Improved Generalized Calibration of an Impedance Probe for Soil Moisture Measurement at Regional Scale Using Bayesian Neural Network and Soil Physical Properties

Regional-scale precise soil moisture measurements are required for remote sensing-based soil moisture product validation besides, complimenting in several hydrological and agricultural applications. Though the gravimetric method provides the most accurate soil moisture measurements, it cannot be extended to the regional-scale due to the large number of sampling requirements. An impedance probe is a suitable substitute for the time-intensive gravimetric method; however, it needs soil/field-specific calibrations for precise measurements. The present study aims to develop a generalized calibration of an impedance probe (i.e., ThetaProbe) for precise measurements of soil moisture at the regional-scale within the root-mean-square-error (RMSE) of 0.04 m(3) m(-3) to fulfil the accuracy requirement of current satellite missions. A few methods for calibrating impedance probe were investigated using 496 gravimetric samples and coincident impedance probe measurements collected over 83 locations through field campaigns in a paddy dominated tropical Indian watershed that covers an area of 500 km(2). The manufacturer generalized calibration was found to have high RMSE (0.0523 m(3) m(-3)) and considerable bias (0.0241 m(3) m(-3)) in soil moisture measurements. Developed generalized and soil-specific calibration based on a linear regression technique that resulted in RMSE values of 0.0468 and 0.0422 m(3) m(-3), respectively. Further, a Bayesian neural network (BNN) based method, a nonlinear technique, was used for developing a generalized calibration of the impedance probe. The results illustrated that BNN-based generalized calibration (RMSE < 0.04 m(3) m(-3)) performs better than the linear regression-based calibrations (RMSE > 0.04 m(3) m(-3)). Moreover, the performance of BNN-based generalized calibration was further improved by the inclusion of soil physical properties as input and yielded an RMSE value up to 0.0352 and 0.0366 m(3) m(-3) during training and cross-validation process, respectively. (C) 2020 American Society of Civil Engineers.