The Effect of Region of Interest Size on Model Calibration for Soil Organic Carbon Prediction from Hyperspectral Images of Prepared Soils

Abstract

Hyperspectral imaging is an attractive technique for soil analysis that provides both spectral and spatial information in a three-dimensional image. Scanning a larger sample area than that permitted in soil spectroscopy allows a larger spatial area to be selected to represent the “average spectra” by means of an interactive region of interest (ROI) tool. The objective of this study was to assess the effect of ROI size on the prediction accuracy of soil organic carbon (SOC) for homogenised soils from a diverse dataset collected on a national scale. Five ROI sizes, 72 × 72 pixels, 54 × 54 pixels, 36 × 36 pixels, 18 × 18 pixels and 7 × 7 pixels were selected in the near infrared (NIR) region and partial least square calibrations were developed for each ROI size and compared. Cross-validation results demonstrated that increasing the area of sample considered for partial least square regression modelling improved SOC accuracy. Increasing the dimensions of a ROI size by 100-fold reduced root mean square error of cross-validation from 4.60% to 3.88% SOC, a 16% improvement, while R2 increased from 0.62 to 0.75. Independently validated models showed further improved accuracy whereby root mean square error of prediction was reduced to 3.49% SOC overall, comparable to that reported elsewhere for geographically diverse samples. The spectral variables contributing to the SOC prediction ( P<0.05) compared for each ROI size showed that the 7 × 7 pixels ROI could not differentiate between important and unimportant variables indicating loss of information at this spatial scale.