Object tracking-based deep learning for calculating the quantity of eel seedlings (Anguilla bicolor)

Abstract

Anguilla bicolor is a highly economically valuable species of eel, whose cultivation needs to be carried out to the maximum. One of the factors supporting successful aquaculture is the optimal stocking density. The current calculation method for determining seedling quantity for stocking density is ineffective for large-scale cultivation. Therefore, in this study, an object-tracking-based calculation method using deep learning was applied to address this issue. Deep learning using YOLO v8 is an advanced method for counting and detecting objects including fish. This algorithm was used with the addition of the ByteTrack algorithm for object tracking and supervision for object counting. Object tracking is applied such that during seed counting, moving objects can be tracked and not counted repeatedly. Training was conducted on two datasets with different labels, namely the head and tail parts, using 300 epochs, 16 batches, and learning rate of 0,001 and 0,01. Based on the training results, the model with the head label has higher precision and recall than the tail-labelled model, with values of 0,91 and 0,93 respectively. The model performed well at a stocking density of 63 individuals/m², with accuracy and F1 scores of 90,91% and 0,95 respectively.