Univariate deep learning framework for short-term SST forecasting at high spatio-temporal scales

Abstract We present a univariate hybrid machine learning framework to predict daily high resolution Sea Surface Temperature (SST) near the Gulf of Kutch region at a resolution of ∼5.5 km. The hybrid model integrates Intrinsic Mode Functions (IMF) derived from variational mode decomposition with a Long Short-Term Memory (LSTM) network to augment predictive skill. The predicted SST demonstrates impressive performance up to lead times of 7 days. Using statistical metrics like Kullback-Leibler divergence and mutual information, we show that the SST predicted from the hybrid model with a lead time of 3 days decisively outperforms the high-resolution GLORYS SST reanalysis let alone forecast skills of a data-assimilated dynamical model. Using conditional probability, we show that the SST forecasts from the hybrid model are quite reliable over the entire range of SST observations in the study domain. In contrast, the reliability of GLORYS falters in the lower range of SST observations. Also, the hybrid model excels in capturing fine-scale SST features, such as SST fronts, and detecting Marine Heatwaves (MHWs) up to 3 days in advance. These capabilities hold significant applications for Potential Fishing Zone identification and coral bleaching alerts. The hybrid model framework is also adept at forecasting location specific high frequency (3 hourly) SST with a lead time of a day.