macro-average F1 score and Spearman rank correlation coefficient.

[0143] The graphical user interface module provides a human-computer interaction interface for parameter setting, result visualization, and data export.

[0144] A terminal includes a processor and a memory, wherein the memory stores a computer program, and the processor invokes the computer program to perform the steps of the evaluation method of Embodiment 2.

[0145] A readable storage medium storing a computer program, wherein the computer program is invoked by a processor to perform the steps of the evaluation method of Embodiment 2.

[0146] The results above demonstrate that the intelligent risk assessment method for water pollution by plastic waste, based on multi-index fusion, utilizes an optimized YOLO26 target detection model to identify and output the total number of surface debris and the amount of plastic waste in images or videos of the water body to be assessed. This significantly improves detection accuracy, recall, and precision, while significantly reducing inference latency. Specifically, detection accuracy (mAP@0.5) is improved by 0.45%-1.36%, recall by 3.86%, and inference latency is reduced to within 1.5 mm, significantly reducing the false negative rate, especially for small targets such as Small Plastic, and greatly shortening the detection time. Furthermore, based on the collected images or videos, the area of ​​the water body to be assessed is obtained. Combined with the total number of surface debris and the amount of plastic waste, the Clean Coast Index and Plastic Abundance Index can be obtained. Simultaneously, by combining the preset Clean Coast Index with the pollution level... The system establishes mapping relationships between plastic abundance index and abundance level, obtaining the pollution level and abundance level of the water body to be assessed. By fusing the Clean Coast Index and Plastic Abundance Index, it achieves fine semantic separation of 7 types of plastics and 4 types of non-plastics, distinguishing different pollution scenarios and supporting differentiated emergency response decisions. This makes the assessment more comprehensive and completely solves the problem of misleading single indicators. At the same time, the area calculation is more accurate, with photogrammetric area error <0.5%, which is about 30 times more accurate than the existing GPS error >15%. In particular, it supports accurate calculation of non-repeating coverage area using four flight modes of UAV video. Finally, based on the pollution level and abundance level, combined with weighted fusion rules, the pollution risk level of plastic waste in water bodies can be quickly and accurately assessed. It has the advantages of high overall assessment efficiency, comprehensive assessment dimensions, good reliability, and strong applicability. It can directly support emergency response decisions, has high use value, and good application prospects.

[0147] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A smart risk assessment method for water pollution by plastic waste based on multi-indicator fusion, characterized in that, Includes the following steps: S1. Collect images or videos of the water body area to be evaluated; S2. Input the image or video obtained in step S1 into the optimized YOLO26 target detection model to identify and output the total number of surface debris and the amount of plastic debris. Before inputting the image or video into the optimized YOLO26 target detection model, the following processing is also included: performing reflection suppression preprocessing on the image. After completing the reflection suppression preprocessing, the processed image is input into the optimized YOLO26 target detection model. The reflection suppression pretreatment includes the following steps: T1. Estimate the pseudo-normal field corresponding to the image based on the brightness gradient of the image to be processed; T2. Combining the preset light source direction and viewing angle direction, based on the bidirectional reflectance distribution function model, predict the specular component of the image according to the pseudo-normal field. The bidirectional reflectance distribution function model includes a Lambertian diffuse reflectance term and a Blinn-Phong specular term. T3. Subtract the image to be processed from the predicted highlight component to obtain the image after eliminating water surface reflection interference; The method for constructing the optimized YOLO26 object detection model includes the following steps: (1) Collect images of different water areas, label the types and quantities of surface debris in the images to form a dataset, and perform data augmentation on the dataset; the types of surface debris are foam blocks, plastic buoys, fishing nets, ropes, plastic buoys, plastic bottles, foam boxes, other plastics, foam buoys, glass, and metal; (2) Divide the labeled dataset into training set, validation set and test set; (3) Input the training set, validation set and test set into the YOLO26 object detection model for training, testing and evaluation until the detection accuracy is ≥0.890 and the inference delay time is ≤1.5 ms, and obtain the optimized YOLO26 object detection model; In the method for constructing the optimized YOLO26 object detection model, a small object perception label allocation method is used to determine positive and negative samples during the training of the YOLO26 object detection model. The small object perception label allocation method includes the following steps: U1. Determine whether the annotation box is a small target based on the ratio of the area of ​​the annotation box to the area of ​​the image. When the ratio is less than a preset small target area threshold, the annotation box is determined to be a small target. U2. Calculate the alignment score between the candidate anchor point and the bounding box. The alignment score is calculated based on the classification prediction probability of the candidate anchor point and the intersection-union ratio of the candidate anchor point and the bounding box. For the bounding box that is determined to be a small target, the alignment score of its corresponding target is increased by a preset score amplification factor. U3. Select candidate anchor points based on alignment scores. For the bounding boxes that are determined to be small targets, the number of candidate anchor points selected is increased by the preset small target candidate anchor point increment on the basis of the preset base candidate anchor point number. U4. Candidate anchor points are filtered according to the intersection-union ratio (IU) threshold to obtain positive sample anchor points. Among them, the IU threshold of the bounding box that is determined to be a small target is lower than the IU threshold of other bounding boxes. In the optimized YOLO26 object detection model construction method, a progressive loss weight scheduling method is used to determine the loss weights for each training epoch during training. This progressive loss weight scheduling method includes: dynamically adjusting the weights of the bounding box regression loss, classification loss, and distribution focus loss according to a preset scheduling strategy based on the training progress of the current training epoch relative to the total number of training epochs. Specifically, the bounding box regression loss weight is greater in the early training phase than in the later training phase, and the classification loss weight is less in the early training phase than in the later training phase. The preset scheduling strategy includes any one of cosine annealing scheduling, linear scheduling, and step-wise scheduling. S3. Based on the images or videos collected in step S1, obtain the area of ​​the water body to be evaluated. Combine the total amount of surface debris and the amount of plastic debris obtained in step S2 to obtain the Clean Coast Index and Plastic Abundance Index. At the same time, combine the preset mapping relationship between the Clean Coast Index and the pollution level, and the mapping relationship between the Plastic Abundance Index and the abundance level to obtain the pollution level and abundance level of the water body to be evaluated. S4. Based on the pollution level and abundance level, and combined with the weighted fusion rules, assess the pollution risk level of plastic waste in the water body.

2. The intelligent risk assessment method for water plastic waste pollution based on multi-index fusion as described in claim 1, characterized in that, In step S2, based on the image obtained in step S1, the area of ​​the water body to be evaluated is obtained using the following formula: ， Where A represents the area of ​​the water body to be assessed, in meters (m²). 2 ; This refers to the camera's shooting height, in meters (m). This refers to the camera's focal length, in mm. These are the width and height of the sensor in the camera, respectively, in mm; In step S2, based on the video obtained in step S1, the area of ​​the water body to be evaluated is obtained using the following formula: ， Where A represents the area of ​​the water body to be assessed, in units of... The shooting height of the camera, in units of The width of the sensor in the camera, in mm; The relative flight speed of the camera, in units of t represents the camera's flight time, in units of... ; The focal length of the camera, in units of ; When the camera's flight mode is set to fly with the current, Obtained by the following formula: ， in, The camera's flight speed, in units of The water flow velocity in the area to be evaluated is expressed in units of... ; When the camera's flight mode is reverse flight, Obtained by the following formula: ， in, The camera's flight speed, in units of The water flow velocity in the area to be evaluated is expressed in units of... ;when Normal progress is being made, when When hovering; When the camera's flight mode is traverse flight, Obtained by the following formula: ， in, The camera's flight speed, in units of The water flow velocity in the area to be evaluated is expressed in units of... ; When the camera's flight mode is hovering, Obtained by the following formula: The water flow velocity in the area to be evaluated is expressed in units of... .

3. The intelligent risk assessment method for water plastic waste pollution based on multi-index fusion as described in claim 2, characterized in that, In step S3, the Clean Coast Index is obtained by the following formula: CCI = (N_total / A) × K, Wherein, CCI is the Clean Coast Index, N_total is the total amount of surface debris, A is the area of ​​the water body to be assessed, and K is the standardization coefficient, K=20; In step S3, the plastic abundance index is obtained by the following formula: PAI=[N_plastic / log 10 (N_total)] / A×K, Wherein, PAI is the plastic abundance index, N_plastic is the amount of plastic waste, N_total is the total amount of surface debris, A is the area of ​​the water body to be evaluated, and K is the standardization coefficient, K=20; In step S3, the preset mapping relationship between the Clean Coast Index (CCI) and pollution levels includes five pollution levels based on the CCI: when 0 ≤ CCI < 2, the pollution level is classified as Level 1, and the corresponding pollution description is extremely low pollution; when 2 ≤ CCI < 5, the pollution level is classified as Level 2, and the corresponding pollution description is slightly polluted; when 5 ≤ CCI < 10, the pollution level is classified as Level 3, and the corresponding pollution description is moderate pollution; when 10 ≤ CCI < 20, the pollution level is classified as Level 4, and the corresponding pollution description is heavily polluted; when CCI ≥ 20, the pollution level is classified as Level 5, and the corresponding pollution description is extremely heavily polluted. In step S3, the preset mapping relationship between the plastic abundance index and the abundance level includes dividing the plastic abundance index into 5 levels, specifically: when 0 ≤ PAI < 0.1, the abundance level is level 1, corresponding to a plastic threat level of almost no plastic threat; when 0.1 ≤ PAI < 1, the abundance level is level 2, corresponding to a plastic threat level of slight plastic threat; when 1 ≤ PAI < 4, the abundance level is level 3, corresponding to a plastic threat level of moderate plastic threat; when 4 ≤ PAI < 8, the abundance level is level 4, corresponding to a plastic threat level of severe plastic threat; when PAI ≥ 8, the abundance level is level 5, corresponding to a plastic threat level of extremely severe plastic threat. In step S4, the weighted fusion rule is as follows: when the pollution level is ≥4 and the abundance level is ≥4, it is determined to be extremely high risk; when the pollution level is ≥3 or the abundance level is ≥4, it is determined to be high risk; when the pollution level is ≥2 or the abundance level is ≥2, it is determined to be medium risk; and in other cases, it is determined to be low risk.

4. The intelligent risk assessment method for water plastic waste pollution based on multi-index fusion according to any one of claims 1 to 3, characterized in that, In step S1, a camera is used to take pictures or record videos of the water body to be evaluated directly above the water body to be evaluated, and images or videos of the water body to be evaluated are collected. In step S2, the surface debris includes plastic debris and non-plastic debris; the plastic debris includes foam blocks, plastic buoys, fishing nets, ropes, plastic buoys, plastic bottles, foam boxes, other plastics, and foam buoys; the non-plastic debris includes glass and metal. In step S2, before inputting the video into the optimized YOLO26 target detection model, the following processing is also included: decoding the video into a continuous video frame sequence, and performing reflection suppression preprocessing on each video frame; After completing the reflection suppression preprocessing, the processed image frames are re-encoded / stitched into a video in their original temporal order; the re-encoded / stitched video is then input into the optimized YOLO26 target detection model. Step S4 also includes verifying the evaluation results: S41. Based on the Clean Coast Index and Plastic Abundance Index of multiple sampling points, obtain the pollution level and abundance level. Construct a 5×5 confusion matrix with the pollution level as the row and the abundance level as the column, and obtain the precision and recall for each level. S42. Based on the precision and recall of each level, obtain the macro average F1 score and Spearman rank correlation coefficient; S43. Based on the macro-average F1 score and Spearman rank correlation coefficient, assess the degree of pollution and the consistency of indicators.

5. The intelligent risk assessment method for water plastic waste pollution based on multi-index fusion as described in claim 4, characterized in that, In step S41, the accuracy rate is calculated using the following formula: ， in, For the first Precision rate for each category; The element in the k-th row and k-th column of the confusion matrix, where k ≤ 5; Let i be the element in the i-th row and k-th column of the confusion matrix, where i ≤ 5; In step S41, the recall rate is calculated using the following formula: ， in, For the first Recall rate for each category; The element in the k-th row and k-th column of the confusion matrix, where k ≤ 5; Let j be the element in the k-th row and j-th column of the confusion matrix, where j≤5; In step S42, the macro-average F1 score is calculated using the following formula: ， ， in, The macro average F1 score, For the first F1 scores for each category For the first Precision rate for each category For the first Recall rate for each category; In step S42, the Spearman rank correlation coefficient is calculated using the following formula: ， in, The Spearman rank correlation coefficient. ∈[ [1, 1], the closer ρ is to 1, the higher the consistency between the two indicators; d i The difference between the CCI rank and the PAI rank for each sample, where n is the total number of samples; In step S43, when using the Spearman rank correlation coefficient as the evaluation index, the standard for assessing the pollution level and the consistency of the indicators is as follows: when |ρ|≥0.7, the evaluation results are highly consistent, indicating that the evaluation results are objective and reliable; when 0.4≤|ρ|<0.7, the evaluation results are moderately consistent; when |ρ|<0.4, the two indicators need to be evaluated separately; ρ<0 indicates a negative correlation, requiring on-site verification. When using the macro-average F1 score as the evaluation index, the standard for assessing the pollution level and the consistency of the indicators is as follows: when F1... macro When F1 is ≥0.7, the evaluation results are highly consistent, indicating that the classification results of the two indicators are highly consistent; when F1 is ≤0.4, the evaluation results are highly consistent. macro When F1 < 0.7, the evaluation results are moderately consistent; when F1 macro When the value is less than 0.4, the consistency of the evaluation results is low, and the two indicators need to be evaluated separately. When the consistency judgment results given by the macro average F1 score and the Spearman rank correlation coefficient are inconsistent, the lower consistency level of the two shall be taken as the final judgment result, and the strict principle shall be followed. When |ρ| < 0.4 or F1 macro When <0.4, any of the following methods shall be used to assess the degree of pollution and the consistency of the indicators: Method 1: The Clean Coast Index independently determines the pollution risk level of the sampling point according to its preset pollution level classification standards; Method 2: The plastic abundance index independently determines the pollution risk level of the sampling point according to its preset abundance level classification standard; Method 3: If the risk levels determined by Method 1 and Method 2 are inconsistent, the higher risk level shall be taken as the final pollution risk level of the sampling point; Method 4: Trigger the on-site verification process for the sampling point, collect on-site measured data, and verify and correct the judgment results.

6. An intelligent risk assessment system for water pollution caused by plastic waste, characterized in that, Used for risk assessment of plastic waste pollution in water bodies according to the assessment method according to any one of claims 1 to 5; The intelligent risk assessment system includes: The acquisition module is used to acquire images or videos of the water body area to be assessed and to calculate the area of ​​the water body area to be assessed. The target detection module is used to identify images or videos and output the total amount of debris and the amount of plastic waste on the water surface; The index calculation module is used to calculate the Clean Coast Index / Plastic Abundance Index and assess the pollution risk level of plastic waste in water bodies; The confusion matrix and evaluation module is used to construct a two-index confusion matrix and calculate the macro-average F1 score and Spearman rank correlation coefficient. The graphical user interface module provides a human-computer interaction interface, enabling parameter setting, result visualization, and data export.

7. A terminal, characterized in that, It includes a processor and a memory, the memory storing a computer program, the processor calling the computer program to perform the steps of the evaluation method according to any one of claims 1 to 5.

8. A readable storage medium, characterized in that, A computer program is stored, which is invoked by a processor to perform the steps of the evaluation method according to any one of claims 1 to 5.

Images