Real-time measurement of soil attributes using on-the-go near infrared reflectance spectroscopy

The spatial variability of soil attributes is cost prohibitive to characterize using traditional soil sampling and laboratory analysis. Yet, the potential benefit of managing soils on a site-specific basis has been recognized. in addition, measurement of terrestrial carbon stocks and their site-specific response to particular management schemes is needed. This paper presents an on-the-go spectrophotometer for in situ measurement of reflectance spectra and evaluates the potential of the system for making real-time predictions of various soil attributes using near infrared reflectance spectroscopy (NIRS). The evaluation was conducted using spectrophotometer data and soil samples from eight fields in central Kansas. For each of the eight fields, a clustering algorithm was used to select 15 sample locations that best represented the spectral data space. Spectral pre-treatments including a derivative and the standard normal variate were considered and calibrations were created using principal components regression (PCR). A one-field-out validation scheme was shown to be a more stringent test than one-sample-out or (1/m)-out, where m is the number of fields represented in the calibration set. Validation using a one-field-out scheme was emphasized because it is identical to the prediction problem encountered in a real-time context. The best one-field-out validation results were obtained for organic matter (OM), which was predicted with a root-mean-square error (RMSE) of 0.52% and a coefficient of determination (R-2) of 0.67. Furthermore, the number of fields used for OM one-field-out validation was varied from 3 up to 8 to test the effect of adding samples from more fields to the calibration set. The results indicate that the prediction accuracy and percentage of locations predicted will increase as fields are added. (C) 2007 Elsevier B.V. All rights reserved.