Journal of Water and Climate Change (Apr 2024)
Improving the statistical downscaling performance of climatic parameters with convolutional neural networks
Abstract
This study examines two downscaling techniques, convolutional neural networks (CNNs) and feedforward neural networks for predicting precipitation and temperature, alongside statistical downscaling as a benchmark model. The daily climate predictors were extracted from the European Center for Medium-range Weather Forecast (ECMWF) ERA5 dataset spanning from 1979 to 2010 for Tabriz city, located in the northwest of Iran. The biases in precipitation data of ERA5 predictors were corrected through the empirical quantile mapping method. Also, two nonlinear predictor screening methods, random forest and mutual information, were employed, alongside linear correlation coefficient. While these methods facilitate identification of dominant regional climate change drivers, it is essential to consider their limitations, such as sensitivity to parameter settings, assumptions about data relationships, potential biases in handling redundancy and correlation, challenges in generalizability across datasets, and computational complexity. Evaluation results indicated that CNN, when applied without predictor screening, achieves coefficient of determination of 0.98 for temperature and 0.71 for precipitation. Ultimately, future projections were employed under two shared socioeconomic pathways (SSPs), SSP2-4.5 and SSP5-8.5, and concluded that the most increase in temperature by 2.9 °C and decrease in precipitation by 3.5 mm may occur under SSP5-8.5. HIGHLIGHTS Convolutional neural networks (CNNs) and feedforward neural networks (FFNNs) were used for downscaling general circulation model.; The empirical quantile mapping method was used for bias correction.; Future projections were employed under two shared socioeconomic pathways (SSPs), i.e., SSP2-4.5 and SSP5-8.5.; The results show the superiority of CNN over other AI methods.;
Keywords