A multiple-step scheme for the improvement of satellite precipitation products over the Tibetan Plateau from multisource information.

Precipitation data with high accuracy and spatial resolution characteristics play significant roles in the regional hydrological and eco-environmental system applications. Thus, satellite-based precipitation products (SPPs) with high spatial resolution and high accuracy must be developed. This study proposed a multiple-step scheme to improve the global precipitation measurement (GPM) at daily and monthly timescales over the Tibetan Plateau (TP) from 2014 to 2017. First, combined with the geographically weighted regression (GWR) method using geographic and topographic factors, the gamma-distribution mapping and local intensity scaling (GDM-LOCI) method is applied to effectively merge the observed data attributes and the spatial representation of SPPs at the daily scale by correcting precipitation volumes and frequencies. Second, the areal merged precipitation, normalized difference vegetation index (NDVI), and reanalyzed atmospheric data are used to improve the spatial resolution of monthly GPM with a random forest (RF) model that uses the 17 land cover types to establish the local downscaling model windows. The results show that daily merged precipitation can better reflect the spatial and temporal variability of precipitation than can satellite estimates, and the correlation coefficient (R), and critical success index (CSI) increased by 0.12 and 0.26, respectively. In the merged downscaling model, the merged precipitation factor can weaken the negative effect of the other auxiliary predictors due to its spatial autocorrelation in precipitation estimation. Most importantly, by using the land cover types to establish local model windows for the downscaling model, not only the spatial resolution of the GPM product is downscaled to 1 km, but also the spatial structure of the downscaled products is enhanced, with less deviation and a higher spatial correlation. The R, the root mean square error (RMSE) and the relative bias (BIAS) were 0.89, 50.19 mm and 0.57, respectively. This study presents a promising scheme for generating high-quality precipitation data for regional hydrometeorological research in data-scarce regions.