Developing machine learning models for air temperature estimation using MODIS data

Abstract

Air temperature is a key variable in a wide range of environmental applications, including land–atmosphere interaction, climate change research and hydrology and crop growth models, among others. The objective of this study was to estimate daily maximum (Tmax) and minimum (Tmin) temperatures, based on MODIS AQUA/TERRA land surface temperature (LST), NDVI, extraterrestrial solar radiation and precipitation data. Artificial neural networks (ANN) and random forests (RF) models were developed to predict these temperatures covering weather stations in Córdoba (Argentina) for 2018-2020. The results show that RF and ANN machine learning algorithms are capable of modeling non-linear relationships between registered temperatures and LST MODIS data, in a very robust way. The validation of the models confirms that Tmax and Tmin can be accurately estimated using, jointly or separately, AQUA and TERRA LST. The best models present determination coefficients equal to 0.81/0.91 and root mean square error of 2.7/2.1 ºC for Tmax/Tmin, when using AQUA LST day/night satellite overpass time data, respectively. The robustness and confidence of the models developed, and the ease and free accessibility of input data at a global scale, suggest that these methodologies have the potential to be applied to other regions.