Employing deep learning for chum salmon (Oncorhynchus keta) age estimation through scale image analysis

Accurate age estimates of salmon produced from scale growth patterns yield demographic information essential for population assessment and fisheries management. Conventionally, salmon ages are estimated from scales by trained personnel where demand often exceeds staffing capacity and financial resources. Convolutional neural network (CNN) models are a promising approach to augment manual age estimation. We compiled an extensive scale image repository from chum salmon (Oncorhynchus keta, Salmonidae) stocks throughout the North Pacific rim, evaluated two image preprocessing algorithms—Segment Anything Model (SAM) and You Only Look Once (YOLO)—to remove or reduce background noise, and trained an ensemble of three CNN models (EfficientNetV2L, NASNetLarge, and ConvNeXtLarge, weighted by their MSE) to infer chum salmon age. Agreement with expert estimates was higher for the SAM preprocessed images (93.2%) than YOLO (91.4%) across all major age classes (ages 3–5). Both preprocessing approaches showed reduced agreement for less common age classes because of inadequate representation in the model training and validation datasets. This approach may be integrated into fisheries management workflows to accelerate the age estimation process, complement manual reading, and bolster quality control protocols.