INTERCOMPARISON OF MACHINE LEARNING AND STATISTICAL DOWNSCALING FOR CLIMATIC PARAMETERS: A CASE STUDY OF BHUNTAR, HIMACHAL PRADESH

This paper describes various methods of statistical downscaling used to downscale the temperatures of the city of Bhuntar in Himachal Pradesh. We have used NCEP data for this study. Statistical downscaling has been done in three main ways. First, three machine learning methods have been used. Simple linear regression just uses temperature as a predictor and multiple linear regression uses the NCEP parameters fas, vas, 500as, 5thas, and humas. Non-linear regression was performed using the same variables. Among the machine learning methods, non-linear regression had the best correlation with an R2 of 0.857. Second, a deep learning method was developed for climate downscaling, and it has shown significant performance based on an R2 value of 0.84. Third, SDSM (Statistical Downscaling Model) was used, a tool that is based on multiple linear regression for prediction. It gave an R2 value of 0.9895 in the A2 scenario and 0.9901 in the B2 scenario. This tool gives the highest value of correlation among all the methods carried out in this study. SDSM has also predicted future values of Bhuntar based on given GCM values. Hence, SDSM has proven to be the most efficient.