A pragmatic, automated approach for retroactive calibration of soil moisture sensors using a two-step, soil-specific correction

作者:Gasch Caley K; Brown David J; Brooks Erin S; Yourek Matt; Poggio Matteo; Cobos Douglas R; Campbell Colin S
来源:Computers and Electronics in Agriculture, 2017, 137: 29-40.
DOI:10.1016/j.compag.2017.03.018

摘要

Soil moisture sensors are increasingly deployed in sensor networks for both agronomic research and precision agriculture. Soil-specific calibration improves the accuracy of soil water content sensors, but laboratory calibration of individual sensors is not practical for networks installed across heterogeneous settings. Using daily water content readings collected from a sensor network (42 locations x 5 depths = 210 sensors) installed at the Cook Agronomy Farm (CAF) near Pullman, Washington, we developed an automated calibration approach that can be applied to individual sensors after installation. As a first step, we converted sensor-based estimates of apparent dielectric permittivity to volumetric water content using three different calibration equations (Topp equation, CAF laboratory calibration, and the complex refractive index model, or CRIM). In a second, "re-calibration" step, we used two pedotransfer functions based upon particle size fractions and/or bulk density to estimate water content at wilting point, field capacity, and saturation at each sensor insertion point. Using an automated routine, we extracted the same three reference points, when present, from each sensor's record, and then bias corrected and re-scaled the sensor data to match the estimated reference points. Based on validation with field-collected cores, the Topp equation provided the most accurate calibration with an RMSE of 0.074 m(3) m(-3), but automated re-calibration with a local pedotransfer function outperformed any of the calibrations alone, yielding a network-wide RMSE of 0.055 m(3) m(-3). The initial calibration equation used in the first step was irrelevant when the re-calibration was applied. After correcting for the reference core measurement error of 0.026 m(3) m(-3) used for calibration and validation, the error of the sensors alone (RMSEadj) was computed as 0.049 m(3) m(-3). Sixty-five percent of individual sensors exhibited re calibration errors less than or equal to the network RMSEadj. The incorporation of soil physical information at sensor installation sites, applied retroactively via an automated routine to in situ soil water content sensors, substantially improved network sensor accuracy.

  • 出版日期2017-5