Climate Projections over Punjab, Pakistan through Machine Learning Approaches using CMIP6 Data.

Abstract

The downscaling techniques have been regarded for a long time as an essential component of assessing climate change for impact assessment studies. General circulation models (GCMs), known as the climate models, are frequently used to examine the influence of climate change on a coarse level. As a result, the data produced by GCMs is not primarily useful for making predictions, therefore downscaling is employed to produce high-resolution results. In this study, the downscaling models for the future projections of daily minimum temperature, maximum temperature, and precipitation were constructed for ten stations of Punjab. The three downscaling methods, namely, the Multivariate Multiple Linear Regression model (MMLR), Support Vector Machine (SVM), and Artificial Neural Networks (ANN) were utilized for the construction of models. The best-identified model was employed for future projection of the predictands (minimum, maximum temperature, and precipitation) using the Coupled Model Intercomparison Project Phase 6 (CMIP6) GCMs, under two future scenarios Shared Socioeconomic Pathways 4.5 (SSP 4.5) and SSP8.5 for the time period 2015-2045. The performance of each model was assessed graphically and numerically. For numerical evaluation, four statistical performance indices were used including the Root Mean Square Error (RMSE), Nash-Sutcliffe Efficiency (NSE), Mean Bias Error (MBE), and Difference (D) in the variations between the observed and model predicted data. Meanwhile, for graphical evaluation, monthly box plot and the Taylor diagram were used. The evaluation results suggest that ANN outperforms the other two methods. Therefore, the best-identified ANN model is then used to project the future minimum temperature, maximum temperature,