Machine learning approach to investigate the influence of water quality on aquatic livestock in freshwater ponds

Water quality (WQ) is a key factor that affects harvest outcome from freshwater ponds. Irregular or aperiodic variations of different WQ variables can influence the growth, survival, and yield of aquatic livestock in the ponds. In this research, WQ and harvest data collected from an Australian prawn farm over a whole grow-out season is used to investigate how the variations of WQ influence the harvest outcome of prawns from the ponds. We present a set of approaches based on machine learning to: (i) understand the effect of five WQ variables in differentiating high and low performing ponds (in terms for harvest performance); and (ii) identify how the variations in these WQ variables over the grow-out season contributed to final harvest outcome (growth and yield). To develop the ponds classification approach, we apply eight different machine learning classifiers: neural networks, support vector machine, k-nearest neighbours, logistic regression, gaussian naive bayes, decision tree, random forest, and AdaBoost. To identify the driving factors (in terms of variations of WQ) that affect growth and yield of aquatic livestock in ponds, we apply three feature selection methods: mutual information, correlation-based feature selection, and ReliefF. Results demonstrate that dissolved oxygen, salinity, and temperature are the three WQ variables that have the greatest influence on overall harvest performance of the ponds. Changes in dissolved oxygen and salinity in the last quarter of the grow-out season, and variations of temperature immediately after stocking contributed the most to differentiate the high and low performing ponds.