Yang et al. (2025) A novel Improved Geographically Weighted Random Forest (IGWRF) model for low-resolution soil moisture data downscaling in Africa

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-12-04
• Authors: Yunjie Yang, Lihui Wang, Xu Zhai, Xiaodi Zheng, Guosong Zhao, Qichi Yang, Yun Du, Feng Ling
• DOI: 10.1016/j.agwat.2025.110034

Research Groups

• School of Geography and Information Engineering, China University of Geosciences, Wuhan, China
• Key Laboratory for Environment and Disaster Monitoring and Evaluation, Innovation Academy for Precision Measurement Science and Technology, Chinese Academy of Sciences, Wuhan, China
• College of Resources and Environment, Anhui Agricultural University, Hefei, China
• University of Chinese Academy of Sciences, Beijing, China

Short Summary

This study proposes an Improved Geographically Weighted Random Forest (IGWRF) model to downscale 9 km SMAP soil moisture data to 1 km in Kenya, effectively addressing spatial heterogeneity and nonlinear relationships. The IGWRF model significantly outperforms traditional Random Forest and Geographically Weighted Random Forest, providing high-accuracy, high-resolution soil moisture data crucial for agricultural management and drought monitoring in Africa.

Objective

• Develop novel Geographically Weighted Random Forest (GWRF) and Improved Geographically Weighted Random Forest (IGWRF) models for soil moisture (SM) downscaling, accounting for spatial heterogeneity and complex nonlinear relationships.
• Assess the performance of Random Forest (RF), GWRF, and IGWRF models, and validate the accuracy of the downscaled results using in-situ measurements.
• Analyze the spatiotemporal variation patterns of SM in Kenya and quantify the contribution of different environmental variables to SM downscaling, providing guidance for variable selection in future studies.

Study Configuration

• Spatial Scale: Kenya, East Africa. Downscaling from 9 km to 1 km spatial resolution.
• Temporal Scale: Monthly data for the period from January 2020 to December 2021.

Methodology and Data

• Models used: Random Forest (RF), Geographically Weighted Random Forest (GWRF), Improved Geographically Weighted Random Forest (IGWRF).
• Data sources:
  • Satellite: SMAP Enhanced L3 Radiometer product (L3SMP_E) (9 km, 0-5 cm surface SM, daily descending orbit 06:00 AM, composited to monthly). MODIS products (MOD11A1 Land Surface Temperature, MCD12Q1 Land Cover, MCD43A3 Albedo, MOD09A1 Surface Reflectance) used to derive 14 SM related remote sensing indices (SMRSIs). GPM IMERG precipitation (0.1° x 0.1°, 30 min, converted to monthly total, resampled to 1 km).
  • Topographic: Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (30 m, derived slope, aspect, and Topographic Wetness Index (TWI), resampled to 1 km).
  • Soil: SoilGrids system (soil organic carbon, pH, clay, silt, and sand content at 0-5 cm depth, 250 m, resampled to 1 km).
  • Observation: International Soil Moisture Network (ISMN) - Trans-African Hydro-Meteorological Observatory (TAHMO) network (7 stations in Kenya, 10 cm depth, hourly, monthly means at 06:00 AM).

Main Results

• All downscaling methods (RF, GWRF, IGWRF) produced results strongly correlated with the original SMAP SM (Pearson correlation coefficient R > 0.9) and significantly enriched spatial detail and texture.
• In complex terrain, GWRF and IGWRF achieved higher accuracy than RF, with IGWRF showing the greatest consistency with in-situ measurements (R = 0.771).
• Relative to RF, IGWRF increased R by 4.5 % and reduced RMSE and ubRMSE by 7 % and 5.8 %, respectively, demonstrating superior performance. IGWRF also outperformed the original SMAP SM data.
• Soil moisture in Kenya exhibited significant spatiotemporal variability, with higher values in the western region (high elevation, forest/cropland, near Lake Victoria, high clay content) and lower values in the eastern region (plains, sparse vegetation, sandy soil).
• Soil moisture showed clear seasonal fluctuations, increasing significantly during rainy seasons (April-May, November-December) and remaining low during dry seasons (January-March, June-October). SM levels in 2020 were generally higher than in 2021, consistent with precipitation patterns.
• Downscaling factor importance analysis showed Digital Elevation Model (DEM) as the dominant factor, followed by Land Surface Temperature (LST) and Clay content. LST contribution was significantly higher during dry periods, while NDWI and NDDI contributions varied with rainfall. Spatial contributions of factors exhibited clear heterogeneity.

Contributions

• Proposes a novel Improved Geographically Weighted Random Forest (IGWRF) model that integrates local adaptation (GWRF) and global generalization (RF) to simultaneously address spatial heterogeneity and high-dimensional nonlinear relationships in soil moisture downscaling.
• Provides high-resolution (1 km) soil moisture data for Kenya, filling a critical gap for agricultural decision-making, water resource management, and drought monitoring in drought-prone regions of Africa.
• Demonstrates superior performance of IGWRF over traditional RF and GWRF, with quantitative improvements in accuracy metrics (e.g., R, RMSE, ubRMSE) validated against in-situ measurements.
• Analyzes the spatiotemporal variation patterns of soil moisture in Kenya and quantifies the contribution of different environmental variables, offering insights for future variable selection.
• Offers a valuable methodological reference for soil moisture downscaling research in other drought-prone regions globally.

Funding

• National Key Research and Development Program of China (2024YFE0214500)
• International Science and Technology Cooperation Project of Hubei Province, China (2024EHA035)
• CAS-ANSO Sustainable Development Research Project (CAS-ANSO-SDRP-2024–04)
• Sino-Africa Joint Research Center, CAS, China (SAJC202527ZD02)
• Hubei Provincial Science and Technology Innovation Base and Platform Program (2025CSA050)

Citation

@article{Yang2025novel,
  author = {Yang, Yunjie and Wang, Lihui and Zhai, Xu and Zheng, Xiaodi and Zhao, Guosong and Yang, Qichi and Du, Yun and Ling, Feng},
  title = {A novel Improved Geographically Weighted Random Forest (IGWRF) model for low-resolution soil moisture data downscaling in Africa},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.110034},
  url = {https://doi.org/10.1016/j.agwat.2025.110034}
}

Generated by BiblioAssistant using gemini-2.5-flash (Google API)