High-resolution maximum air temperature estimation over India from MODIS data using machine learning

Abstract

Maximum air temperature is one of the crucial parameters required for climate change research, public health, agriculture, energy consumption, and urban planning. Observations of maximum temperature are limited spatially and temporally, and available observations are discontinuous due to operational constraints. With the advent of the satellite era, the land surface temperature is available continuously across the globe over space and time. This study explores the potential of using land surface temperature data from MODIS satellites to estimate maximum air temperature across India, particularly in areas with low cloud cover. The research employs advanced machine learning techniques to estimate the maximum temperature over India from MODIS land surface temperature and other inputs like NDVI, elevation, land use, geographical location, and the Julian day. We have assessed the capability of three machine learning techniques: XGBoost, Neural network, Generalized additive model and multiple linear regression in estimating maximum temperature over India using MODIS and Insitu data spanning from 2010 to 2022. Results indicate that XGBoost outperforms the other techniques, achieving the lowest RMSE and R² values of 1.79 °C and 0.90, respectively. Our findings reveal that land surface temperature is the most influential predictor of maximum air temperature, followed by Julian day, elevation, latitude, distance to coast, NDVI, and land cover type, in order of importance. This research demonstrates the potential of satellite-derived data and machine learning in addressing gaps in maxiumum temperature observations, which could significantly benefit various sectors reliant on accurate data.