Abstract
Identification of indigenous fish species is crucial for sustainable fisheries management and biodiversity conservation. The traditional expert systems of fish identification are inconsistent and time-consuming. In an attempt to address this weakness, we leveraged cutting-edge deep learning techniques, namely various iterations of the You Only Look Once (YOLO) algorithm, i.e., YOLOv5, YOLOv7, YOLOv8, and YOLOv11. Our study introduces a new data set of 13,000 images of 16 indigenous species of fish acquired from the Bahir Dar Fishery and Other Aquatic Life Research Center in Ethiopia. We used a whole preprocessing pipeline that included image enhancement techniques, CSPDarkNet, Histogram of Oriented Gradients (HOG), and segmentation-based image feature extraction, along with Roboflow annotations. The new approach significantly improved the dataset quality and usability to train models. Besides, we experimented using various hyperparameters, such as activation functions (Softmax, ReLU), optimizers (AdamW, SGD), batch sizes (32, 16), learning rates (0.01, 0.001), and epochs (100, 50). Our experiment confirms that YOLOv11n achieved the best mean Average Precision (mAP) of 94.7% when trained using 100 epochs with the AdamW optimizer. This study confirms that the combination of a well-designed dataset, sophisticated preprocessing techniques, and effective deep learning architectures provides a robust and practical paradigm for indigenous fish species identification, thereby enhancing aquatic biodiversity conservation.