[A-HW28] Hydrology and Water Environment

[AHW28-P08] AI-based Near-Real-Time Soil Moisture Downscaling

*Soumita Sengupta¹, Hone-Jay Chu¹ (1.Department of Geomatics, National Cheng Kung University, No. 1, University Road, Tainan City 701, Taiwan.)

Keywords:Soil Moisture, Downscaling, Machine Learning, Deep Learning, Remote Sensing, Near Real-Time

Soil moisture (SM) plays a critical role in hydrological processes, drought and flood management, climate dynamics, and agricultural productivity. High-resolution (HR) SM data provide fine-scale spatial details but are prone to high uncertainties. While low-resolution (LR) SM data capture large-scale hydrological patterns, offering greater spatial stability and ensuring the continuity of large-scale soil moisture dynamics. Thus, the advantages of HR and LR data are essential for accurate representation of water balance and prevent artificial distortions in soil moisture distribution. Existing downscaling methods integrate both LR and HR data but their optimization remains HR-dominated, leading to excessive reliance on HR features, which disrupts large-scale hydrological consistency and reduces the stability of downscaled estimates. To overcome this limitation, we propose a hybrid machine learning (ML) and residual learning framework for near real-time SM downscaling, featuring a residual learning architecture equipped with a novel combined loss function to ensure equal LR-HR contributions. The framework integrates multi-source data—including in-situ measurements, satellite observations (SMAP, AMSR2/GCOM-W1, SMOPS), meteorological variables, and environmental parameters—to generate high-precision, near real-time SM estimates across the United Kingdom (UK). It follows a two-stage approach: (1) an ensemble ML model (Random Forest, Gradient Boosting, and Extreme Gradient Boosting) producing 9 km HR SM estimates, and (2) a residual learning model refining these estimates while aligning them with LR hydrological patterns. Results demonstrate that the ensemble ML model effectively captures regional SM variability, with high SM in the northwest, moderate SM in the northeast, low SM in the narrow neck region, moderate SM in central and southwestern UK, and low SM in the southeast with localized moderate values. The model achieved an R² of 0.77 across 40 stations, with 36 stations exceeding 75% accuracy, 2 stations between 50–75%, and 4 stations below 50%. The advanced model further enhances spatial consistency, particularly in southwestern UK, where it aligns better with LR hydrological patterns, achieving an RMSE of 0.0538. This study focuses on enhancing single-time SM predictions for near real-time applications through a scalable, computationally efficient framework. The findings demonstrate that the novel combined loss function in the residual learning model architecture effectively reduces uncertainty, enhances fine-scale resolution, and preserves physical realism, offering a robust solution for near real-time SM estimation. This research supports climate resilience, water resource management, and precision agriculture.