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

Abstract

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 m3 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 km2. The manufacturer generalized calibration was found to have high RMSE (0.0523 m3 m−3) and considerable bias (0.0241 m3 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 m3 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 m3 m−3) performs better than the linear regression–based calibrations (RMSE >0.04 m3 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 m3 m−3 during training and cross-validation process, respectively.