Estimating Diurnal Patterns of Land Surface Temperature using Vision Transformers and Satellite Images

Abstract

Diurnal cycles, the recurring 24–hour patterns produced by Earth’s rotation shape a wide range of environmental processes including temperature variation, evapotranspiration, and soil thermal dynamics. Land surface temperature (LST), one of the 54 Essential Climate Variables defined by the Global Climate Observing System, serves as a central parameter in climatological, hydrological, agricultural, and ecological studies. However, obtaining complete diurnal LST patterns remains difficult. The sparse coverage of in-situ stations, together with cloud contamination, environmental factors, sensor outages, and scan mismatches in satellite imagery, interrupt temporal continuity and leave large gaps in the record. This study introduces DayView, a spatiotemporal deep learning framework designed to reconstruct full diurnal cycles of LST from a single satellite observation, regardless of acquisition time. The methodology draws on hourly products from the GOES–R satellite series over the contiguous United States and integrates ancillary information such as climatic zones and elevation. Built on a Vision Transformer (ViT) architecture with a Masked Autoencoder strategy, \textsc{DayView} directly addresses three core challenges: (1) estimating diurnal cycles from sparse observations, (2) incorporating environmental context to refine fluctuation modeling, and (3) extending predictions reliably across continental scales. Empirical validation using remote sensing datasets demonstrates that DayView achieves high accuracy and strong robustness across diverse spatial and temporal conditions.Because the method is not limited to temperature alone, it can also be applied to other diurnal phenomena, such as solar–induced fluorescence, thus advancing environmental monitoring, climate analysis, and decision making at scale.