Increasing the Number of Calibration Samples in DSM by In-situ Vis-NIR Spectroscopy

Abstract

Remote and proximal sensors allow for the collection of data at a high resolution and to low costs, but for the approach to be cost-effective the number of calibration samples needs to be kept low. On the other hand, too few calibration samples can lead to unstable calibration models. In a small study on three adjacent fields (55 ha) at a farm in southwest Sweden, in-situ vis-NIR spectroscopy was used to increase the number of calibration samples used in a multiple adaptive regression spline (MARS) model for mapping clay and sand content. The present study did not find support for an improvement of MARS models when the number of calibration samples was increased from 20 to 100 by vis-NIR predictions of the 80 extra samples. This was probably because the 20 soil samples carried enough information to calibrate the exhaustive predictor data to sand and clay and all the extra samples did was only introduced noise. There were some indications of more stable models when the number of reference samples was reduced to 10 or when the best single predictor was excluded. In this study the number of calibration samples seemed to be less critical than their accuracy.