**1. Introduction**

Precipitation is the main component of the global water cycle and plays a critical role in Earth's energy balance [1,2]. Therefore, accurate information on the spatial and temporal distribution of precipitation is essential to improve our understanding of the Earth system and to better predict weather and climate conditions and natural disasters. Ground observations are the most direct source of precipitation data, but most stations are unevenly distributed in low-altitude zones, which makes it difficult to capture the full distribution of large-scale precipitation. In contrast, satellite-retrieved precipitation products have the unique advantages of global coverage and spatiotemporal continuity [3,4] and have

consequently promoted a more complete understanding of the patterns of and changes in regional and global precipitation.

However, because of instrument limitations and imperfect retrieval algorithms, satellite precipitation products have drawbacks in terms of spatial resolution and data precision [5–9]. At present, because of the coarse spatial resolution of satellite precipitation products, their application in hydrological and climatic models at the watershed scale is restricted. For this reason, many researchers have focused on developing statistical downscaling methods for satellite or reanalysis precipitation products [10–12]. These methods usually involve building a relationship between coarse-resolution precipitation data and high-resolution variables to improve the spatial resolution of satellite precipitation data. For example, Immerzeel et al. [13] established an exponential regression (ER) model by integrating 1-km normalized difference vegetation index (NDVI) data with Tropical Rainfall Measuring Mission (TRMM) 3B43 precipitation data and obtained high-resolution annual precipitation data over the Iberian Peninsula. Based on the method of Immerzeel, Jia et al. [14] established a functional relationship between 3B43 precipitation data and other variables (i.e., altitude and NDVI) using multiple linear regression (MLR) and obtained downscaled annual data at a 1 km resolution for the Qaidam Basin in China. Duan et al. [15] developed a further modified downscaling algorithm by introducing calibration methods based on geographic difference analysis (GDA) and geographic ratio analysis (GRA) and obtained 1-km monthly precipitation data over the Tana Lake Basin in Africa and the coast of the Caspian Sea in Asia. Zhang et al. [16] applied the abovementioned methods in Xinjiang, China, and obtained 1-km annual precipitation data. Considering the spatial variations exhibited by the relationship between precipitation and environmental variables, geographically weighted regression (GWR) has been introduced into precipitation analyses to achieve improved downscaling performance [17–20]. However, although environmental variables play a vital role in the monthly or yearly downscaling of precipitation, they have limited applicability in daily and hourly downscaling, which is more strongly reliant on cloud properties. For example, Sharifi et al. [21] obtained 1-km daily precipitation data in northeast Austria using MLR, artificial neural networks (ANNs) and spline interpolation based on 1-km cloud optical thickness (COT), cloud effective radius (CER), and cloud water path (CWP) data. Ma et al. [22] obtained 1-km hourly precipitation data in the southeast coast region of China based on COT, CER, and cloud top height (CTH) data from Himawari 8.

Traditional downscaling methods can improve the spatial resolution of satellite precipitation data. Many studies have shown that the accuracy of the satellite precipitation products is the most important factor affecting the quality of the downscaled estimates [13,14,17] if the environmental variables can satisfactorily reproduce the pattern of the satellite precipitation data. However, all previous downscaling methods have been applied to original satellite precipitation data, which contain large uncertainties that limit the accuracy of the downscaled precipitation estimates. Correcting the original satellite products before applying them in downscaling analyses can potentially contribute to the improvement of the downscaled precipitation estimates. Nevertheless, to our knowledge, no such study has previously been performed.

To address this gap, this study proposes a two-step merging and downscaling framework called OI-GWR to improve the accuracy of downscaled precipitation estimates. The merging procedure is based on optimum interpolation (OI), and the downscaling procedure is based on GWR. The remaining sections of this paper are organized as follows: Section 2 introduces an overview of the study area; Section 3 provides detailed information about the data and methods; Section 4 reports the results of OI and GWR; and finally, a discussion and conclusions are presented in Sections 5 and 6, respectively. The method proposed in this study will contribute to the production of high-quality and high-resolution regional gridded precipitation datasets. In particular, the method can serve as a useful reference for the development of grid data at the daily or hourly scale.
