Comparison between Random Forests, Artificial Neural Networks and Gradient Boosted Machines Methods of On-Line Vis-NIR Spectroscopy Measurements of Soil Total Nitrogen and Total Carbon.

摘要: Accurate and detailed spatial soil information about within-field variability is essential for variable-rate applications of farm resources. Soil total nitrogen (TN) carbon (TC) are important fertility parameters that can be measured with on-line (mobile) visible near infrared (vis-NIR) spectroscopy. This study compares the performance local scale calibrations those based on spiking selected samples from both fields into an European dataset TN TC estimation using three modelling techniques, namely gradient boosted machines (GBM), artificial neural networks (ANNs) random forests (RF). The measurements were carried out a mobile, fiber type, vis-NIR spectrophotometer (305-2200 nm) (AgroSpec tec5, Germany), during which spectra recorded in diffuse reflectance mode two UK. After pre-processing, entire datasets then divided calibration (75%) prediction (25%) sets, models developed GBM, ANN RF leave-one-out cross-validation. Results cross-validation showed effect collected field when combined has resulted highest coefficients determination (R²) values 0.97 0.98, lowest root mean square error (RMSE) 0.01% 0.10%, residual deviations (RPD) 5.58 7.54, TC, respectively. laboratory predictions generally followed same trend as one field, where spiked dataset-based outperformed corresponding GBM models. In second replaced being best performing. However, provided lower R² RPD most cases. Therefore, cost-effective point view, it recommended to adopt RF/ANN successful under measurement conditions.