A Robust Method to Derive and Predict Soil-Specific Calibration Coefficients for Soil Water Content Sensors

We developed an approach to make consistent and comparable soil water content sensor calibrations across a continental-scale network in a production framework that incorporates a thorough accounting of uncertainties. Over one-hundred soil samples with varying characteristics from 27 National Ecological Observatory Network (NEON) sites across the United States were used to generate soil-specific calibration curves for a capacitance-type sensor (Sentek EnviroSCAN TriSCAN). We found the manufacturer’s nominal equation poorly fit the data and resulted in negative (84% of soil samples) and positive (9% of soil samples) systematic biases relative to ‘true’ volumetric water content for the soil types within this study; the mean absolute bias was 0.103 cm³ cm^-3. Use of soil-specific calibration coefficients with the manufacturer’s equation corrected the systematic biases and resulted in a mean, calibration fit uncertainty of ± 0.016 cm³ cm^-3 (± 1 SD), with the tail ends of the calibration comprising the largest percentage of uncertainty. For many soils, the relationship between sensor output and water content was S-shaped, which could not be reproduced by the manufacturer’s equation. Therefore, we developed a new calibration equation to mitigate the uncertainties at the tail ends of the calibration spectrum. Overall, this resulted in a mean, calibration fit uncertainty of ± 0.013 cm³ cm^-3 (± 1 SD).

Using results from the above investigation, we then related the soil-specific calibration coefficients to soil properties to predict soil-specific calibrations for a wide range of soils. We used random forest ensemble machine learning to train models to predict soil-specific calibration coefficients based on soil properties for the manufacturer’s equation and our new equation. We evaluated the ability of the models with a separate testing dataset and found that we could predict soil-specific calibrations with a root mean square error (RMSE) of 0.076 cm³ cm^-3 for our new equation. In contrast, the corresponding RMSE for the manufacturer equation was 0.196 cm³ cm^-3, which further supports the notion that our new equation is superior to the manufacturer’s equation. Our robust calibration method is suitable for adoption by other large soil moisture monitoring networks to derive reliable calibrations with quantified uncertainty across a wide range of soils. Moreover, users of EnviroSCAN sensors could use our models to predict soil-specific calibration coefficients for their site based on a few simple soil properties if they want to avoid the labor-intensive, time-consuming, and costly process of empirically generating soil-specific calibration coefficients.