Balancing accuracy and feasibility in diurnal temperature modeling: A comparison of data-driven and physical-based models using geostationary satellite observations

The derivation of Diurnal Temperature Cycle (DTC) models from geostationary satellite data plays a critical role in temperature monitoring of the landscape and thermal anomaly applications such as wildfire detection. This study compares the performance of physical-based and data-driven DTC models on 1,305 study sites across Australia, leveraging Himawari-8 AHI middle-infrared (MIR) band 7 data. The physical-based model, GOT09 (based on G & ouml;ttsche and Olesen study), achieved the highest accuracy, with a mean validation Root Mean Square Error (RMSE) of 2.41 K, but its practical application was limited by a lower model generation rate (48.77%), especially under high cloud cover conditions. Among data-driven methods, the proposed TRI model (named after the first author) balances accuracy and practical feasibility, achieving a validation RMSE of 3.62 K and a generation rate of 85.07%. The TRI model consistently generated reliable DTCs under various environmental conditions, including high cloud cover, outperforming alternative data-driven models such as RW (from Roberts-Wooster study), XIE (from Xie et al. study), and HAL (from Hally et al. study). Additionally, the TRI model maintained reliability across diverse land cover and climate types, showing only minimal variations in performance. The study further highlights strategies for addressing cloud and data availability challenges, proposing methods such as the use of previous day's DTC or adjusting training data criteria in cloudy conditions. These approaches ensure a continuous temperature background where continuity of measurements is required, such as for wildfire detection. Overall, the research underscores the importance of balancing accuracy and model generation rates in DTC modeling, particularly for real-time applications. Future work could explore hybrid models and additional factors to further improve performance.