IMPROVING THE ACCURACY OF SATELLITE-BASED NEAR SURFACE AIR TEMPERATURE AND PRECIPITATION PRODUCTS

Abstract

Abstract. In this study, we evaluate the performance of several reanalyses and satellite-based products of near-surface air temperature and precipitation to determine the best product in estimating daily and monthly variables across the complex terrain of Turkey. Each product’s performance was evaluated using 1120 ground-based gauge stations from 2015 to 2019, covering a range of complex topography with different climate classes according to the Köppen-Geiger classification scheme and land surface types according to the Moderate Resolution Imaging Spectroradiometer (MODIS). Furthermore, various traditional and more advanced machine learning downscaling algorithms were applied to improve the spatial resolution of the products. We used distance-based interpolation, classical Random Forest, and more innovative Random Forest Spatial Interpolation (RFSI) algorithms. We also investigated several satellite-based covariates as a proxy to downscale the precipitation and near-surface air temperature, including MODIS Land Surface Temperature, Vegetation Index (NDVI and EVI), Cloud Properties (Cloud Optical Properties, Cloud Effective Radius, Cloud Water Path), and topography-related features. The agreement between the ground observations and the different products, as well as the downscaled temperature products, was examined using a range of commonly employed measures. The results showed that AgERA5 was the best-performing product for air temperature estimation, while MSWEP V2.2 was superior for precipitation estimation. Spatial downscaling using bicubic interpolation improved air temperature product performance, and the Random Forest (RF) machine learning algorithm outperformed all other methods in certain seasons. The study suggests that combining ground-based measurements, precipitation products, and features related to topography can substantially improve the representation of spatiotemporal precipitation distribution in data-scarce regions.