资源环境科技发展态势分析平台

Current gridded precipitation datasets are hard to meet the requirements of hydrological and meteorological applications in complex-terrain areas due to their coarse spatial resolution and large uncertainties. High-resolution atmospheric simulations are capable of describing the influence of topography on precipitation but are difficult to be used to obtain long-term precipitation datasets because they are computationally expensive, while reanalysis data has a long-term coverage and can provide reasonable large-scale spatial and temporal variability of precipitation. This study presents an approach to obtain long-term high-resolution precipitation datasets over complex-terrain areas by combining the ERA5 reanalysis with short-term high-resolution atmospheric simulation. The approach consists of two main steps: first, the ERA5 precipitation is corrected by the high-resolution simulation at the coarse spatial resolution; second, the corrected data is downscaled using a convolution neural network (CNN) based model at daily scale. The proposed approach is applied to the Tibetan Plateau (TP). The downscaled results from ERA5 have a finer spatial structure than ERA5 and can reproduce the spatial patterns of precipitation revealed by the high-resolution simulation. An evaluation based on rain gauge data shows that the downscaled ERA5 has remarkably lower biases than the original ERA5 which overestimates precipitation a lot, and even higher accuracy than the high-resolution simulation data over the TP. The downscaled ERA5 preserves the temporal characteristics of ERA5 which are more consistent with the rain gauge data than that of high-resolution simulation. Since this approach is much less computing resources consuming than the high-resolution simulation, it is an effective method to obtain long-term high-resolution precipitation datasets in complex-terrain areas and is expected to have extensive applications.