Performance of European artificial neural network (ANN) calibrations for moisture and protein in cereals using the Danish near-infrared transmission (NIT) network

Three problems need to be addressed in networks of Infratec Grain Analysers: 1) the networks are not interconnected, 2) the partial least squares (PLS) calibrations used so far have to be individually adjusted for bias when transferred to the slave instruments, and 3) the calibrations are not entirely stable over time. Nonlinear artificial neural network (ANN) calibrations based on a large common European data set (approximate to4,000 samples in the training sets and approximate to1,000 samples in the stop sets) were introduced to overcome these constraints. The performance of these ANN calibrations was compared with Danish PLS models for protein and moisture in cereals during the 1998 harvest in Denmark, and subsequently with PLS models based on the same European data set. ANN models were more accurate than PLS and, unlike PLS, were linear and transferable up to 25% moisture. It is suggested that the improved performance of the ANN models is attributable to the modeling technique rather than the size and nature of the European data set. In most cases, ANN models could he applied directly and without bias adjustment to slave instruments. The ANN models were also more stable, they required fewer bias adjustments or remodeling over Lime compared with Danish PLS models. ANN calibrations using shared data have been adopted for commercial use in several European countries and work is in progress to develop global ANN, models for determination of protein in wheat and barley.