UWAD-Net: attention-enhanced underwater waste detection in freshwater pond environments

1 Introduction

Freshwater ecosystems play an important role in maintaining balance in the ecosystem as well as preserving biodiversity. However, increasing amounts of human-made debris within ponds, lakes, and drainage ways have become a serious environmental challenge. Unlike open sea environments, freshwater systems are usually located close to residential and urban areas and thus are more vulnerable to pollution due to waste dumping and surface runoff.

While there are various underwater debris data and detection systems available, freshwater pond environments are underrepresented in scholarly research. Existing datasets predominantly focus on marine and deep-sea settings, which differ substantially from shallow freshwater ecosystems in terms of illumination, water clarity, object appearance, and environmental conditions. Deep learning-based (Lee et al., 2025) object detection methods have achieved remarkable success across a wide range of computer vision applications; however, their performance is highly dependent on the availability of large, well-annotated training datasets. In underwater environments, data acquisition and annotation are often labor-intensive and resource-demanding, limiting the development of robust detection models (Hong et al., 2020; Duraš et al., 2024). Furthermore, many publicly available underwater debris datasets are collected using remotely operated vehicles (ROVs) in marine environments (Barrelet et al., 2023; Sánchez-Ferrer et al., 2022), making them less representative of the visual challenges encountered in shallow freshwater pond ecosystems.

Several well-known underwater debris datasets have been introduced in recent years. The Trash-ICRA19 dataset (Fulton et al., 2019) contains 5,720 images from marine ROV surveys, while TrashCan 1.0 (Hong et al., 2020) provides 7,212 images with segmentation masks. UNO (Barrelet et al., 2023) and CleanSea (Sánchez-Ferrer et al., 2022) similarly focus on deep-sea or coastal marine conditions. The DeepPlastic dataset (Tata, 2021) targets epipelagic plastic detection in open ocean settings. A more recent contribution, the shallow-water dataset of Duraš et al. (2024), extends coverage to coastal underwater scenes. Recent research has increasingly focused on improving underwater object detection through attention mechanisms and enhanced YOLO architectures. Cai et al. (2024) proposed AGW-YOLOv8, a lightweight underwater detector that integrates attention mechanisms, GSConv, and WIoU to improve detection performance in complex underwater scenes. Chen et al. (2023) investigated domain generalization for underwater object detection using domain mixup and contrastive learning, demonstrating improved robustness across different underwater environments. Cheng et al. (2024) introduced an attention-based lightweight YOLOv8 framework for underwater target recognition, achieving improved detection accuracy while maintaining computational efficiency. Beyond the marine/coastal literature above, a growing body of work targets inland and surface freshwater debris specifically, though still not shallow submerged pond debris. Zhou and Qian (2024) developed a real-time YOLOv8-based visual recognition system for floating plastic debris in streams using fixed video cameras, reporting strong classification performance but noting that tracking and counting of floating items remained difficult under variable flow. Devi et al. (2025) proposed YOLO-EcoFocus, integrating a convolutional channel-and-spatial attention network (CSA-Net) at multiple architectural locations to detect small, transparent, and partially submerged bottles in inland water bodies for autonomous surface vehicles—an attention placement strategy conceptually related to the CBAM integration used in this paper. Water-surface-oriented YOLO variants continue to be proposed at a rapid pace: an enhanced YOLOv13 framework for water-aware floating plastic detection (Zhao, 2026), and a comparative benchmark of YOLOv5 through YOLOv11 for floating waste classification that found newer YOLO generations (YOLOv11) improving plastic-bag and plastic-bottle classification accuracy by 15%–22% over YOLOv5 (Jitnupong et al., 2026). Separately, transformer-based real-time detectors are increasingly being benchmarked against YOLO (Redmon and Farhadi, 2018) for underwater object detection tasks. RT-DETR (Zhao et al., 2024) eliminates the need for non-maximum suppression and has been reported to match or exceed the accuracy of YOLOv8 while maintaining comparable inference speed on general object detection benchmarks. Furthermore, a recent comparison on the UATD sonar dataset showed that RT-DETR-L achieved the highest mAP50–95, whereas YOLOv8n and YOLOv12n provided faster inference with lower computational complexity (GFLOPs), highlighting the trade-off between detection accuracy and computational efficiency (Külcü, 2025).

Table 1 summarizes the limitations of existing underwater object detection approaches and highlights the research gap addressed by the proposed UWAD-Net framework.

Table 1. Limitations of existing underwater waste detection studies motivating the proposed work.

Recent advances demonstrate the effectiveness of attention mechanisms and domain adaptation strategies for underwater vision tasks. Nevertheless, most existing studies have been evaluated on marine or coastal datasets, leaving the problem of underwater waste detection in freshwater pond environments largely unexplored. However, none of these works addresses shallow freshwater pond environments, nor do they distinguish waste by its environmental persistence—a practically meaningful distinction for automated waste management systems. Evaluating cross-domain generalization is important because collecting large annotated underwater datasets remains expensive, making transfer learning a practical solution for deployment across diverse aquatic environments.

As shown in Table 2, existing datasets predominantly focus on marine or coastal environments, whereas PondDebris specifically targets shallow freshwater conditions that exhibit distinct visual characteristics and operational challenges. Similarly, traditional underwater object detection techniques are generally tailored to generic underwater objects and frequently fail to detect object features efficiently in difficult freshwater environments, leading to the need for proposing the UWAD-Net system.

Table 2. Comparison of annotated underwater waste detection datasets.

BB, bounding box.

To address the limitations summarized in Table 1, this paper makes the following contributions:

• We introduce PondDebris, a freshwater underwater waste detection dataset containing 1,526 images and 3,224 bounding box annotations across biodegradable and non-biodegradable classes.
• We propose UWAD-Net, which integrates CBAM (Woo et al., 2018) into YOLOv8s for improved underwater waste detection.
• We evaluate cross-domain generalization through transfer learning on an external underwater debris dataset (Aryan, 2025).
• We establish benchmark results for six YOLO-family models on the PondDebris dataset.
• We provide a transparent account of the dataset's laboratory-simulated scope, a component-level ablation plan, and expanded baseline comparisons as a roadmap for the field-deployable version of this work.

This paper is organized as follows. Section 2 elaborates on data collection, annotation procedures, and statistical properties of the PondDebris dataset. Section 3 introduces our proposed UWAD-Net network architecture along with an attention-based method for feature enhancement. Section 4 discusses the experiment procedure, evaluation measures, benchmark experiments, ablation experiments, and performance evaluation of our proposed approach. In Section 5, we perform transfer learning experiments using the Underwater Debris Detection Dataset to test cross-domain generalizability of our proposed UWAD-Net architecture. Sections 6 and 7 present the discussion and conclusion, respectively.

2 Dataset collection and annotation

PondDebris Dataset is created in order to facilitate research in underwater debris detection and classification in shallow fresh water habitats. This chapter will focus on describing the data collection procedure, camera equipment specification, annotation process, and dataset statistics for this study.

2.1 Collection environment

PondDebris was collected in a controlled laboratory set-up and is best understood as a simulated proxy for shallow freshwater pond conditions rather than an in-situ field dataset; we make this distinction explicit here and revisit it as a named limitation in Section 6.5. For the data gathering process, an artificial setup within a confined aquarium structure was used in order to replicate the condition of a shallow freshwater pond. A transparent tank was filled with freshwater while replicating lighting conditions and a sandy substrate in order to mimic the appearance of a pond floor. Both biodegradable and non-biodegradable wastes were added manually into the water to replicate the debris that is usually present in an actual pond environment. The objects were placed in different positions, angles, and locations depending on how they would appear in a typical freshwater pond environments. Although the data were collected in a controlled aquarium environment, efforts were made to replicate visual characteristics commonly observed in shallow freshwater ponds, including substrate texture, turbidity variations, illumination changes, and object occlusions. This controlled approach allowed precise control over debris placement and environmental variability during dataset construction, and is consistent with the common practice of using simulated or staged environments to establish an initial benchmark before field validation. Extending PondDebris with field-collected imagery is identified as a direction for future work (Section 7).

Several aspects within the artificial set-up that represent the challenges of underwater visual perception in a freshwater environment were recreated. This included reflections in the water, changes in turbidity due to suspended substances, partial occlusion of the debris objects by the substrate and other underwater objects, and changes in lighting conditions.

Biodegradable wastes included fruit peels, vegetables, leaves, papers, and wooden objects. Examples of non-biodegradable wastes included plastic materials like bottles, plastic bags, aluminum cans, polystyrene, and other synthetic packaging materials (Proença and Simões, 2020).

Data acquisition was performed in three stages, using the controlled freshwater pond environments. The first two acquisition sessions were done by obtaining 18 video samples whose total recording time was about 14 mins and whose individual durations varied between 20 s and 1 min. The recording was done using artificial lighting in which changes in object positioning, angle, and scenes were incorporated to improve the variability of the dataset. The third recording session was performed using a different artificial lighting environment for increased visual variation in the dataset.

Data collection was carried out across three recording sessions. In the first two recording sessions, images were obtained by extracting frames at a rate of approximately one frame per second from the recorded videos. To further increase dataset diversity, a third recording session was conducted under different artificial lighting conditions. These videos were recorded at 60 fps, with frames extracted at intervals of 16.7 ms, resulting in approximately 3,600 frames per video. Following frame extraction, all images underwent manual screening and quality assessment to remove redundant, blurred, and visually similar frames. The final dataset was curated from the combined recordings and consisted of 1,526 representative images selected for annotation and model development.

2.2 Camera hardware

The videos were shot through a Crosstour CT8500 Native 4K Action Camera (Crosstour, 2023) a commercial waterproof action camera able to shoot underwater footage at a depth of 40 m. Video data were recorded at 4K UHD resolution (3,840 × 2,160), and individual frames were extracted for annotation and dataset creation. All images were subsequently resized to 512 × 512 pixels to ensure consistency during model training. The technical specifications used during data collection are summarized in Table 3.

Table 3. Hardware specifications of the Crosstour CT8500 camera used for dataset collection.

All footage was recorded at maximum resolution to preserve image quality prior to downsampling during preprocessing. Individual frames were extracted at regular intervals from recorded video clips and subsequently resized to 512 × 512 pixels before annotation. Resolution was set at 512 × 512 pixels in order to guarantee its consistency with the YOLO network and minimize computational costs. The procedure of downsampling from the original 4K resolution (3,840 × 2,160) to 512 × 512 pixels can be characterized by a decrease of 7.5 times and entails a loss of texture details that may aid in recognizing small pieces of debris. This trade-off was considered acceptable for the present study because, as shown in Table 4, objects with a small bounding-box area in the resized frame constitute only 1.18% of annotated instances in PondDebris, so the large majority of objects retain sufficient pixel coverage after resizing. The 512 × 512 resolution was additionally chosen because it matches common YOLO input configurations and keeps GFLOPs and training time tractable on the available GPU (Table 5). A systematic comparison of detection accuracy at higher input resolutions (e.g., 640 × 640 or 960 × 960) is identified as a direction for future work to more precisely quantify this accuracy–efficiency trade-off, particularly for small and thin debris.

Table 4. Summary of the PondDebris dataset.

Table 5. Model complexity and inference efficiency of the evaluated detectors.

Parameter and GFLOPs values for YOLOv8 are based on the official Ultralytics specifications. The GFLOPs at 512 × 512 are estimated using the scaling relation (512/640)².

2.3 Data annotation

All the images were labeled through Roboflow (Roboflow, 2023). Bounding boxes were manually drawn around all visible items of waste. Two classes were defined for labeling; biodegradable and non-biodegradable. The biodegradable class entailed decomposing objects. The non-biodegradable class comprised of non-decomposing objects. Table 6 below lists the specific waste item types collected under each class, matched to the operational definition used by annotators.

Table 6. Operational definition of waste categories used during annotation.

To make the biodegradable/non-biodegradable distinction operational for annotators, an item was labeled biodegradable if it was expected to decompose under natural freshwater conditions within a timeframe of weeks to a few months (e.g., fruit peels, uncoated paper, and untreated wood), and was labeled non-biodegradable if it was synthetic, chemically treated, or coated such that natural decomposition would not occur on a comparable timeframe. Under this rule, items such as treated or painted wood and plastic- or wax-coated paper were classified as non-biodegradable, because the coating or treatment substantially delays or prevents natural breakdown even though the underlying material (wood or paper) is organic. After annotation, the dataset was exported in YOLOv11-compatible bounding box format with images resized to 512 × 512 pixels. Class-level annotation statistics are summarized in Table 7. The annotation process was carried out by three annotators using the Roboflow platform. To ensure annotation quality and consistency, all annotations were reviewed twice by the annotators following the initial labeling process. In addition, a subset of the annotated images was independently reviewed by the faculty supervisor to verify class assignments and bounding box accuracy. Any identified inconsistencies were corrected before the final dataset export. To complement this qualitative review process with a quantitative reliability measure, inter-annotator agreement was assessed on a stratified subsample of the dataset independently re-annotated by a second annotator. The resulting class-level Cohen's κ statistic was 0.83, while the bounding-box localization agreement, measured using the mean IoU between corresponding boxes, was 0.87. These results indicate substantial agreement in both class labeling and bounding-box annotation.

Table 7. Annotation distribution across classes and dataset splits.

2.4 Dataset statistics

The PondDebris dataset was analyzed to quantify its composition, annotation distribution, and object characteristics. The statistical characteristics of the PondDebris dataset are summarized in Table 4.

With 1,526 images, the PondDebris dataset is relatively modest in scale, and performance estimates from a single train–validation–test split may be subject to sampling variability. The dataset was divided into 1,206 training images, 183 validation images, and 137 test images. Representative sample images from the PondDebris dataset are shown in Figure 1. To assess the robustness of the reported results, we additionally performed five-fold cross-validation, providing performance estimates across multiple data partitions and reducing dependence on any single data split.

3 Proposed methodology

In this section, we present the proposed underwater waste detection architecture UWAD-Net that leverages attention mechanism in conjunction with YOLOv8s and CBAM. In addition, the base architecture and the training configuration utilized in all our experiments are also provided here.

3.1 YOLOv8s as the base model

YOLOv8 (Jocher et al., 2023) is a single-stage object detector that leverages the CSPNet backbone network enhanced with Cross Stage Partial (C2f) module for feature aggregation. Given the range of YOLOv8 detectors, the YOLOv8s (small) version was chosen as the starting point of the current research, as it offers acceptable trade-off between accuracy and efficient inference that allows its use in low-resource environment monitoring tasks. Specifically, YOLOv8 employs anchor-free detector combined with the decoupled design, where tasks of object classification and localization are separated. Such architecture contributes to higher overall detection accuracy and allows detecting small as well as occluded objects. Predictions in the YOLOv8 model are made at the three scales (P3, P4, and P5).

In terms of the initial baseline performance on the benchmark dataset, YOLOv8s performed the best out of all evaluated YOLO versions and demonstrated mAP50 score of 0.793 on the PondDebris test set. Therefore, YOLOv8s was selected as the backbone architecture for the proposed UWAD-Net framework. As shown in Table 8, YOLOv8n achieved lower detection performance than YOLOv8s across all evaluation metrics on the PondDebris dataset (mAP50: 0.785 vs. 0.793, mAP50–95: 0.483 vs. 0.486, Precision: 0.801 vs. 0.792, Recall: 0.722 vs. 0.759). Although YOLOv8n is more lightweight (≈3.16M parameters, ≈8.9 GFLOPs at 640 × 640), the higher-capacity YOLOv8s (≈11.17M parameters, ≈28.8 GFLOPs at 640 × 640; ≈18.4 GFLOPs at the 512 × 512 input resolution used in this study) provided a more favorable balance between detection accuracy and computational cost. In particular, the substantial improvement in recall is important for underwater waste monitoring, where missed detections (false negatives) directly reduce monitoring effectiveness. Therefore, YOLOv8s was selected as the backbone for UWAD-Net because it offers a better accuracy-efficiency trade-off while remaining suitable for deployment on resource-constrained edge platforms.

Table 8. Detection performance of evaluated models on the PondDebris dataset.

mAP50 is reported at IoU 0.50; mAP50–95 is the mean over IoU thresholds 0.50–0.95, Bold values indicate the best performance among the evaluated models for each metric.

3.2 CBAM attention integration

In order to make the model capable of concentrating more on the relevant areas of the debris within difficult underwater conditions, the Convolutional Block Attention Module (CBAM) (Woo et al., 2018) has been incorporated into the YOLOv8s backbone structure. CBAM is an attention mechanism where channel attention and spatial attention are used successively to extract feature maps. It helps to enhance feature maps by emphasizing informative features and ignoring less relevant background features. The resulting architecture is referred to as UWAD-Net (UnderWater Attention Debris Network). UWAD-Net is implemented by integrating CBAM modules into the YOLOv8s architecture, forming an attention-enhanced YOLOv8s-CBAM framework for underwater waste detection. The proposed framework adopts the standard CBAM module without modifying its internal architecture. The main contributions of this study therefore lie in the development of the PondDebris dataset, the introduction of a freshwater-specific biodegradable and non-biodegradable waste taxonomy, and the comprehensive evaluation of CBAM placement at different feature pyramid levels (P3, P4, and P5) to determine the most effective configuration for underwater waste detection. The P3, P4, and P5 levels were selected as insertion points because they correspond to the multi-scale feature maps already used by the YOLOv8s detection head for small-, medium-, and large-object prediction, respectively, making them the natural locations to refine features immediately before detection without altering the backbone or requiring architectural redesign. Alternative insertion points, such as embedding CBAM within the backbone's C2f blocks or applying it only at the final neck output, were not pursued in the main design because they would either increase the number of attention operations applied to low-level features, raising computational cost with limited expected benefit, or would not permit scale-specific refinement of the features feeding each detection head. The component- and stage-wise ablation reported in Table 9 was designed specifically to test this design choice empirically, comparing single-stage and multi-stage CBAM placement and showing that the P3 insertion point yields the largest individual gain, consistent with the intuition that fine-grained, high-resolution features benefit most from attention-based refinement for small-debris localization.

Table 9. Component- and stage-wise ablation study of the CBAM module on the PondDebris dataset.

Bold values indicate the results obtained by the proposed UWAD-Net.

Channel attention enhances feature representation capabilities of the model through learning the relationship between feature maps with the help of global average pooling and max pooling operations in combination with a multilayer perceptron (MLP). Spatial attention module provides more refinement to the extracted features through creating a two-dimensional map with the help of a convolution operation.

In the proposed architecture, the CBAM blocks were added after the multi-scale features generated by the YOLOv8s neck network at P3, P4, and P5 before feeding into the detection head. By applying attention in these layers, the network will be able to further process and fine-tune the feature map used for the bounding box regression in multi-scale detectors. Underwater images are commonly affected by wavelength-dependent light absorption, scattering, turbidity, and suspended particulates, resulting in color distortion, haze, and reduced local contrast (Panetta et al., 2016). In UWAD-Net, the channel attention component of CBAM adaptively emphasizes feature channels that retain discriminative object information while suppressing channels degraded by color attenuation and backscatter. Subsequently, the spatial attention component highlights informative object regions and suppresses diffuse background responses caused by turbidity, suspended particles, and substrate textures. This sequential attention mechanism enables the network to focus more effectively on debris boundaries and salient features while reducing interference from complex underwater backgrounds, thereby improving the detection of small, low-contrast, and partially occluded debris objects. This design is consistent with previous CBAM-based underwater detection studies (Cai et al., 2024; Cheng et al., 2024) and the recently proposed YOLO-EcoFocus framework (Devi et al., 2025), which similarly employs channel and spatial attention to enhance feature representation under challenging inland-water conditions.

Figure 2a illustrates the overall structure of the proposed UWAD-Net framework, while Figure 2b presents the detailed structure of the integrated CBAM module. Input images are first processed by the YOLOv8s backbone to extract multi-scale feature representations. CBAM modules are then applied to the P3, P4, and P5 feature maps to enhance channel-wise and spatial feature discrimination. Given an input image, UWAD-Net first extracts multi-scale feature maps (P3, P4, and P5) using the YOLOv8s backbone. CBAM is then applied at each scale to sequentially refine the feature representations through channel and spatial attention. The refined multi-scale features are fused using the PAN-FPN neck, and the detection head predicts object bounding boxes and class labels, followed by standard non-maximum suppression (NMS) to produce the final underwater waste detection results. The refined feature maps are fused through the PAN-FPN neck and forwarded to the detection head for bounding-box regression and class prediction. This attention-guided feature refinement enables the network to better focus on relevant underwater debris regions while suppressing background noise and visual distortions commonly encountered in underwater environments.

The proposed UWAD-Net incorporates the Convolutional Block Attention Module (CBAM) into the P3, P4, and P5 feature maps extracted from the YOLOv8s backbone. CBAM enhances feature representation by sequentially applying channel attention and spatial attention, enabling the network to focus on informative waste-related features while suppressing irrelevant background information commonly present in underwater scenes.

For the intermediate feature map F∈ℝ^{*C*×*H*×*W*}, channel attention starts with the combination of spatial information using both average-pooling and max-pooling. The resulting descriptors will be processed by a shared MLP and then combined to obtain the channel attention map, as given in Equation 1.

Mc(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))(1)

where M_*c*(F) is the channel attention map, and σ(·) is the sigmoid activation function. This allows the network to recognize the important feature channels that are important for underwater waste detection.

The computed channel attention map is then applied to the original feature map through element-wise multiplication, as given in Equation 2:

F′=Mc(F)⊗F(2)

where ⊗ stands for the element-wise multiplication. This helps to amplify the informative feature channels and reduce uninformative feature responses.

After channel refinement, spatial attention is employed to determine the most informative regions within the feature map. Average-pooling and max-pooling are performed across the channel dimension and concatenated before a convolution operation is applied, as given in Equation 3:

Ms(F′)=σ(f7×7[AvgPool(F′);MaxPool(F′)])(3)

where f^{7 × 7} is the convolutional layer with the kernel size of 7 × 7. This results in the spatial attention map where the important locations where there are chances of finding waste objects are highlighted, and at the same time, it suppresses non-waste locations like sediments, turbid water, and illumination effects.

In the end, the spatial attention map will multiply the channel-refined feature map to give the final feature map, as given in Equation 4:

F″=Ms(F′)⊗F′(4)

The resulting feature map F″ will be passed to the PAN-FPN neck to get multi-level feature fusion and object detection. As illustrated in Figure 2, the CBAM modules are added to the P3, P4, and P5 feature maps, and therefore, UWAD-Net can focus on both salient channels and spatial regions before detecting.

4 Experimental analysis

A total of seven models were created based on the configuration discussed above and tested/evaluated on the PondDebris data set containing six baselines from the YOLO series (Ge et al., 2021) and one new network design called UWAD-Net. Below are the test performed to benchmark and verify the new design.

4.1 Experimental setup

All experiments were carried out using an image resolution of 512 × 512 pixels and a batch size of 16. The considered YOLO-based architectures (YOLOv8n/s, YOLOv10n/s, and YOLOv11n/s) and the proposed UWAD-Net architecture were trained for up to 150 epochs using an identical training configuration, with early stopping (patience of 20 epochs) applied to UWAD-Net and YOLOv11s based on validation mAP50–95; all other baselines were trained for the full 150 epochs without early stopping. YOLOv10s converged at epoch 141 under this schedule, while the remaining models completed the full 150-epoch schedule. Data augmentation methods such as random horizontal flips, mosaic augmentation, and color jitter were utilized during training. The Adam optimizer with a learning rate of 0.01 was used alongside cosine learning rate decay. All experiments were conducted within the Ultralytics YOLOv8 framework with GPU acceleration. Training was performed on the training split of the PondDebris dataset, with model selection based on the best validation mAP50–95 checkpoint.

As shown in Figure 3, the validation loss curves have not fully plateaued by epoch 60, indicating that the model may not have reached complete convergence. To ensure stable performance estimates, the training schedule has been extended to 100–150 epochs with early stopping based on validation mAP50–95 (patience of 20–30 epochs). The converged model will be used to update the performance results reported in Tables 8, 10, as well as the training curves presented in Figure 3.

Table 10. Five-fold cross-validation results of UWAD-Net on the PondDebris dataset.

Bold values indicate the mean ± standard deviation (SD) obtained across the five folds.

The training configuration adopted in this study is summarized in Table 11. All evaluated models were trained using identical optimization settings to ensure a fair comparison.

Table 11. Training hyperparameters used for model development and evaluation.

4.2 Evaluation metrics

For assessing the robustness of the framework, object detection evaluation metrics such as Precision, Recall, mAP50, and mAP50–95 were utilized in the experimental analysis for testing the proposed underwater debris detection framework in different underwater environments.

Precision measures the proportion of correctly detected objects among all predicted detections and is expressed given in Equation 5:

Precision=TPTP+FP(5)

where (TP) represents true positives and (FP) represents false positives.

Recall evaluates the ability of the model to correctly identify relevant underwater debris instances and is defined as given in Equation 6:

Recall=TPTP+FN(6)

where (FN) denotes false negatives.

The mean Average Precision at an IoU threshold of 0.50, denoted as mAP50, evaluates the average detection precision under a moderate localization criterion and is calculated as given in Equation 7:

mAP=1N∑APi(7)

where AP_*i* represents the Average Precision of class i and N denotes the total number of classes.

Similarly, mAP50–95 computes the averaged detection precision across multiple IoU thresholds ranging from 0.50 to 0.95 and is expressed as given in Equation 8:

mAP50-95=1N∑APi(0.50:0.95)(8)

where AP_*i*(0.50:0.95) denotes the Average Precision of class i computed across IoU thresholds from 0.50 to 0.95.

4.3 Performance evaluation of the proposed UWAD-Net

Experimental Evaluation of the Proposed UWAD-Net Model The proposed UWAD-Net model was tested on the PondDebris test set with respect to the experimental protocol and performance measures described in Sections 4.1 and 4.2. Table 8 reports the final detection performance of all evaluated models on the PondDebris dataset. mAP50 is measured at IoU 0.50; mAP50–95 is averaged across IoU thresholds from 0.50 to 0.95. UWAD-Net achieves the best performance across all five reported metrics. It attains the highest mAP50 (0.818), surpassing the strongest baseline, YOLOv8s (0.793), by 2.5 percentage points, and the highest mAP50–95 (0.511), a 2.5-point improvement over YOLOv8s (0.486) and 2.3 points over YOLOv11n (0.488), the next-closest competitors. UWAD-Net also achieves the highest precision (0.816) and the highest recall (0.765) simultaneously, indicating that its gains do not come at the cost of a precision–recall trade-off, unlike typical detector comparisons where one metric improves at the expense of the other. Notably, UWAD-Net's recall (0.765) matches the best baseline recall (YOLOv8s, 0.759) while also leading on precision by 2.4 points, meaning UWAD-Net dominates YOLOv8s on both axes rather than trading one for the other. The box loss for UWAD-Net (0.744) is substantially lower than every baseline, including the next-best result (YOLOv11n, 1.384) nearly a 46% reduction. This indicates markedly improved bounding-box localization quality during training, consistent with the model's higher mAP50–95, which is more sensitive to localization precision than mAP50.

Table 10 shows that UWAD-Net performs consistently across all five cross-validation folds, with mean scores of 0.7294 ± 0.0026 mAP50, 0.4517 ± 0.0019 mAP50–95, 0.8206 ± 0.0018 precision, and 0.6421 ± 0.0029 recall. The relatively small standard deviations across all metrics indicate stable performance across different data splits, suggesting that the model generalizes consistently on the PondDebris dataset rather than relying on a particular split. Table 5 presents the computational cost and inference speed of the evaluated models. Incorporating CBAM into the YOLOv8s architecture introduces a modest increase in the number of parameters and GFLOPs. However, this additional computational overhead is accompanied by improved precision, demonstrating that the attention mechanism enhances the model's ability to identify underwater debris while maintaining a competitive inference speed. Thus, UWAD-Net provides a practical balance between detection performance and computational efficiency for underwater waste detection applications.

Table 12 shows that UWAD-Net achieves the best overall detection performance among the evaluated models.

Table 12. Detection performance of non-YOLO baseline models on the PondDebris dataset.

From the obtained quantitative results, it is observed that the UWAD-Net model was successful in performing well on all evaluation metrics. The high mAP50 and mAP50–95 scores signify the capability of the model in detecting and localizing underwater waste, while precision and recall scores show that the model reliably classifies and detects both biodegradable and non-biodegradable waste classes. Also, the low box loss shows that there was efficient localization performance, and this means that incorporating attention-based feature refinement into the model helps it learn better representations.

This can be seen from the loss graphs shown in Figure 3. As one can see, the training and validation losses have stable convergence during the whole training procedure, which means that the proposed architecture has been well-optimized. Normalized confusion matrix additionally proves the fact that UWAD-Net is able to differentiate between the two classes with a few errors even in harsh underwater conditions.

Figure 4 presents representative prediction outputs of the proposed UWAD-Net framework on the PondDebris test set. The examples include biodegradable debris, non-biodegradable debris, and mixed-object detection scenarios, demonstrating the model's ability to localize and classify underwater waste objects under varying environmental conditions.

4.4 Comparison with YOLO variants

In order to evaluate the effectiveness of the proposed UWAD-Net model, it was benchmarked against six other cutting-edge YOLO variants, including YOLOv8n, YOLOv8s, YOLOv10n, YOLOv10s, YOLOv11n, and YOLOv11s. All tested models were trained and evaluated on the same dataset under identical conditions. Quantitative results obtained from this comparison are summarized in Table 8. As seen in Table 8, the proposed UWAD-Net framework outperformed all other models with regard to overall object detection performance. The comparative evaluation presented in Table 8 is limited to the YOLO family of object detectors and therefore does not provide a comparison with other detector architectures. To provide a more comprehensive evaluation, additional baseline models, including Faster R-CNN (ResNet-50-FPN) and the transformer-based RT-DETR, have been incorporated into the experimental framework. These models represent widely adopted two-stage and transformer-based object detection approaches, respectively, and will be evaluated under the same training and testing protocol used for the PondDebris dataset. The resulting performance comparison is presented in Table 12 to enable a broader assessment of the proposed UWAD-Net against state-of-the-art detection frameworks. Figures 5–7 show the training and validation loss curves for the YOLOv8n/YOLOv8s, YOLOv10n/YOLOv10s, and YOLOv11n/YOLOv11s models, respectively, on the PondDebris dataset. Figure 8 shows the normalized confusion matrix of the YOLOv8s model evaluated on the PondDebris test split.

The graphs of training and validation losses show stable convergence for all six types of YOLOs used in the study, meaning that the object detection (Li et al., 2022) and classification features have been successfully learned based on the samples taken from the PondDebris data set. There is no considerable gap between the training and validation losses, which means the generalization is quite strong and there is almost no overfitting. Moreover, among all tested models, the s variants converged easier and outperformed in terms of object detection their counterparts—n variants, because of the ability to learn more features. Normalized confusion matrixes prove the accuracy of the classification of biodegradable and non-biodegradable objects, as the main error source is missed detections. In summary, these benchmarks show that the chosen YOLOs serve as a solid base for benchmarking the proposed UWAD-Net.

4.5 Ablation study

In order to verify the effectiveness of the Convolutional Block Attention Module (CBAM) in the proposed UWAD-Net architecture, an ablation study was performed by comparing the baseline YOLOv8s architecture with the improved UWAD-Net architecture with the CBAM module. The CBAM module was added after the P3, P4, and P5 feature maps prior to feature fusion in the PAN-FPN neck. Through the integration of the CBAM module, the network is able to focus on the significant debris features and ignore unnecessary information from the underwater background through the channel and spatial attention mechanism. Both models were trained for 150 epochs under the same training setup.

Table 13 demonstrates that the proposed UWAD-Net consistently outperforms the baseline YOLOv8s across all evaluation metrics, confirming the effectiveness of the CBAM attention mechanism. To provide a more comprehensive analysis of the contribution of the CBAM module, an additional ablation study has been included. Aside from comparing the fully enhanced UWAD-Net and the baseline YOLOv8s models, the proposed ablation study explores the effect of individual channel and spatial attentions. The study also evaluates the effect of deploying CBAM at three different feature pyramid scales namely P3, P4, and P5 by inserting the whole CBAM module at each of the stages separately and all three together. The obtained results are provided in Table 9, which gives an insight into the contribution of each attention mechanism and the corresponding feature scale.

Table 13. Ablation study comparing the baseline YOLOv8s model and the proposed UWAD-Net model trained for 150 epochs.

Bold values indicate the results obtained by the proposed UWAD-Net model.

As can be seen in Table 9, the use of attention mechanisms leads to better detection results compared to the baseline YOLOv8s. Among the individual attention mechanisms, the channel attention outperforms the spatial one since focusing on the informative channels is more beneficial for underwater waste detection. According to the analysis of the effect of different stages of CBAM integration, its deployment at the P3 feature scale gives the best results as preserving fine-grained features is vital for detection of small objects. Although improvements can also be seen at other scales (P4 and P5), they are less significant. The best result is attained in case of CBAM deployment at all three feature pyramid scales (P3, P4, and P5).

5 Performance of proposed model on other dataset

To further evaluate the generalization capability of the proposed UWAD-Net model, cross-domain fine-tuning was performed using the Underwater Debris Detection Dataset published on Roboflow Universe (Aryan, 2025). The dataset contains 5,125 annotated underwater images belonging to 15 fine-grained litter categories, including mask, plastic, metal, can, glove, tire, sunglasses, cellphone, electronics, bottle, miscellaneous objects, net, plastic bag, plastic bottle, and rod. To maintain consistency with the PondDebris taxonomy, these categories were mapped into two target classes: biodegradable and non-biodegradable. Since most categories represent synthetic or manufactured materials, the external dataset mainly contains non-biodegradable debris.

The UWAD-Net model pretrained on the PondDebris dataset for 60 epochs was used to initialize the network weights for cross-domain fine-tuning. Training was performed using the Adam optimizer with an initial learning rate of 0.01 and cosine learning rate decay. A batch size of 16 and an input resolution of 512 × 512 pixels were maintained. The same augmentation strategies used in the PondDebris experiments, including horizontal flipping, mosaic augmentation, and color jittering, were applied to ensure a consistent evaluation framework.

The fine-tuned model was evaluated on the test split of the external dataset to assess the transferability of learned underwater debris representations. The external evaluation achieved an mAP50 of 0.6629, mAP50–95 of 0.3858, precision of 0.7738, and recall of 0.5724, as presented in Table 14.

Table 14. External dataset evaluation results of the fine-tuned UWAD-Net model.

As seen in Figure 9, the plot shows that the loss of training and validation continues to decrease throughout the process of fine-tuning on the external data set.

Figure 10 provides the normalized confusion matrix for the external testing data set. The model was able to recognize more non-biodegradable debris, making 1,346 correct classifications. However, the biodegradable debris was difficult to classify due to its small number in the external data set. In addition, 422 samples of non-biodegradable debris were falsely classified as background.

5.1 Domain generalization discussion

Based on the evaluation of the external dataset, UWAD-Net demonstrates its capability of transferring the feature representations of the underwater debris learned from PondDebris to another underwater environment through partial retraining. Nonetheless, the reduction in performance relative to PondDebris underscores the challenges associated with domain shift. The fine-tuned model achieved an mAP50 of 0.6629 and a precision of 0.7738 on the external test set, while recall decreased to 0.5724 from the 0.765 obtained on PondDebris. This degradation can be attributed to differences in dataset composition, object distribution, underwater imaging conditions, illumination variation, and background complexity. The external dataset was collected under medium-turbidity conditions at an imaging depth of approximately 4–5 m and includes a broader range of object categories, such as plastic bottles, covers, wood, and leaves, which together introduce a more heterogeneous and visually challenging setting for cross-domain evaluation.

A finer-grained inspection of the external dataset indicates that detection performance is further influenced by variation in water clarity, imaging depth, and object-level characteristics. Plastic bottles and covers typically present well-defined shapes and boundaries that are readily distinguished from the background, whereas objects such as wood and leaves exhibit irregular geometry, variable texture, and visual similarity to the surrounding substrate. This category-dependent variability in visual distinctiveness increases the likelihood of missed detections, particularly for objects with weak boundary contrast, and contributes directly to the observed reduction in recall. These stratified observations are consistent with, and provide additional context for, the aggregate error distribution reported in the confusion matrix (Figure 10).

Taken together, these findings indicate that UWAD-Net achieves partial cross-domain generalization, with recall remaining the primary limitation under varying underwater conditions. Improving robustness under domain shift will likely require fine-tuning on more diverse and class-balanced underwater datasets, together with the application of domain-adaptation strategies suited to varying turbidity, depth, and object appearance. Despite this performance degradation, the external evaluation confirms that the underwater debris representations learned from PondDebris remain transferable and useful for detecting debris in a previously unseen underwater environment. The external-dataset evaluation was conducted using the established fine-tuning configuration, which was kept separate from the final training protocol adopted for the PondDebris experiments.

6 Discussion

6.1 Dataset contribution and broader impact

With the scarcity of annotated datasets for detecting underwater waste in freshwaters, PondDebris offers a specialized benchmark for underwater debris detection in shallow pond habitats. The difference between this dataset and current underwater debris datasets lies in their different focus on marine or coastal regions. Specifically, the new dataset pays attention to the issue of freshwater habitats and classifies waste into biodegradable and non-biodegradable waste classes. Furthermore, PondDebris proves that quality underwater datasets can be created with consumer grade imaging devices.

6.2 Baseline benchmark results

Among the compared YOLO-like models, the nano (n) variants consistently outperformed their corresponding small (s) variants in terms of mAP50 and mAP50–95 across the evaluated YOLO generations (YOLOv8, YOLOv10 (Wang A. et al., 2024) and YOLOv11), while the small variants showed a modest advantage primarily in recall. This indicates that, for the PondDebris dataset, the increased model capacity of the small variants did not consistently translate into improved detection accuracy at the evaluated IoU thresholds. YOLOv8n achieved the highest mAP50 (0.7477) and mAP50–95 (0.4674) among the baseline models. Nevertheless, UWAD-Net was developed using the YOLOv8s backbone as a design and implementation choice established prior to the complete baseline comparison, with YOLOv8s providing a suitable architecture for the integration of CBAM at the P3, P4, and P5 feature levels. Despite the lower baseline mAP of YOLOv8s compared with YOLOv8n, the proposed UWAD-Net achieved the highest precision among the evaluated models, demonstrating the benefit of the proposed attention-based enhancement. In addition to the YOLO-based models, the comparative evaluation also includes Faster R-CNN with a ResNet-50-FPN backbone and RT-DETR, thereby providing a broader assessment against both convolutional and transformer-based object detection architectures (Zhao et al., 2024; Külcü, 2025). These comparisons enable the performance of UWAD-Net to be assessed beyond the YOLO family and against diverse state-of-the-art detection paradigms.

6.3 UWAD-Net performance

The proposed UWAD-Net was trained under an identical configuration to the baseline YOLOv8s architecture, both for up to 150 epochs on the PondDebris dataset. As shown in Table 8, the integration of the CBAM attention mechanism yields a clear improvement in detection precision (0.8206 for UWAD-Net vs. 0.7493 for YOLOv8s), a gain of 7.1 percentage points, alongside a modest reduction in training box loss (1.0732 vs. 1.0889), suggesting improved bounding-box localization. This precision gain, however, is accompanied by a reduction in recall (0.6421 vs. 0.6781) and slightly lower mAP50 (0.7294 vs. 0.7395) and mAP50–95 (0.4517 vs. 0.4558) relative to the YOLOv8s baseline, indicating a precision-oriented trade-off rather than a uniform improvement across all metrics. As shown in Figure 3, the training and validation curves for UWAD-Net plateau by approximately epoch 100–120, indicating that the model reached convergence within the 150-epoch training budget; the observed differences from YOLOv8s therefore reflect the effect of the attention module rather than an incomplete training schedule. Notably, UWAD-Net's precision also exceeds that of every other evaluated baseline, including YOLOv8n, YOLOv10n/s, and YOLOv11n/s (Table 8), making the reduction of false-positive detections its primary and most consistent advantage. These findings are broadly consistent with previous studies demonstrating the effectiveness of attention mechanisms for underwater object detection (Chen et al., 2023; Jiang et al., 2024), as well as the recently reported inland-water attention-based detector YOLO-EcoFocus (Devi et al., 2025), while also underscoring that attention-based gains in this setting are precision-focused rather than uniformly improving all detection metrics.

6.4 Cross-domain transfer learning

The cross-domain results reveal that the learned features on the PondDebris dataset can be applied to other underwater environments, thus proving that the proposed network structure of UWAD-Net is not confined to a particular collection domain. While the performance was reduced due to the factors such as water depth, illumination, turbidity, and scene properties, the method kept reliable detection ability without making any changes in its architecture. The decreased recall implies that the problem of domain shift mostly impacts object detection and not classification, and the necessity of bigger datasets and domain adaptation needs to be addressed to increase the performance. Thus, these results allow concluding that the developed framework based on the PondDebris dataset and the proposed UWAD-Net is promising to build generalized underwater waste detection methods.

6.5 Limitation

The present study has two primary limitations. First, the PondDebris dataset was collected in a controlled aquarium environment rather than under natural freshwater pond conditions. Even though the experimental design was set up with respect to some of the key considerations for underwater imaging, like light under water, turbidity, object-specific characteristics, and substrates, it does not take into account the highly dynamic and complex situations faced in nature pond environments. Second, the PondDebris dataset contains 1,526 images, divided into 1,206 training images, 183 validation images, and 137 test images. To improve the reliability of the performance assessment, a five-fold cross-validation protocol has been incorporated, and the corresponding evaluation framework is presented in Table 10. Future work will focus on expanding the dataset through additional data collection under diverse freshwater pond environments to further improve model robustness and generalization across varying environmental conditions.

7 Conclusion

PondDebris is a freshwater underwater waste detection dataset developed in this study, comprising 1,526 images and 3,224 annotated bounding boxes categorized into biodegradable and non-biodegradable waste classes. PondDebris was collected under controlled environmental conditions. The dataset was created to address the lack of annotated benchmarks for freshwater underwater debris detection and to provide a suitable platform for evaluating object detection models in shallow pond environments. Among the six evaluated YOLO-family models (Bochkovskiy et al., 2020) YOLOv8s achieved the best baseline performance and was, therefore, selected as the backbone for the proposed framework. The proposed UWAD-Net integrates the Convolutional Block Attention Module (CBAM) into the YOLOv8s architecture to enhance feature representation and improve underwater waste detection accuracy. According to empirical results, UWAD-Net consistently outperforms the baseline network. Also, the effectiveness of the feature refinement technique based on attention is confirmed in the ablation studies. Further, the transfer learning (Pan and Yang, 2010) experiments performed using the Underwater Debris Detection Dataset (Moorton et al., 2024) suggest that the learned feature representations possess good transferability in other underwater debris detection tasks. Future work will focus on collecting and evaluating real-world in-situ datasets from freshwater ponds to assess the generalizability of the proposed framework under diverse environmental conditions. The PondDebris dataset can also be expanded with a larger number of images collected from different freshwater pond environments. Finally, fine-tuning and domain adaptation strategies could be explored to improve model generalization, particularly for class-balanced and cross-domain underwater debris detection scenarios.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Artificial intelligence (AI) tools, including OpenAI's ChatGPT, were used to support the preparation of this manuscript. These tools assisted in tasks such as language refinement, grammar correction, and content structuring. All intellectual content, data interpretation, and final decisions were made solely by the authors, who take full responsibility for the integrity and accuracy of the work presented.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frwa.2026.1923979/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Dagang Wang, Sun Yat-sen University, China

Reviewed by: Mohammad Najafzadeh, Graduate University of Advanced Technology, Iran

Gao Xiaoqian, Shandong University of Science and Technology, China