Prediction of fingerling biomass with deep learning

Abstract

Organic nitrogenous nutrients present in proteins like fish are important and indispensable for cell formation, growth, and reproduction of living beings. The search for quality protein is a challenge for aquaculture especially for the breeding, fattening, and slaughtering of fish ensuring environmental sustainability with strict management of resources and waste reduction. The main contribution of this research is in the development of a system for predicting the body biomass of live fingerlings and juveniles as an alternative to the traditional invasive estimation methods used by small and medium-sized fish farms. This paper provides an evaluation based on automatic frame selection with data augmentation by 10° rotation for biomass estimation of Pintado Real ®using video frames collected by a computer vision system that resulted in a new dataset called ALEV400P with 400 fingerlings ranging from approximately 1 to 32 grams. Three different deep learning frameworks were evaluated and also compared with five shallow learning algorithms. The results showed that the Deep Belief Network outperformed the others with R2 of 0.7 for batch 1. The combination of the automatic frame selection strategy by Euclidean distance from head to tail with data augmentation by 10-degree rotation optimized the prediction result using a dataset with high variability of fingerling sizes.