Lamichhane et al. (2025) Multi‐layer root zone soil moisture estimation using field and remote sensing data fusion with machine learning in semi‐arid croplands

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Vadose Zone Journal
• Year: 2025
• Authors: Manoj Lamichhane, Sushant Mehan, Kyle R. Mankin
• DOI: 10.1002/vzj2.70047

Short Summary

This study developed an Extreme Gradient Boosting model integrating PlanetScope optical data, climate variables, and soil properties to estimate multi-layer soil moisture (SM) down to 1.8 m at 3 m spatial resolution, achieving high accuracy ($R^2$ up to 0.89) and demonstrating that incorporating SM from the adjacent upper layer significantly improves deep SM prediction.

Objective

• Estimate multi-layer soil moisture (SM) in 0.3 m increments down to 1.8 m depth by integrating multimodal remote sensing and sub-field data using machine learning algorithms.

Study Configuration

• Spatial Scale: Sub-field scale, high spatial resolution (3 m).
• Temporal Scale: Single-period estimation based on input feature availability (e.g., climate variables, soil properties).

Methodology and Data

• Models used: Extreme Gradient Boosting (XGBoost).
• Data sources: PlanetScope (optical sensor), in situ soil properties (texture, organic matter, soil carbon), climatic variables (precipitation, reference evapotranspiration), soil temperature.
• Modeling Approaches: (1) Using only input features (RS, climate, soil properties); (2) Adding predicted SM from the adjacent upper layer to (1).

Main Results

• Model performance across six depths (0 m to 1.8 m) yielded $R^2$ values ranging from 0.78 to 0.89.
• The Root Mean Squared Error (RMSE) for SM predictions varied between $0.021 \text{ m}^3/\text{m}^3$ and $0.028 \text{ m}^3/\text{m}^3$, with the Ratio of RMSE to observed mean (RRMSE) ranging from 11.7% to 14.6%.
• Incorporating predicted SM from the adjacent upper layer significantly enhanced performance, increasing $R^2$ by 8%–24% and reducing RMSE by 10%–27% across various depths.
• Surface SM estimation was primarily influenced by remote sensing bands (blue, near-infrared), vegetation indices (NDVI green, VARI), and climatic variables, while deeper SM estimation relied heavily on soil properties (texture, organic matter, soil carbon).
• A simplified model excluding soil property input features still captured considerable root zone SM variability ($R^2$ of 0.66–0.77).

Contributions

• Development of a robust machine learning methodology to generate high-resolution (3 m), multi-layer (up to 1.8 m) soil moisture maps.
• Demonstrated the critical role of adjacent layer soil moisture information in improving the prediction accuracy of deeper soil moisture layers.
• Provided a practical alternative for root zone soil moisture analysis applications where detailed soil property data is limited or unavailable.

Funding

• Not specified in the abstract.

Citation

@article{Lamichhane2025Multilayer,
  author = {Lamichhane, Manoj and Mehan, Sushant and Mankin, Kyle R.},
  title = {Multi‐layer root zone soil moisture estimation using field and remote sensing data fusion with machine learning in semi‐arid croplands},
  journal = {Vadose Zone Journal},
  year = {2025},
  doi = {10.1002/vzj2.70047},
  url = {https://doi.org/10.1002/vzj2.70047}
}

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