End-to-End Reconstruction of High-Resolution Temperature Data Using Physics-Guided Deep Learning

High-resolution land surface temperature data with fine spatiotemporal granularity is essential for real-world applications. While satellites provide observations at 100 m every 16 days and coarser resolution hourly, these observations are incomplete due to cloud cover and long revisit times. Earth system models provide continuous hourly temperature data but at much coarser spatial resolution (0.1$^\circ$ to 0.25$^\circ$). In this study, we present an end-to-end, physics-guided deep learning approach for temperature data reconstruction. The approach is a convolutional neural network that incorporates the annual temperature cycle and includes a linear term to amplify the coarse Earth system model temperatures using fine-scale satellite observations. We evaluate the approach using data from GOES-16 (2 km, hourly) and Landsat (100 m, every 16 days), demonstrating effective temperature reconstruction across selected areas. This simple yet effective approach, enabled by physics-guided deep learning, presents a promising direction for reconstructing temperature data under all weather conditions globally.