MACHINE LEARNING ON GCM ATMOSPHERIC VARIABLES FOR SPATIAL DOWNSCALING OF PRECIPITATION

Abstract

 This study presents a downscaling method that utilizes a machine learning algorithm to improve the regional analysis capabilities of general circulation models (GCMs) in hydrological prediction. The U-Net algorithm, a representative approach based on Convolutional Neural Networks (CNNs), is employed to emulate the dynamic downscaling process of regional circulation models (RCMs). In this process, the GCM output is utilized as the input, while the output from the regional climate model (RCM) acts as the label data. The accuracy of the model is evaluated by considering various factors, including the input datasets, model structures, and output formats. The input datasets consist of essential atmospheric variables such as precipitation, temperature, and humidity, which are examined in both two-dimensional and three-dimensional formats. The model structure is assessed by testing different hyperparameters and analyzing their impact on the accuracy of the predictions. Furthermore, the effect of the label data is also investigated. The results of the experiments demonstrate that the U-Net algorithm successfully emulates the dynamic downscaling process, even in the absence of precipitation input data. Increasing the number of convolutions in the model structure contributes to improved accuracy; however, enlarging the receptive filter size does not necessarily yield better results. Moreover, incorporating temperature data along with precipitation in the labeled dataset enhances the accuracy of the predictions compared to using precipitation alone.