Machine learning based high-resolution air temperature modelling from landsat-8, MODIS, and In-Situ measurements with ERA-5 inter-comparison in the data sparse regions of Himachal Pradesh

Accurately estimating air temperature across spatially dispersed areas holds significant value across a range of disciplines, including hydrology, meteorology, and ecology. For this purpose, ERA-5 air temperature data is globally used but it has a coarse spatial resolution and has a higher bias in the mountainous regions that presents a challenge when local variations in air temperature are significant. In this work, we employed techniques such as Multilayer-Perceptron (MLP), Generalized-Additive-Models (GAM), and Multiple-Linear-Regression (MLR) to predict air temperature using Land-Surface-Temperature (LST) and sun geometry data from the high spatial resolution Landsat-8 and the high temporal resolution MODIS Aqua and Terra satellites along with auxiliary data and in-situ air temperature measurements given by the Indian Meteorological Department (IMD). Our findings indicate that MLP outperforms GAM and MLR in accuracy for predictions using MODIS data. High correlations are detected in Bhuntar station (R² = 0.94, RMSE = 1.13 °C and MAE = 0.83 °C) between IMD observed daily mean air temperature and Landsat-8 predicted mean air temperature. In the Kalpa station, while a strong correlation exists between observed and predicted air temperatures for all satellite products, the RMSE and MAE values are comparatively higher. Additionally, we compared the ERA-5 data with our predicted results and in-situ data. It was observed that RMSE, MAE and bias are significantly higher for the ERA-5 data. Thus, the method and the coefficient derived in this work could be used for estimation of air temperature from satellite data at high resolution.