Estimating Field-Scale, Depth-Specific Soil Moisture from NOAA’s GEFS Soil Moisture Products Using Machine Learning

Soil moisture (SM) information from forecasting systems is essential for scheduling irrigation, monitoring drought, and making hydrological decisions. However, model-derived SM estimates are rarely calibrated to field-scale conditions. Most existing machine learning studies have focused on correcting meteorological variables rather than directly calibrating soil moisture products. This study evaluated the ability to employ machine learning to calibrate soil moisture estimates at specific depths from the Global Ensemble Forecast System (GEFS), using in-situ observations taken at depths of 15, 20, 25, and 30 cm. In-situ sensor measurements were collected across five monitoring sites in Alabama, USA, and three experiments were conducted. First, we compared Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Multiple Linear Regression (MLR) using an expanding-window nested cross-validation framework. Second, the predictive value of the GEFS’s surface (0–10 cm), deep (10–40 cm), and combined soil moisture layers was evaluated. Finally, global and site-specific calibration strategies were assessed using an independent holdout dataset. XGBoost and RF generally achieved higher mean R2 values than MLR, particularly at deeper soil depths, although differences were not statistically significant. Combining the surface and deep GEFS layers achieved the highest mean R2 at depths of 20, 25, and 30 cm. In contrast, the surface layer alone achieved the highest accuracy at a depth of 15 cm. Feature importance analysis identified site identity and rolling-average GEFS SM variables as the dominant predictors. Using independent evaluation across four sites, the global XGBoost model achieved site-averaged R2 values ranging from 0.446 to 0.516 and reduced RMSE by approximately 46–54% compared with raw GEFS soil moisture estimates. The model achieved an additional 10–21% reduction in RMSE when compared to additive mean-bias correction. The highest mean R2 was also observed at 30 cm. These results show the potential of machine learning-based bias correction to improve GEFS SM analysis data for field-scale water management applications.

Authors

Institutions

Publication Details

Journal
Hydrology
Published
2026-11-19
DOI
https://doi.org/10.3390/hydrology13110310
Primary Topic
Soil Moisture and Remote Sensing
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Estimating Field-Scale, Depth-Specific Soil Moisture from NOAA’s GEFS Soil Moisture Products Using Machine Learning

Souleymane Fall, Eniola Emmanuel Olakanmi, Joseph E. Quansah, Sara Gutierrez Diaz
Hydrology
Soil Moisture and Remote Sensing
article

Estimating Field-Scale, Depth-Specific Soil Moisture from NOAA’s GEFS Soil Moisture Products Using Machine Learning

Souleymane Fall, Eniola Emmanuel Olakanmi, Joseph E. Quansah, Sara Gutierrez Diaz
article en

Abstract

Soil moisture (SM) information from forecasting systems is essential for scheduling irrigation, monitoring drought, and making hydrological decisions. However, model-derived SM estimates are rarely calibrated to field-scale conditions. Most existing machine learning studies have focused on correcting meteorological variables rather than directly calibrating soil moisture products. This study evaluated the ability to employ machine learning to calibrate soil moisture estimates at specific depths from the Global Ensemble Forecast System (GEFS), using in-situ observations taken at depths of 15, 20, 25, and 30 cm. In-situ sensor measurements were collected across five monitoring sites in Alabama, USA, and three experiments were conducted. First, we compared Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Multiple Linear Regression (MLR) using an expanding-window nested cross-validation framework. Second, the predictive value of the GEFS’s surface (0–10 cm), deep (10–40 cm), and combined soil moisture layers was evaluated. Finally, global and site-specific calibration strategies were assessed using an independent holdout dataset. XGBoost and RF generally achieved higher mean R2 values than MLR, particularly at deeper soil depths, although differences were not statistically significant. Combining the surface and deep GEFS layers achieved the highest mean R2 at depths of 20, 25, and 30 cm. In contrast, the surface layer alone achieved the highest accuracy at a depth of 15 cm. Feature importance analysis identified site identity and rolling-average GEFS SM variables as the dominant predictors. Using independent evaluation across four sites, the global XGBoost model achieved site-averaged R2 values ranging from 0.446 to 0.516 and reduced RMSE by approximately 46–54% compared with raw GEFS soil moisture estimates. The model achieved an additional 10–21% reduction in RMSE when compared to additive mean-bias correction. The highest mean R2 was also observed at 30 cm. These results show the potential of machine learning-based bias correction to improve GEFS SM analysis data for field-scale water management applications.

HydrologyVol. 13(11)
Tuskegee University (US)
Soil Moisture and Remote Sensing
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon