Inter-comparison of statistical downscaling and machine learning based post-processors for improving Indian summer monsoon rainfall

Abstract:

Downscaling precipitation obtained from climate models rectifies the bias and the mismatch with the observed data. Downscaled daily precipitation data is useful in predicting many climate extremes like floods and droughts. Statistical downscaling with big data and machine learning approaches is gaining a wide interest in this field. The conventional statistical downscaling (SD) methods and recent machine learning (ML) approaches are compared in this study.

The study is carried out with Indian monsoon precipitation data from the year 2007 to 2015 for the state of Madhya Pradesh. The observed dataset is obtained from IMD (Indian Meteorological Department) at 0.25°× 0.25° spatial resolution and the model output from ECMWF (European Centre for Medium-Range Weather Forecasts) at 0.5°× 0.5°. Two SD approach (Bias-correction and spatial disaggregation (BCSD) and Statistical downscaling and bias correction (SDBC)) and two ML approach (Convolutional Neural Network (CNN) and Recursive Neural Network (RNN)) are applied. The ML methods are applied taking covariates like temperature, sea level pressure, and specific humidity. These methods are applied with data from 2007 to 2014 and the accuracy is tested on 2015 data. These approaches are evaluated based on various error metrics and comparison with the original observed data.

The analysis is currently in progress. Preliminary results show that the error metrics obtained from the SDBC method is better than the error metrics of the BCSD method. These approaches underestimate extreme precipitation events in many places. The inter-comparison of ML methods with SD methods will be presented at the conference.