Revamping extended range forecast of Indian summer monsoon

Abstract

Forecast of the Indian summer monsoon on an extended range (beyond the conventional one-week lead time) is critical for an agronomic economy like India. Although dynamic models have been quite successful in capturing and representing monsoon circulation, they fail to sustain this skill beyond the standard weather time scale (7–10 days). As such, the present study is directed at developing a computationally feasible, yet reliable method of statistical downscaling that further improves the present skill of global dynamic models for extended range forecasting. We quantitatively demonstrate the feasibility of this post-processing statistical module for improving the predictability of the dynamic Extended Range Prediction System (ERPAS), which is developed by Indian Institute of Tropical Meteorology, Pune, India and now operational in the country. It first develops climate cluster(s) (rainfall states in the present case), then builds a statistical relationship between these clusters and a set of appropriate climate variables using a robust and advanced classification technique known as Extreme Gradient Boosting (XGboost), and eventually delivers the real-valued precipitation at individual grid cells via a non-parametric regression. The module is able to skilfully capture the active and break phases of the Indian summer monsoon, and also subsequently project them for the ensuing regression component of the module. This approach shows to significantly boost the prediction skill over the Core Monsoon Zone of Indian mainland up to a lead time of 4 weeks. Our statistical downscaled model is comparable in the week 1 lead time but outperforms global models for 2nd, 3rd, and 4th weeks lead time. Nonetheless, this performance is retained only in the northern and central region, and not ubiquitous over the rest of India.