A predictive model of recreational water quality based on adaptive synthetic sampling algorithms and machine learning

Predicting recreational water quality is one of the most difficult tasks in water management with major implications for humans and society. Many data-driven models have been used to predict water quality indicators to allow a real time assessment of public health risk. This assessment is most commonly based on Faecal Indicator Bacteria (FIB), with the value of FIB compared with thresholds published in guidelines. However, FIB values usually tend to be unbalanced within water quality datasets, with small proportions of data exceeding guideline thresholds and far larger numbers that do not. This can be a limiting factor in the uptake of model predictions since, even if the overall accuracy is high, the sensitivity of the predictions can be low. To address this issue, this paper proposes an adaptive synthetic sampling algorithm (ADASYN) to generate synthetic above-threshold FIB instances and test the validity of the approach for the prediction of recreational water quality. The models in this paper are based on four machine learning techniques: k-mean nearest neighbour, boosting decision tree, support vector machine, and multi-layer perceptron artificial neural network and are applied to five different locations in Auckland, New Zealand. Aside from support vector machine, all models provide favourable predictions with relatively high sensitivity (around 75%) and overall accuracy (over 90%), indicating that both the compliant and exceedance conditions can be effectively predicted through the use of more sophisticated model training which involves artificial data. Considering the model accuracy and stability, boosting decision trees (BDT) and multi-layer perceptron artificial neural (MLP-ANN) network are the best two models and the multi-layer perceptron is the most efficient with the shortest computation time.