Reconstructing high-resolution gridded precipitation data in the southwest China highland canyon area using an improved (MGWR) downscaling method
Lihong Wang,Yuechen Li,Yushi Gan,Long Zhao,Lei Fan,Wei Qin,Lin Ding
DOI: https://doi.org/10.1016/j.scitotenv.2024.174866
2024-10-20
Abstract:High-spatiotemporal-resolution rainfall data are vital for investigating local terrestrial water cycles. Although remote-sensing satellite retrieval of precipitation products effectively reproduces spatial patterns of rainfall, it suffers from low spatial resolution. To overcome such limitations, a two-step downscaling approach is proposed here. First, 80 % of the meteorological-station data is utilized to calibrate the original Global Precipitation Measurement (GPM) data, enhancing the correlation between GPM and station data. Subsequently, utilizing elevation, slope, aspect, the normalized difference vegetation index (NDVI), wind direction, water vapor, and land surface temperature, as well as slope and aspect correction factors, as independent variables, multiscale geographically weighted regression (MGWR) and temporal lag MGWR (TL-MGWR) models were constructed. Through the aforementioned steps, downscaled monthly and daily precipitation data for the geographic region under investigation in 2022 at a spatial resolution of 0.01° were obtained. Our findings indicate that selectively employing suitable MGWR or TL-MGWR models on a monthly basis can effectively downscale monthly GPM rainfall data. The downscaled (original) monthly precipitation data exhibited a correlation of 0.94 (0.768), with a mean absolute error (MAE) of 16.233 mm/month, root-mean-square error (RMSE) of 27.106 mm/month, and bias of -0.043. Similar enhancement was likewise noted in daily precipitation, displaying a correlation coefficient of 0.863 (0.318) for downscaled (original) data, and a RMSE of 3.209 mm/day, MAE of 1.082 mm/day, and bias of -0.06. The downscaled results show a correlation increase of 0.172 monthly and 0.545 daily, with MAE reductions of 18.43 mm/month and 1.658 mm/day, RMSE reductions of 26.172 mm/month and 4.183 mm/day, and bias reductions of 82.7 % and 56.8 %. In summary, the data after downscaling, both for monthly and daily datasets, was markedly improved in accuracy. The proposed downscaling method is applicable for reconstructing high-resolution grid data in the complex terrain of the southwest China highland canyon area.