Spatio temporal hydrological extreme forecasting framework using LSTM deep learning model

Abstract

Hydrological extremes occupy a large spatial extent, with a temporal sequence, both of which can be influenced by a range of climatological and geographical phenomena. Understanding the key information in the spatial and temporal domain is essential to make accurate forecasts. The capabilities of deep learning methods can be applied in such instances due to their enhanced ability in learning complex relationships. Given its success in other domains, this study presents a framework that features a long short-term memory deep learning model for spatio temporal hydrological extreme forecasting in the South Pacific region. The data consists of satellite rainfall estimates and sea surface temperature (SST) anomalies. We use the satellite rainfall estimate to calculate the effective drought index (EDI), an indicator of hydrological extreme events. The framework is developed to forecast monthly EDI using three different approaches: (i) univariate (ii) multivariate with neighbouring spatial points (iii) multivariate with neighbouring spatial points and the eigenvector values of SST. Additionally, better identification of extreme wet events is noted with the inclusion of the eigenvector values of SST. By establishing the framework for the multivariate approach in two forms, it is evident that the model accuracy is contingent on understanding the dominant feature which influences precipitation regimes in the Pacific. The framework can be used to better understand linear and non-linear relationships within multi-dimensional data in other study regions, and provide long-term climate outlooks.