A Deep Learning Framework for Cetacean Detection: Integrating YOLOv11 With ADRIF Post‐Processing

Abstract

Cetaceans are widely regarded as sentinels of marine ecosystem health, yet their monitoring in the Indian Ocean remains challenging because conventional survey techniques often lack adequate spatial and temporal resolution. In this study, we developed and evaluated a deep learning‐based detection framework for cetaceans in Indian waters using high‐resolution imagery collected during research cruises conducted between 2024 and 2025. The framework employs the YOLOv11 object detection architecture, trained and validated on field‐based images of three representative species: Balaenoptera musculus (blue whale), Globicephala macrorhynchus (short‐finned pilot whale), and Stenella longirostris (spinner dolphin). To improve detection reliability, the pipeline integrates ensemble‐ and rule‐based post‐processing strategies, including weighted box fusion (WBF) and a novel adaptive distance‐aware refinement and intelligent fusion (ADRIF) algorithm designed to reduce duplicate detections and false positives in complex marine science applications. The optimized framework achieved a mean average precision (mAP@0.5) of 81.6% and an F 1‐score of 80.05% on the test dataset, demonstrating robust detection performance under variable oceanographic conditions. These results highlight the potential of deep learning–based approaches to enhance the efficiency and consistency of visual cetacean detection from vessel‐based imagery. The proposed framework provides a reproducible methodological foundation for integrating computer vision tools into marine mammal research and future large‐scale monitoring efforts.