Intercomparison of downscaling methods for daily precipitation with emphasis on wavelet-based hybrid models

Downscaling of local daily precipitation from large-scale climatic variables is required for assessing the impact of climate change on hydrology and water resources. This study proposes wavelet transform (WT)-based Feed-Forward Neural Network (FF-NN) and Nonlinear Auto Regressive with exogenous inputs Network (NARX-NN) models for downscaling daily precipitation. The models are applied to a large river basin, the Krishna River basin, in the Indian subcontinent. Several climatic variables, including geo-potential heights, wind direction, vorticity, humidity, air temperature, mean sea level pressure, meridional velocity at surface, and 500hpa and 850hpa levels, are considered based on their statistical correlations. The results are evaluated using different performance measures and the ability of the models to capture the extreme events at five selected grid points (in different locations) having varying climatic characteristics is assessed. The performance of the proposed wavelet-based models is also compared with that of four different traditional and recent downscaling methods: Multiple Linear Regression (MLR), Statistical Downscaling Model (SDSM), Genetic Programming (GP), and Artificial Neural Networks (ANNs). The results reveal that the wavelet-based neural network models (WT-FF-NN and WT-NARX-NN) are robust compared to the other methods in terms of their ability to capture the regional precipitation patterns and the extreme events. The improvement in the wavelet-based models can be attributed to their ability to unravel the hidden relationship between the predictors and precipitation. It is also observed that there is considerable increase in the correlation between precipitation and the decomposed climatic variables. All these results suggest that wavelets aid in unravelling the relationship between local precipitation and large-scale climatic variables and improving the overall performance of the downscaling models.