Feng et al. (2026) Global daily 9 km remotely sensed soil moisture (2015–2025) with microwave radiative transfer-guided learning

Identification

• Journal: Scientific Data
• Year: 2026
• Date: 2026-02-12
• Authors: Sijia Feng, Andi Li, Rui Zhou, Ralf Kiese, Kaiyu Guan, Zhenong Jin, Majken C. Looms, Sherrie Wang, Christian Igel, Claire C. Treat, Jørgen Eivind Olesen, Sheng Wang
• DOI: 10.1038/s41597-026-06721-6

Research Groups

• Pioneer Center Land-CRAFT, Department of Agroecology, Aarhus University, Aarhus, Denmark
• Agroecosystem Sustainability Center, Institute for Sustainability, Energy, and Environment, University of Illinois Urbana-Champaign, Urbana, IL, USA
• Department of Natural Resources and Environmental Sciences, College of Agricultural, Consumers, and Environmental Sciences, University of Illinois Urbana-Champaign, Urbana, IL, USA
• National Center for Supercomputing Applications, University of Illinois Urbana-Champaign, Urbana, IL, USA
• Department of Computer Science, University of Illinois Urbana-Champaign, Urbana, IL, USA
• Karlsruhe Institute of Technology, Institute for Meteorology and Climate Research, Atmospheric Environmental Research (IMK-IFU), Garmisch-Partenkirchen, Germany
• Department of Bioproducts and Biosystems Engineering, University of Minnesota, St Paul, MN, USA
• Department of Geosciences and Natural Resource Management (IGN), University of Copenhagen, Copenhagen, Denmark
• Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA
• Institute for Data, Systems, and Society, Massachusetts Institute of Technology, Cambridge, MA, USA
• Department of Computer Science, University of Copenhagen, Copenhagen, Denmark
• Pioneer Center for Artificial Intelligence, Copenhagen, Denmark

Short Summary

This study developed a Process-Guided Machine Learning (PGML) framework, integrating microwave radiative transfer theories with deep learning, to generate a global daily 9 km surface soil moisture (SM) dataset from 2015 to 2025. The resulting PGML SM product demonstrates superior accuracy (R=0.923, ubRMSE=0.040 m³/m³) compared to existing satellite and reanalysis products, particularly in regions with dense vegetation and complex surface conditions.

Objective

• To develop a microwave radiative transfer process-guided machine learning (PGML) framework that integrates microwave radiative transfer model (RTM) theories and deep learning to predict global daily 9 km surface soil moisture (SM) from April 2015 to June 2025, overcoming limitations of empirical RTM-based algorithms and purely data-driven machine learning approaches.

Study Configuration

• Spatial Scale: Global, 9 km spatial resolution.
• Temporal Scale: Daily, from April 2015 to June 2025.

Methodology and Data

• Models used: Process-Guided Machine Learning (PGML) framework based on a Multi-Layer Perceptron (MLP) architecture, guided by microwave radiative transfer model (RTM) theories (specifically, the τ-ω model). The PGML model was pre-trained with synthetic data from the τ-ω model and fine-tuned with in-situ measurements.
• Data sources:
  • Satellite observations: SMAP L-band Brightness Temperature (Tb) (SPL3SMP_E, 9 km, daily), MODIS Normalized Difference Vegetation Index (NDVI) (MOD13C1, 0.05°, 16-day).
  • Reanalysis data: ERA5-Land (Surface Temperature, Precipitation, Evaporation; 0.1°, hourly/daily).
  • Ancillary data: SoilGrid250 (Clay fraction, Soil bulk density; 250 m), MCD12C1 (Land cover types; 0.05°, yearly), Köppen-Geiger climate zones (1 km).
  • In-situ measurements: Global soil moisture measurements from the International Soil Moisture Network (ISMN), AmeriFlux, ICOS, JapanFlux, and other published studies (daily, top 5 cm).

Main Results

• The PGML SM dataset shows strong agreement with independent in-situ measurements, achieving a Pearson correlation coefficient (R) of 0.868 and an unbiased Root Mean Square Error (ubRMSE) of 0.054 m³/m³.
• When compared against seven other widely used global SM products (ESA CCI, SMOS-IC, SMOS L3, SMAP DCA, SMAP SCAV, SMAP-IB, ERA5-Land) using 69,115 in-situ records, PGML SM demonstrated superior performance with R = 0.923, bias = -0.001 m³/m³, and ubRMSE = 0.040 m³/m³.
• PGML significantly improved SM accuracy across diverse land cover types, particularly in regions with dense vegetation (e.g., North American forests) and croplands, by effectively correcting systematic biases.
• The PGML SM product exhibits similar global spatial distributions and spatiotemporal patterns (e.g., meridional migration of the Intertropical Convergence Zone) to other products, while capturing finer spatial variability in mid-latitude regions.
• PGML SM accurately captured drought signals during the 2018 European drought, showing higher R and lower systematic bias compared to ESA CCI SM, highlighting its potential for drought detection.

Contributions

• Development of a novel microwave radiative transfer process-guided machine learning (PGML) framework that deeply integrates physical RTM theories with deep learning, enhancing the accuracy and generalizability of global soil moisture (SM) estimation.
• Generation of a new, high-resolution (9 km) global daily surface SM dataset spanning from April 2015 to June 2025, which provides a valuable complement to existing satellite SM products.
• Demonstration that the PGML framework effectively overcomes the limitations of traditional RTM-based algorithms (e.g., empirical parameterization, oversimplified processes) and purely data-driven machine learning approaches (e.g., lack of physical knowledge, limited generalizability).
• The study's innovative approach includes designing the model architecture based on RTM and hydrological theories, utilizing a Kling-Gupta efficiency-based cost function, pre-training with RTM simulations, and fine-tuning with extensive in-situ measurements.

Funding

• Danish Data Science Academy (Novo Nordisk Foundation NNF21SA0069429, VILLUM FONDEN 40516)
• VILLUM Young Investigator 2024 project (grant no. 00072051)
• Novo Nordisk starting grant (grant no. NNF23OC0087612)
• SCALE project in AgriFoodTure, Innovation Fund Denmark
• NASA ECOSTRESS Science and Applications Program (80NSSC23K0308)
• NASA Early Career Investigator Program in Earth Science (80NSSC24K1057)
• Global Wetland Center (grant no. NNF23OC0081089) from the Novo Nordisk Foundation
• Pioneer Center for Research in Sustainable Agricultural Futures (Land-CRAFT), DNRF grant number P2, Aarhus University, Denmark

Citation

@article{Feng2026Global,
  author = {Feng, Sijia and Li, Andi and Zhou, Rui and Kiese, Ralf and Guan, Kaiyu and Jin, Zhenong and Looms, Majken C. and Wang, Sherrie and Igel, Christian and Treat, Claire C. and Olesen, Jørgen Eivind and Wang, Sheng},
  title = {Global daily 9 km remotely sensed soil moisture (2015–2025) with microwave radiative transfer-guided learning},
  journal = {Scientific Data},
  year = {2026},
  doi = {10.1038/s41597-026-06721-6},
  url = {https://doi.org/10.1038/s41597-026-06721-6}
}