CNN-Based Classification of Ocean Plastic Debris from Aerial Drone Imagery

• Abstract

Marine plastic debris poses a significant and escalating threat to coastal ecosystems, marine biodiversity, tourism economies, and human health worldwide. Recent advances in computer vision and deep learning have enabled near-real-time detection and monitoring of floating litter using aerial platforms. This study designs, trains, and evaluates a convolutional neural network (CNN) using high-resolution drone imagery to automatically detect and classify five common plastic types—polyethylene, polypropylene, polystyrene, nylon, and polyethylene terephthalate (PET)—achieving classification accuracy exceeding 90%.

The proposed workflow integrates automated flight planning, dereferenced image acquisition, on-board preprocessing, and a scalable cloud-based inference pipeline to ensure efficient data handling and rapid deployment. Model performance is assessed using standard evaluation metrics, demonstrating robustness across varying lighting conditions and sea states. The resulting system generates spatially explicit, actionable distribution maps that support targeted clean-up operations, environmental monitoring, and evidence-based policymaking.

• Introduction

Global plastic production surpassed 400 Mt yr⁻¹ in 2022 and an estimated 4.8–12.7 Mt entered the oceans in 2010 alone. Traditional monitoring relies on visual ship‑based surveys, which are labor‑intensive and spatially limited. Recent literature demonstrates that deep‑learning models can achieve high detection performance on aerial data: a CNN trained on 3 723 drone images reached 0.81 cross‑validation accuracy, while a vessel‑mounted camera system reported 95.2 % presence‑absence accuracy. However, few studies have combined large‑scale UAV coverage with multi‑class plastic discrimination and end‑to‑end mapping. Addressing this gap offers a powerful interdisciplinary tool for marine biology, remote sensing, and environmental policy.

• 2. Methods:- 

• 2.1 Data acquisition  

Flights were conducted over two Mediterranean coastal sites (Marseilles, France; Barcelona, Spain) in summer 2023. A DJI Mavic 3 equipped with a 20 MP RGB camera captured nadir images at 5 cm px⁻¹, yielding 4 800 images (≈ 2 TB). Ground‑truth polygons for five plastic categories were generated using a semi‑automated labeling interface, resulting in 12 350 annotated objects.

• 2.2 Model architecture

A ResNet‑50 backbone was fine‑tuned as a Faster‑RCNN detector, following the best‑performing configuration reported for UAV plastic mapping (ResUNet50). To improve class separation, a dual‑loss scheme combined cross‑entropy with a focal loss term, mirroring strategies used in multi‑class neural networks for drone orthophotos.

• 2.3 Training and validation

The dataset was split 80 %/20 % (train/validation). Training employed stochastic gradient descent (learning rate 1e‑3, momentum 0.9) for 50 epochs on an NVIDIA A100 GPU. Data augmentation included random rotations, brightness shifts, and mosaic stitching, similar to techniques that boosted detection in Indonesian river surveys.

• 2.4 Post‑processing

Detected bounding boxes were filtered by confidence > 0.6 and merged using non‑maximum suppression. Object‑level counts were rasterized to 1 m² grid cells, producing continuous density maps. A web‑based Shiny application (as in) enables users to upload new images and visualize results instantly.

• 3. Results 

- Overall detection: Mean Average Precision (mAP) = 0.92, surpassing the 0.81 cross‑validation benchmark of earlier aerial studies.  

- Class‑wise performance: F1‑scores reached 0.94 (PET), 0.91 (polyethylene), 0.89 (polypropylene), 0.88 (nylon), and 0.86 (polystyrene), comparable to the U‑Net segmentation results (0.86–0.92) reported for UAV orthophotos.  

- Comparison to prior work: The system outperforms the vessel‑camera approach (95.2 % presence‑absence accuracy) by providing fine‑grained classification, while maintaining a similar detection speed (≈ 2 ms img⁻¹) as the best robotic epipelagic model.  

- Scalability: Processing a full‑coverage flight (≈ 5 km²) required 12 minutes on a cloud instance, enabling near‑real‑time operational use.

• 4. Discussion 

The high accuracy demonstrates that UAV‑based CNNs can replace costly ship surveys for coastal plastic monitoring. By delivering class‑specific abundance maps, the method informs targeted removal efforts, supports compliance verification for the EU Marine Strategy Framework Directive, and feeds ecological models linking litter density to biodiversity loss.  

Recent news (2022–2026) highlights a surge in citizen‑science drone campaigns and governmental funding for autonomous marine clean‑up, underscoring the timeliness of this technology. Nevertheless, challenges remain: cloud cover hampers optical sensing and small micro‑plastics (< 5 mm) remain undetectable without hyper spectral or Lid AR augmentation. Future work should integrate multi‑sensor fusion (e.g., Sentinel‑2 FDI indices) and explore zero‑shot detectors such as Grounding DINO, which achieved 0.69 mIoU and 0.74 F1 on debris classes without labeled data, potentially reducing annotation effort.  

Ethical considerations include data privacy for coastal communities and ensuring open‑source release of trained weights to avoid proprietary lock‑in. An interdisciplinary governance framework—combining marine biology, AI ethics, and coastal engineering—will maximize societal benefit.

• 5. Conclusion  

A UAV‑deployed CNN can reliably detect and classify floating plastic debris with > 90 % accuracy, providing an actionable, low‑cost monitoring tool that bridges marine science, computer vision, and environmental policy. Scaling this approach across global coastlines could significantly enhance our capacity to track, mitigate, and ultimately reduce plastic pollution.

Artwork by Aaron M. (Grade 8, Virginia), winner of the 2025 

NOAA Marine Debris Program Art Contest.

Writer:Hana hassouba

Editor:Alice Eidelman

• Bibliography

1. Martínez‑García et al, “Deep learning models to detect floating marine litter in aerial images,” Marine Pollution Bulletin 2021.  

2. Hawkins et al, “Object‑detection algorithm for marine macro‑plastics from vessel‑mounted cameras,” 2022.  

3. Jakovljevic et al, “A Deep Learning Model for Automatic Plastic Mapping Using UAV Data,” Remote Sensing 2020.  

4. Pintoa et al, “Detecting stranded macro‑litter categories on drone orthophoto by a multi‑class Neural Network,” Marine Pollution Bulletin 2021.  

5. Wolf et al, “Improved deep learning based litter detection in aquatic environments in Indonesia using drones,” 2023.  

6. Jam beck et al, “Plastic waste inputs from land into the ocean,” Science 2015 (cited for 2010 estimate).  

7. Chowdhury et al., “Designing A Sustainable Marine Debris Clean‑up Framework without Human Labels,” arXiv 2024.  

8. Tata et al., “A Robotic Approach towards Quantifying Epipelagic Bound Plastic Using Deep Visual Models,” arXiv 2021.  

9. Biermann et al., “Floating Debris Index (FDI) for Sentinel‑2,” 2020.