Deep learning-based detection of submerged debris and plastics in underwater environments: a systematic review

Purpose To synthesize and critically appraise deep learning approaches for detecting submerged plastics and marine debris, with a focus on architectures, image enhancement and deployment on resource-constrained autonomous platforms. Design/methodology/approach A systematic review of 50 peer-reviewed, Scopus-indexed articles (2020–2026) was conducted following PRISMA guidelines, including dual-stage screening, standardized data extraction and qualitative synthesis. Methodological quality was assessed in terms of architectural transparency, dataset reporting, evaluation metrics and comparative baselines. Findings Single-stage detectors (particularly YOLOv11 ensembles) achieve state-of-the-art accuracy [up to 93.9% mean average precision (mAP) @0.5], whereas transformer-based Lightweight Cross-Stage Aggregation Detection Transformer - Pruned version (LCSA-DETR-P) defines the efficiency frontier (105.1 Frames Per Second (FPS) with 31.7% parameter retention). Frequency-domain enhancement (e.g. frequency-domain feature fusion) yields 12–18% mAP gains under high turbidity with modest computational cost, while traditional enhancement offers smaller, context-dependent benefits. Transfer learning and semi/self-supervised strategies can reduce labeled data requirements by approximately 80% and, combined with pruning and quantization, enable real-time inference (30–60 FPS, 85–90% accuracy) on edge devices such as Jetson Nano. Persistent challenges include severe performance degradation in extremely turbid or low-light conditions, limited cross-geographic generalization, and strict onboard power budgets for autonomous underwater vehicles and remotely operated vehicles. Originality/value To the best of the authors’ knowledge, this review is the first to integrate architectural choices, enhancement strategies and optimization/deployment techniques into a unified framework for submerged plastic detection, explicitly linking model design decisions to environmental conditions and edge-computing constraints. It identifies concrete research priorities, including turbidity-resilient multimodal sensing, standardized underwater benchmarks and foundation models tailored to diverse aquatic environments.