Beyond satellite-based precipitation data: a novel soil moisture physics framework with Green-Ampt and Bayesian optimization for rainfall estimation

Abstract

Reliable rainfall constraints over land remain a limitation for climate diagnostics, land–atmosphere coupling analyses, and evaluation of gridded products, where heterogeneity and sparse observations challenge the reliability of satellite products. The SM2RAIN addresses this gap by inverting the soil-moisture water balance, but its skill is governed by how infiltration and drainage physics are represented and by the depth at which soil moisture is observed. Here we show that infiltration–drainage parameterization and soil-moisture depth constitute bottlenecks for soil-moisture-based rainfall inversion. We introduce SM2RAIN–GreenAmpt, which embeds wetting-front Green–Ampt infiltration as a physically based inverse constraint, and we evaluate it using multi-depth in situ soil moisture across four layers (0–20 cm). In parallel, we enhance SM2RAIN–NWF by learning site-specific physical parameters via Bayesian optimization and we evaluate the model response to parameter variability for both schemes using one-at-a-time stochastic sensitivity analysis. Across depths, both bottom-up retrievals improve Nash–Sutcliffe efficiency and reduce RMSE relative to a satellite precipitation benchmark, while performance degrades with depth as the rainfall imprint attenuates. Stratified analyses by land cover further reveal regime-dependent gains and help interpret where physics-constrained inversion is most transferable. Together, these developments yield more defensible rainfall constraints for climate-oriented analyses and benchmarking of gridded precipitation products, with explicit uncertainty characterization. © The Author(s) 2026.