Tramblay et al. (2022) Estimating soil moisture conditions for drought monitoring with random forests and a simple soil moisture accounting scheme

Identification

• Journal: Natural hazards and earth system sciences
• Year: 2022
• Authors: Yves Tramblay, Pere Quintana Seguí
• DOI: 10.5194/nhess-22-1325-2022

Research Groups

• HSM, University of Montpellier, CNRS, IRD, IMT, Montpellier, France (Yves Tramblay)
• Observatori de l'Ebre (OE), Ramon Llull University, CSIC, Roquetes, Spain (Pere Quintana Seguí)

Short Summary

This study regionalized a simple Soil Moisture Accounting (SMA) model over the Iberian Peninsula for drought monitoring by estimating the soil water holding capacity parameter ($A$). It found that using Random Forests (RF) with climatic and physiographic covariates provided a more robust and stable estimation of $A$, particularly for detecting dry soil conditions, compared to direct estimation using European soil maps.

Objective

• To regionalize a simple Soil Moisture Accounting (SMA) scheme for daily soil moisture simulation and drought monitoring across the Iberian Peninsula.
• To compare two methods for estimating the SMA model's single parameter (maximum soil water holding capacity, $A$): direct estimation using pedotransfer functions from soil maps (European Soil Database, ESDB) versus indirect estimation using a machine learning approach (Random Forests, RF) based on climatic and physiographic covariates.

Study Configuration

• Spatial Scale: Iberian Peninsula (Spain and Portugal). Gridded data at 5 km $\times$ 5 km resolution.
• Temporal Scale: Daily time steps.

Methodology and Data

• Models used:
  • Simple Soil Moisture Accounting (SMA) model (conceptual hydrological model, based on the GR4J structure).
  • SURFEX Land-Surface Model (LSM) using the ISBA-DIF scheme (used as the reference/benchmark for soil moisture simulation).
  • Random Forests (RF) machine learning algorithm (used for regionalizing the SMA parameter $A$).
• Data sources:
  • Reference Soil Moisture: Simulated soil moisture (0–60 cm depth) from the SURFEX LSM run over the Iberian Peninsula.
  • Meteorological Forcing: SAFRAN-Spain database (daily precipitation, temperature, and potential evapotranspiration (PET)).
  • Soil Data (for direct estimation): European Soil Database (ESDB), providing Total Available Water Content (TAWC).
  • Covariates (for RF estimation): Altitude, mean annual precipitation, temperature, PET, and land cover (from ECOCLIMAP2).
  • Validation: Comparison of the two regionalization methods using verification scores (Probability of Detection (POD), False-Alarm Ratio (FAR), and Heidke Skill Score (HSS)) focused on detecting daily soil moisture below the 10th percentile (dry conditions).

Main Results

• The calibrated SMA model accurately reproduced the SURFEX soil moisture dynamics, achieving a high mean Nash coefficient of 0.94 across the domain.
• The Random Forest model successfully estimated the parameter $A$ with high robustness (low loss of performance in out-of-bag estimates).
• Predictor Importance (RF): Precipitation and Potential Evapotranspiration (PET) were the most influential predictors of soil water holding capacity ($A$), followed by altitude. Land cover was the least important.
• Overall Performance Comparison: The SMA model using $A$ estimated by RF achieved a mean Nash value of 0.86 (median 0.89) on the testing set, slightly outperforming the ESDB approach (mean Nash 0.81, median 0.85).
• Drought Detection Performance: The RF estimation method showed more stable and robust results for detecting dry soil conditions (below the 10th percentile). RF HSS scores were consistently above 0.4, while ESDB scores dropped close to zero in some grid cells, indicating less reliable detection in certain regions.

Contributions

• Developed and validated an efficient, simple, and regionalizable methodology (SMA + RF) for high-resolution daily soil moisture monitoring, suitable for regions lacking dense in situ networks or high-resolution LSM implementation.
• Demonstrated that machine learning (Random Forests) provides a more robust regionalization of the soil water holding capacity parameter ($A$) compared to traditional pedotransfer functions derived from high-resolution soil maps (ESDB), especially for drought monitoring applications.
• Established that the methodology relies on covariates (climate and elevation) that can be reliably estimated from global databases, making it highly transferable to data-sparse regions (e.g., Northern Africa, Italy, Greece).

Funding

• Ministerio de Ciencia, Innovación y Universidades (HUMID project; grant no. CGL2017-85687-R, AEI/FEDER, UE).
• ANR HILIAISE projects (contribution acknowledged).

Citation

@article{Tramblay2022Estimating,
  author = {Tramblay, Yves and Quintana‐Seguí, Pere},
  title = {Estimating soil moisture conditions for drought monitoring with random forests and a simple soil moisture accounting scheme},
  journal = {Natural hazards and earth system sciences},
  year = {2022},
  doi = {10.5194/nhess-22-1325-2022},
  url = {https://doi.org/10.5194/nhess-22-1325-2022}
}

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