A new approach for neutron moisture meter calibration: artificial neural network

Abstract

The neutron moisture meter (NMM) is a widely used device for sensing soil water content (SWC). Calibration accuracy and precision of the NMM are critical to obtain reliable results, and linear regression analysis of SWC against NMM count data is the most common method of calibration. In this study, artificial neural network (ANN) calibration models were developed and compared with linear regression. For this purposes, training and validation data were obtained from 2 calibration and 16 testing plots, respectively. Calibration plots consist of wet and dry soil water conditions separately. Data measured in dry beans and red pepper plots that have four different water levels were used to determine validity of regression and ANN-based calibration models. Volumetric SWC and NMM count ratio measurements were taken for depth intervals of 30 cm throughout a 120-cm-deep soil profile. Several neural network architectures were explored in order to determine the optimal network architecture. Data analyses were conducted for each soil layer and for the whole profile, separately, based on both linear regression and ANN. Linear regression calibration equation coefficients of determination (r (2)) for the 0-30, 30-60, 60-90 and 90-120 cm depth ranges calculated by regression models were 0.85, 0.84, 0.72 and 0.82, respectively, and r (2) values were 0.94, 0.95, 0.87 and 0.88 based on ANN models, respectively. Using the data set from the entire 120-cm soil profile for calibration by ANN, the r (2) value was raised to 0.97.