Compiling an hourly gridded dataset for surface air temperature at 50-m resolution using radiative cooling scale and numerical weather prediction model outputs

Meteorological data are essential components of decision-making systems for agriculture. Fine-resolution meteorological datasets are important where crops are grown in hilly areas with variable microclimates. A method for compiling daily surface air temperature (SAT) with 50 m resolution was previously developed. The compiled daily air temperature forecasts are available to farmers who produce Ujicha, tea via a web application. However, the complex terrain affects the accuracy of these forecasts. Our goal was to develop a practical system to provide hourly SAT data with 50 m resolution for a day-to-day decision-making system for agriculture. We applied two methods: a method for developing forecast models to forecast potential temperature differences between each grid and an observation site to compile 50 m resolution grid data; and a method for correcting SAT differences between the global spectral model output and observations (referred to as GSM_error). In both these methods, SAT was continuously forecast using the radiative cooling scale (RCS) computed near a public observation site. The RCS is a meteorological factor defined as the difference in potential temperature between an upper air pressure level and that at ground level. Forecast models for GSM_error were developed from observation data using a stepwise multiple regression (SMR), with geographic factors as independent variables in each group classified by the RCS. Statistical techniques were used for observation data to develop forecast models using machine learning, including lasso, ridge, random forest, support vector and deep neural network regression techniques. The SMR produced better forecast models for GSM_error than the other techniques. We found that SMR was better because it selected features relevant to GSM_error compared with machine learning, which fitted all features. Our grid data were compiled for hourly SAT with 50 m resolution and had a root mean square error of 1.67 °C.