Premium
The Legume Content in Multispecies Mixtures as Estimated with Near Infrared Reflectance Spectroscopy Method Validation
Author(s) -
Locher F.,
Heuwinkel H.,
Gutser R.,
Schmidhalter U.
Publication year - 2005
Publication title -
agronomy journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.752
H-Index - 131
eISSN - 1435-0645
pISSN - 0002-1962
DOI - 10.2134/agronj2005.0018
Subject(s) - calibration , near infrared reflectance spectroscopy , monoculture , mathematics , statistics , standard error , near infrared spectroscopy , mean squared error , content (measure theory) , environmental science , analytical chemistry (journal) , chemistry , agronomy , biology , chromatography , optics , physics , mathematical analysis
The legume content in multispecies mixtures can accurately be estimated by means of near infrared reflectance spectroscopy (NIRS) provided that there is a valid calibration. This study was conducted (i) to test the applicability of two narrow‐based calibrations, A and B (origin of the calibration sets: one farm, five harvests), for plant material not present in the calibration set and (ii) to compare the predictive ability of both models to a third, broad‐based calibration, C (different farms and harvests). The prediction accuracy of the NIRS models, A (end‐points calibration) and B (calibration with incremental standards), was tested with defined legume–grass mixtures of nine test sets, which differed in origin and harvesting dates. Most of the test‐set spectra were later used as additional calibration samples for Model C. Calibrations A and B, which differed in calibration design, showed varying root mean square errors of prediction for each of the nine test sets (A: 3.3–12.5%; B: 3.9–10.3%). The slope of the line from regression of NIRS predicted on true values ranged between 0.93 and 1.09 ( r 2 always > 0.92). All predictions were precise but biased. Prediction accuracy was worst for samples mixed of plant material that was grown in monoculture. After a bias correction and the exclusion of the mixtures of monocultural samples from the test sets, both models showed the same error (standard error of prediction after bias correction < 6%). Calibrations A, B, and C were compared by predicting two external test sets. The broad‐based Calibration C offered the same prediction accuracy as Calibrations A and B but did not reduce the bias. With Model C, there was no consistent reduction in the proportion of outliers in the test sets with a Mahalanobis distance > 3 compared with the Models A and B. It is concluded that predictions are biased even if more natural variability is included in the calibration set. Therefore, based on these data, a bias correction is always necessary. After a bias correction, all calibrations offered a highly accurate tool to estimate the legume content of mixtures independent of origin and harvesting dates. Contrary to our expectations, the most simple model, A, which was derived from the least number of calibration spectra and which represented the least natural variability, proved to be as accurate and robust in the prediction of independent test sets as the broader models, B and C.