Abstract & Overview
Accurate soil moisture prediction during extreme events remains a challenge for Earth system modeling, with significant implications for drought monitoring, flood forecasting, and climate adaptation strategies. Although land surface models (LSMs) provide physics-based predictions, their parameterizations struggle to represent the non-linear feedbacks and threshold behaviors that emerge during hydrometeorological extremes, producing systematic biases. Here we present a deep learning bias-correction layer for Noah-MP soil moisture. A 3D U-Net architecture for Noah-MP outputs (soil moisture, latent heat flux, and sensible heat flux) is trained against SMAP L4 soil moisture across two contrasting extreme events: a drought (March-September 2022) and a hurricane (July 2024) over Texas. The bias correction layer achieves substantial improvements over standalone Noah-MP, increasing R² from below the climatological mean to 0.5 during drought while reducing prediction errors by 40-60\% for both events, all while preserving spatial coherence. The framework operates as a post-processing layer on LSM output, addressing forecast windows where direct observational correction is unavailable. This provides a deployable pathway for integrating land surface models with deep learning in Earth system forecasting.