Using unmanned aerial vehicle (UAV) with machine vision systems (MVS) to assess fish weight of red tilapia in river cage culture

Abstract

Abstract Efficiently estimating fish weight poses a major challenge for effective fish feeding and harvesting.This study introduced a novel approach using an Unmanned Aerial Vehicle (UAV) and a Machine Vision System (MVS) to non-intrusively estimate the weight of red tilapia fish within a cultured environment (river-based cage culture). Our proposed method applied image acquisition via the UAV and subsequent image analysis using the MVS. Initially, Decision Tree (DT), Random Forest (RF), Naïve Bayes (NB), K-Nearest Neighbors (KNN), and Artificial Neural Network (ANN) models were trained for image classification across 9 fish classes in 5 cages. Subsequently, these models were tested on another set of 9 fish classes in 3 different cages. The results demonstrated that ANN achieved the highest prediction efficiency during training and validation, having mean (± standard deviation) scores for accuracy, precision, recall, and F1 of 90.39±8.33, 90.13±10.61, 90.05±9.14, and 90.13±9.05 %, respectively. However, during testing, the accuracy was 42 %. Therefore, future implementations are recommended to expand the dataset and to optimize ANN parameters (including K-fold, training cycle, hidden layers, and neurons) to enhance the efficiency of the ANN model. Where a substantial accuracy improvement cannot be achieved, we suggest considering the utilization of Convolutional Neural Network models and image data generator with transfer learning techniques.