Predicting shellfish farm closures using time series classification for aquaculture decision support

•We develop a novel prediction model for shellfish farm closure in aquaculture.•We exploit machine learning using time series classification for prediction model.•We also address data imbalance issue in time series classification for prediction.•The model is used in aquaculture decision support in Tasmania, Australia.

Closing a shellfish farm due to pollutants usually after high rainfall and hence high river flow is an important activity for health authorities and aquaculture industries. Towards this problem, a novel application of time series classification to predict shellfish farm closure for aquaculture decision support is investigated in this research. We exploit feature extraction methods to identify characteristics of both univariate and multivariate time series to predict closing or re-opening of shellfish farms. For univariate time series of rainfall, auto-correlation function and piecewise aggregate approximation feature extraction methods are used. In multivariate time series of rainfall and river flow, we consider features derived using cross-correlation and principal component analysis functions. Experimental studies show that time series without any feature extraction methods give poor accuracy of predicting closure. Feature extraction from rainfall time series using piecewise aggregate approximation and auto-correlation functions improve up to 30% accuracy of prediction over no feature extraction when a support vector machine based classifier is applied. Features extracted from rainfall and river flow using cross-correlation and principal component analysis functions also improve accuracy up to 25% over no feature extraction when a support vector machine technique is used. Overall, statistical features using auto-correlation and cross-correlation functions achieve promising results for univariate and multivariate time series respectively using a support vector machine classifier.