Estimation of monthly air temperature using Random Forest algorithm

Air temperature (Ta) measurements can be obtained from the ground weather stations with high accuracy and temporal frequency. Nevertheless, the weather stations are limited with spatial coverage because of their sparse distribution. Fortunately, satellite data with the advantage of high spatial coverage can provide us land surface temperature (LST) observations for further extracting Ta through their strong relationship. In this study, we applied Random Forest (RF) to estimate monthly Ta with l-km resolution across Taiwan in 2020. The variables considered in the RF model include air temperature observations, LST version 6 from Moderate Resolution Imaging Spectroradiometer (MODIS), Normalized Difference Vegetation Index (NDVI), Normalized Difference Latent Heat Index (NDLI), altitude, longitude, latitude, and albedo. To assess the RF model's performance, we employed 5-fold cross-validation and used the coefficient of determination (R2), root-mean-square of error (RMSE), and mean absolute error (MAE) as the performance measurement. The results show excellent five-fold cross-validated performance of the RF model, with R2 of 0.965, RMSE of 0. 98°C, and MAE of 0. 73°c. The results indicate that Ta can be accurately estimated using the RF prediction model, even in Taiwan, with complex topography and weather patterns.