Using Artificial Neural Networks to Forecast Monthly and Seasonal Sea Surface Temperature Anomalies in the Western Indian Ocean

Abstract

A study implementing Nonlinear Autoregressive with Exogenous Input (NARX) neural network has been undertaken to predict monthly and seasonal SST anomalies in the western Indian Ocean. The study involves a coastal site located along the eastern African seashore, and an oceanic site that lies precisely within the western pole of the Indian Ocean Dipole. Performance of the network is measured by a series of statistical indicators during testing phase (1981–2010), and results are compared with outputs from three other neural networks and a linear system, the Autoregressive Integrated Moving Average with Exogenous Input (ARIMAX) model. The NARX network has provided the best overall performance, but the other four models have also given sufficiently good predictions. The monthly predictions are on average within an error of ±0.09°C for the first 50% and 90% within ±0.22°C. The corresponding errors for the seasonal predictions are ±0.04°C and ±0.09°C, respectively. The RMSE between observations and predictions is about 0.13°C and 0.06°C for the monthly and seasonal SST anomalies, while the average correlation coefficient is about 0.88 and 0.98, respectively.