Field evaluation of semi-automated moisture estimation from geophysics using machine learning.

Area of Science:

• Geophysics
• Soil Science
• Machine Learning

Background:

• Geophysical methods offer 3D soil moisture estimates, but direct comparisons with measurements are often poor, limiting their use to qualitative assessments.
• Limitations include the need for accurate models, uncertainties from data processing, and challenges in integrating data from multiple geophysical techniques.
• Accurate soil moisture estimation is crucial for various environmental and agricultural applications.

Purpose of the Study:

• To investigate the limitations of geophysical methods for quantitative soil moisture estimation.
• To evaluate the effectiveness of combining multiple geophysical datasets (GPR, ERT, FDEM) for soil moisture prediction.
• To explore the application of machine learning for improved soil moisture modeling.

Main Methods:

• An irrigation experiment was conducted, monitoring soil moisture and collecting surface geophysical data (GPR, ERT, FDEM) before and after irrigation.
• Geophysical data were processed, gridded, and used with calibration techniques, multivariate regression, and machine learning for soil moisture prediction.
• A random regression forest model was employed, utilizing a combination of inverted ERT, raw FDEM, and inverted FDEM data.

Main Results:

• A machine learning model combining inverted ERT, raw FDEM, and inverted FDEM data achieved high accuracy in predicting soil moisture (RMSE 0.025-0.046 cm³/cm³).
• The model's performance was validated using cross-validation and a separate test dataset, confirming its robustness.
• Machine learning facilitated a semi-automated process for model selection, adaptable to different sites and datasets.

Conclusions:

• Combining multiple geophysical datasets with machine learning significantly improves quantitative soil moisture estimation accuracy.
• The developed approach overcomes limitations of individual geophysical methods and data integration challenges.
• This methodology offers a pathway for developing localized, accurate soil moisture prediction models for diverse applications.