Optimisation of Machine Learning ARIMA Forecasts for Modelling Future Variations in Earth’s Surface Phenomena Using Sparse Time Series Satellite Data – A Case Study of Sea Surface Salinity in a Coast

Data-poor areas usually have difficulties in meeting the multiple predictor variables requirement of building appropriate multivariate machine learning (ML) regression models with contemporary radar (all-weather) satellites observations of Earth’s surface phenomena (ESP). In this regard, univariate ML autoregressive integrated moving average (ARIMA) models built with such satellite observations have a relative advantage. However, our knowledge of the accuracy of traditional variants of ESP forecasts (Forecast, Lo 95, and Hi 95) by auto-determined best ARIMA models; and the forecasts accuracy optimisation when such models are built with relatively sparse time series satellite data is still limited. Experimental approach was leveraged to exploit three sets (36, 48, and 60 monthly epochs) of relatively sparse sea surface salinity (SSS) datasets from the Soil Moisture Active Passive Mission (SMAP) satellite products (Jan. 2016-Dec. 2021) as a case study. The optimised SSS forecasts (Hybrid of the traditional variants) shows RMSE of 0.3133 to 0.7813 psu; and the validation shows MAPE of 0.6773 to 1.3643% respectively. This implies that the innovative “Hybrid” variant consistently optimised the SSS forecasts accuracy (MAPE) of each of the traditional variants across the three different levels of the temporal data sparsity by about 14.97-81.54%. This suggests that relatively accurate traditional variants of ESP forecasts by such ARIMA models can be reasonably and significantly optimised to better model future variations in any ESP at any geographical location.