A benchmark dataset and ensemble YOLO method for enhanced underwater fish detection

Abstract Fish monitoring is crucial for the aquaculture industry as it addresses economic losses and productivity challenges. Existing manual fish detecting methods are difficult and time consuming, limiting their application in real‐time monitoring. To alleviate this challenge, there is growing interest in intelligent systems for vision‐based fish health analysis. In this study, we introduce DePondFi, a novel dataset designed for various computer vision tasks in underwater pond environments. The dataset comprises underwater images from different habitats in South Indian ponds, with corresponding bounding box annotations. Our experiments analyze the characteristics of the dataset and evaluate the performance of single‐stage object detection models. YOLOv8 demonstrates the highest effectiveness, achieving a mean average precision (mAP@50) of 0.92. By aggregating predictions and applying non‐maximum suppression, our ensemble method improves detection performance and robustness compared to individual models, achieving an mAP@50 of 0.94. DePondFi serves as a valuable benchmark in the underwater computer vision domain for real‐time fish monitoring.