Sang et al. (2026) SMAP Soil Moisture Downscaling Via Multimodal Data Fusion and Machine Learning

Identification

• Journal: Earth Systems and Environment
• Year: 2026
• Date: 2026-02-04
• Authors: Xin Zhu Sang, Xiaoping Lu, Kailun Wang, Junli Zhou, Guosheng Cai, Jinrui Fan
• DOI: 10.1007/s41748-026-01031-8

Research Groups

• Key Laboratory of Spatio-Temporal Information and Ecological Restoration of Mines, Ministry of Natural Resources of the People’s Republic of China, Henan Polytechnic University, Jiaozuo, China
• Henan Zhilian Spatio-Temporal Information Technology Co., Ltd, Zhengzhou, China
• Henan Remote Sensing Institute, Zhengzhou, China

Short Summary

This study proposes a novel method to downscale Soil Moisture Active Passive (SMAP) soil moisture from 9 km to 1 km resolution by integrating Solar-induced chlorophyll fluorescence (SIF) and multi-source remote sensing data with machine learning. The Random Forest model incorporating SIF demonstrated superior performance, significantly enhancing the spatial detail and temporal consistency of the downscaled soil moisture product for improved drought monitoring and agricultural management.

Objective

• To develop and evaluate a downscaling method for SMAP soil moisture (SM) from 9 km to 1 km spatial resolution by integrating vegetation physiological characteristics (Solar-induced chlorophyll fluorescence, SIF) with multi-source remote sensing data and machine learning, addressing the temporal lag issues of traditional vegetation indices.

Study Configuration

• Spatial Scale: Henan Province, China (167,000 km²). SMAP soil moisture downscaled from 9 km to 1 km resolution.
• Temporal Scale: May 2018 to December 2021 (44 months). Monthly scale for the downscaled soil moisture product.

Methodology and Data

• Models used:
  • For SIF reconstruction: Random Forest (RF)
  • For SMAP downscaling: Support Vector Regression (SVR), XGBoost, Random Forest (RF), Deep Neural Networks (DNN), and a Stacking ensemble. The optimal model identified was Random Forest (RF) with SIF.
• Data sources:
  • Satellite/Remote Sensing:
    • SMAP L4-level Soil Moisture (0–5 cm depth, 9 km, 3 h temporal resolution).
    • TROPOMI enhanced Solar-induced chlorophyll fluorescence (eSIF) (0.05° (~5.5 km), 8-day temporal resolution).
    • MODIS products: Land Cover (MCD12Q1, 500 m), Evapotranspiration (MOD16A2, 500 m, 8-day), Leaf Area Index (MOD15A2H, 500 m, 8-day), Surface multispectral reflectance (MCD43A4, 500 m, for NDVI, EVI, NIRV, NDWI, NSDSI), Albedo (MCD43A3, 500 m), Land Surface Temperature (MOD11A2, 1 km, 8-day, for Apparent Thermal Inertia (ATI)).
  • Topographic: SRTM 90 m Digital Elevation Model (DEM) (for DEM and Slope).
  • Soil Properties: SoilGrids V2.0 (clay, sand, silt content at 0–5 cm depth, 250 m).
  • Precipitation: CHIRPS daily precipitation data.
  • Validation Data: China 1 km daily all-weather surface soil moisture dataset (SSM), China 1 km daily soil moisture dataset (SMCI1.0).

Main Results

• SIF Reconstruction Accuracy: The reconstructed SIF data showed high consistency with original SIF, with R² values generally above 0.72 and reaching up to 0.96, along with low RMSE and MAE.
• SMAP Downscaling Model Performance: The Random Forest (RF) model incorporating SIF (RF with SIF) achieved the best performance with an R² of 0.92 and an RMSE of 0.023 m³/m³. This outperformed RF without SIF, Stacking, XGBoost, DNN, and SVR models.
• Downscaled SM Validation:
  • The downscaled SMAP SM product (DSMAP_SIF) exhibited a mean correlation coefficient (R) of 0.864 with 57 in-situ observation sites, with 41 sites showing R values between 0.8 and 1.0.
  • The mean unbiased root mean square error (ubRMSE) was 0.040 m³/m³ (minimum 0.015 m³/m³), and the mean RMSE was 0.047 m³/m³. A slight systematic negative bias (mean -0.046 m³/m³) was observed.
  • DSMAPSIF outperformed the original SMAP L4SM (mean R = 0.821) and downscaled results without SIF.
• Spatial and Temporal Characteristics: The downscaled SMAP SM product showed enhanced spatial detail and heterogeneity, accurately capturing local soil moisture variations. The temporal correlation improved from 0.686 (vegetation index-based) to 0.713 (SIF-based), demonstrating superior responsiveness to precipitation events (R² > 0.8).
• Feature Importance: Precipitation and Evapotranspiration were the most important features. SIF ranked sixth in importance, outperforming traditional vegetation indices like NDVI, EVI, and NIRV, confirming its value as a physiological indicator.

Contributions

• Introduced Solar-induced chlorophyll fluorescence (SIF) as a direct physiological vegetation factor for SMAP soil moisture downscaling, effectively addressing the temporal lag inherent in traditional vegetation index-based methods.
• Developed a robust multi-source data fusion and machine learning framework that significantly enhances the spatial resolution (from 9 km to 1 km) and accuracy of SMAP soil moisture products.
• Demonstrated the superior performance of the Random Forest model when integrating SIF for capturing fine-scale soil moisture heterogeneity and improving temporal consistency.
• Generated a high-resolution (1 km) monthly SMAP soil moisture product for Henan Province, China, providing a valuable tool for agricultural drought monitoring and water resource management.

Funding

• National Key Research and Development Plan (Grant No. 2016YFC0803103)
• Henan Provincial Youth Student Scientific Research Fund Project (Grant No. 252300423933)

Citation

@article{Sang2026SMAP,
  author = {Sang, Xin Zhu and Lu, Xiaoping and Wang, Kailun and Zhou, Junli and Cai, Guosheng and Fan, Jinrui},
  title = {SMAP Soil Moisture Downscaling Via Multimodal Data Fusion and Machine Learning},
  journal = {Earth Systems and Environment},
  year = {2026},
  doi = {10.1007/s41748-026-01031-8},
  url = {https://doi.org/10.1007/s41748-026-01031-8}
}