ZMFISC: Zhu‐Ming data set with a convolutional neural network for identifying Indo‐Pacific humpback dolphins (Sousa chinensis)

Abstract

AbstractThe Indo‐Pacific humpback dolphin (Sousa chinensis) is a small‐toothed whale species that inhabits estuaries and shallow coastal waters from the eastern Indian Ocean to the western Pacific, and faces significant negative impacts from anthropogenic activities. The noninvasive Photo‐identification method enables individual identification and abundance estimation based on natural markings of cetaceans without disrupting their natural behaviors. Currently, the identification of S. chinensis using photographs relies primarily on time‐intensive visual recognition by experienced researchers. Through field surveys conducted in the west Huangmao Sea area from 2012 to 2021, we compiled the Zhu‐Ming data set focusing on S. chinensis (ZMSC), consisting of 479 individuals and 5,196 photos. Utilizing the ZMSC, we proposed a Few‐Shot Identification method for S. chinensis (FISC), which achieved 85.93% identification Top‐1 accuracy. The implementation of proper preprocessing steps and data augmentation techniques has significantly enhanced the performance of FISC, while visualizing network weights has improved its interpretability. Despite the remaining challenges of data imbalance and the inability to automatically allocate new labels, ZMFISC alleviates the challenge of the current heavy reliance on time‐intensive visual recognition methods by researchers for individual identification of S. chinensis and provide a valuable tool to enhance future conservation efforts for S. chinensis.