Neural network models for predicting organic matter content in Saskatchewan soils

Neural networks were developed to predict soil organic carbon content using reflectance data as inputs. Networks were subsequently evaluated through validation testing, and their predictive performance was compared to that of previously developed multiple-linearregression (MLR) models. Neural networks tended to outperform MLR models even though the inputs were selected based on optimum linear model performance. Wavelengths of input reflectance data were the same for the best network and the best MLR model in three of the five fields tested. Investigation into determining optimum inputs for nonlinear network development is recommended.