Predictability assessment of northeast monsoon rainfall in India using sea surface temperature anomaly through statistical and machine learning techniques

Abstract

The socio‐economic growth of India is adversely affected by the abnormal meteorological phenomena of floods and droughts. Thus, rainfall prediction is highly desirable for livelihood and sustainability. In this study, northeast monsoon rainfall (NEMR) is predicted over the Indian peninsular region for the months of October, November, and December using a global sea surface temperature (SST) anomaly as a predictor by linear regression (LR), artificial neural network (ANN), and extreme learning machine (ELM) techniques. The predictions are made by LR, ANN, and ELM models for the period 1990–2016 using the training input (1871–1989) of different time series samples of global SST anomaly data collected from Hadley Centre SST data set (HadSST3). Principal component analysis was used for dimensionality reduction of the data sets, and its application substantially improved the predictive ability of the machine learning techniques. The performance of the PC‐ELM technique with the ensemble method (ESM 8–9) training window is found to be more accurate than the LR and ANN techniques and provides minimal error scores as per the statistical analysis. This study concludes that the global SST anomaly has the potential to be used as a predictor for northeast monsoon rainfall and useful for long‐range climatic projections.

Statistics for climate informatics

“…Samarasinghe, McGraw, Barnes, and Ebert‐Uphoff () provide deeper insights into the climate system, such as Arctic temperatures and midlatitude jet streams, by exploring causal relationships between different phenomena (using the methods of Granger and Pearl causality). Dash, Mishra, and Panigrahi () apply regression and artificial neural networks for improving predictions of northeast monsoon rainfall over the Indian peninsular. Li and Ding () propose a new form of penalized regression, which is an alternative to the existing methods of LASSO, ridge regression, and elastic net, and demonstrate its applicability in statistical downscaling of climate data.…”