An approach for good modelling and forecasting of sea surface salinity in a coastal zone using machine learning LASSO regression models built with sparse satellite time-series datasets

Abstract

Abstract The risks of upstream seawater intrusion from coastal zones, particularly to the environment and people’s health,are gradually becoming serious issues thatrequire proactive environmental monitoring and good modellingapproaches. However, the temporal resolutions of relevant contemporary all-weather satellites that detect SSS are unable to support real-time applicationsthat can provide the required early warning information for mitigating such risks. Our current practical knowledge of the efficiency of machine learning (ML) least absolute shrinkage and selection operator (LASSO) regression modelsbuilt with relatively sparse all-weather satellite data for achieving relatively accurate predictor variable selection,collinearity detection,and high SSS prediction accuracy is still limited. In this paper, we utilized relatively sparse time series all-weather satellite datasets consisting of 6 potential predictor variables (PPVs), wind speed (WS), high wind speed (HWS), sea surface temperature (SST), absolute dynamic topography (ADT), sea level anomalies (SLAs) and precipitation (PRECIP) (January 2016-December 2020) to construct an ML LASSO model (using the forecastMLlibrary in R/R-studio) to predict SSS ona tropical coast (Nigerian coastal zone). We utilized the same datasets for building the L0-regularized regression (L0) model (using the L0Learnlibrary) to determine the relative importance of the PPVs for the ML time series forecasting of the SSS and to detect collinearity. The output was used to determinethe abilityof the LASSO model to determinethe relative importance of the PPVs for forecasting SSS and detecting collinearity. We determinedthe best combination of lookback (LB) and h-step-ahead (H) parametervalues for building a relatively accurate ML LASSO model with the datasets. We determinedand validatedthe relative importance of the PPVs for forecasting the monthly SSS using the LASSO model with the best combination of parametervalues. We predict and validate the monthly SSS values for January-December 2021 with a relatively accurate model. We show that the LB:24 and H:12 parametervalues,with an RMSE of 0.54437, are the best for building a relatively accurate LASSO model with such datasets. We show that the WS, HWS, and SLA are the most important PPVs for achieving relatively accurate SSS forecasts with the model. However, we show the limitations of such a LASSO model in achieving relatively accurate predictor variable selection and collinearity detection. We show practical solutions to such limitations by utilizing the L0 model to assist the LASSO model in achieving relatively high SSS prediction accuracy. Finally, we predict the monthly SSS values using the relatively accurate LASSO model and validate them with the observed SSS (January-December 2021) and obtain an RMSE of 0.7428 and a MAPE of 1.9039%. AMAPE value approximately5 times less than 10% implies a high SSS prediction accuracy that can be replicated to provide useful early warning information for mitigating such risks in any coastal zone. The results imply that the good practice for using such satellite datasets to build a relatively accurate ML LASSO model for forecasting should begin with rigorous supervised-automatic deletion of observation records with null values and outliers,followed by unbiased selection of appropriate parametervalues and important predictor variables and collinearity assessment.