A Robust Machine Learning Framework for Assessment of Hydroclimatic Teleconnections between Monthly ISMR and 18 Circulation Indices and Prediction of Monthly ISMR

Abstract

Rainfall received across the India during monsoon season is very crucial for the country’s economy. Predicting monthly rainfall plays a crucial role in efficient water resources management and in reducing the risk of hydrological disasters. Looking to the drawback of general circulation models (GCMs), it is better to assess hydroclimatic teleconnections (HCTs) of monthly Indian summer monsoon rainfall (ISMR) for its prediction. The climate system is dynamic and nonlinear. These nonlinearities are not well represented by conventional linear statistical techniques. Machine learning (ML) provides in-depth understanding of complex nonlinear data structures. To the best of the authors’ knowledge, no earlier study has performed an assessment of HCTs and prediction of monthly ISMR by formulating robust ML models by employing eighteen circulation indices, which is performed in the present study. Assessment of input significance and input independence, systematic data division in such a way that statistical parameters of subsets and all data are similar, followed by use of Bayesian optimization for development of ML model, are necessary for formulating robust ML models and it is not performed by any earlier study according to the authors’ best knowledge. Therefore, it is performed in the present study. Thus, the present study employed four ML techniques, namely, adaptive boosting, extreme gradient boosting regression, random forest (RF) regression, and support vector machine for assessment of aforesaid HCTs. The aforementioned analysis is performed for two periods viz. 1951–2000 and 1951–2014 to investigate the alteration of aforesaid HCTs along with time. It is observed that HCTs of monthly ISMR vary with time. The present study also showed that all the formulated ML models are robust and do not show any overfitting/underfitting. It is also observed that the performance of the RF technique is better as compared to the other three ML techniques.