Quantifying Historical and Future Surface Soil Moisture Drying Using Deep Learning and Remote Sensing

Abstract

Understanding historical and future surface soil moisture (SSM) drying is pivotal due to its close links with droughts, heatwaves, and wildfires, yet debates regarding its evolution persist. In this study, we leverage advanced deep learning techniques to fill gaps of remote sensing-based SSM data during 1983–2020 and therefore use these gap-filled observations to constrain SSM estimates from 23 Earth System Models (ESMs) during 1901–2100. Our enhanced observations reveal that approximately half of Earth's landmass experienced SSM drying over the past four decades. However, in contrast to projections from current-generation ESMs, observation-constrained simulations indicate a less pronounced drying trend in dry-wet transitions and monsoon margins during 2021–2100 compared to 1901–1980. Current ESMs may overestimate SSM drying in these regions, likely due to their limited representation of soil moisture-atmosphere feedback. These findings highlight the need to integrate remote sensing and artificial intelligence into ESMs to improve projections of future droughts and their socio-economic consequences.