LAUSR

Near-surface air temperature (NSAT) is a key parameter in climate changes, environmental ecosystem monitoring, and human settlement issues. As it is difficult for in situ observations to capture the spatial distribution characteristics of NSAT in great detail, various methods have been developed to use remotely sensed land surface temperature and other auxiliary variables to estimate the NSAT. Among them, machine learning turns out to be an exhilarating choice due to its superior performance. However, for long-term dataset estimation, when using machine learning methods with abundant related variables, computation loads and data storage cannot be ignored. Fortunately, Google Earth Engine (GEE) provides a wealth of geospatial datasets and powerful parallel computation capability. In this article, GEE has been utilized to demonstrate the feasibility of estimating long-term 1-km NSAT with two machine-learning models (random forest and deep neural network). After testing the effects of the input variables and the model portability, the 20-year monthly mean 1 km × 1 km NSAT over the Yellow River basin was generated and analyzed. Compared with in situ observations, the overall RMSE, MAE, R2, and R of the NSAT product are 0.429, 0.302, 0.998, and 0.999, respectively. Specifically, for each observation station, R and R2 are greater than 0.997, and RMSE and MAE are smaller than 0.5 and 0.4, respectively. The comparisons with three existing NSAT products show that our product outperforms the other three products. Once again, it is demonstrated in this study that GEE is a powerful platform to generate valuable products with ready-to-use datasets and computation resources.