Evaluating Machine Learning Techniques for Predicting Salinity

Matthew Harper, Ivy Liu, Bing Xue, Ross Vennell, Mengjie Zhang

Abstract: Oyster farms provide a sustainable and profitable export for New Zealand. Oyster farms are sensitive to changes in salinity that can cause significant crop loss if they persist too long. Recent extreme weather events have been leading to increased periods of low salinity, putting the farms at risk. Machine learning based methods provide a way to predict these low salinity events and provide an early warning system, but this has not been investigated in aquaculture. In this paper, we investigate three different methods to assess the viability of salinity prediction systems. A simple statistical model, a genetic programming (GP) based symbolic regression model and a convolutional neural network (CNN) were compared as ways of solving this problem. The results show that GP based symbolic regression and CNNs are fairly good approaches to predicting salinity. However, as weather events get more extreme, the CNN approach tends to hold up better and can be generalised better, while the G P based symbolic regression models show better potential explainability with the tree based model structure. These results show promise and provide a good stepping off point at creating a generalised approach to predicting salinity in estuaries.