USING NEAR-INFRARED REFLECTANCE PRODUCT LIBRARY FILES TO IMPROVE PREDICTION ACCURACY AND REDUCE CALIBRATION COSTS

Deriving a near infrared reflectance spectroscopy (NIRS) calibration f or new samples is an expensive and time-consuming process. Often a lab oratory analyzing samples with NIRS will receive new samples that are spectrally different from the library samples for that product. This s tudy was performed to develop a method to: (i) identify previously ana lyzed samples in a product library that best match the spectra of the new samples for a local calibration, (ii) compare the prediction accur acy of the broad product library calibration to the local calibration, and (iii) if necessary, expand the local calibrations with a few of t he new samples when greater prediction accuracy is required. Three gro ups of new samples were obtained for the study. They were 68 hay sampl es, 106 grass samples, and 110 whole-plant corn (Zea mays L.) samples. Every third sample was reserved for validation. A product library fil e of 2176 hay, haylage, and fresh forage spectra supplied the product calibration and similar spectra for the hay and grass samples, and a c orn silage product library file of 309 spectra supplied the product ca libration and similar spectra for the whole-plant corn samples. The sp ectra in the library most similar to the new samples were identified a nd selected with a new program, MATCH. Local calibrations were develop ed from these selected samples. Calibration accuracy was tested with t he new samples reserved for validation using equations developed from the product library. In 7 of the 12 comparisons, the local calibration s were more accurate than the broad-product library calibrations, but only two of these calibrations were acceptable. By expanding the local calibrations with 10 samples of each new group and recalibrating, acc uracy of the expanded calibrations was similar to the accuracy of cust om calibrations derived from the new samples not reserved for validati on. Only 30 samples out of 284 needed new reference values to obtain a cceptable prediction accuracy. This would have resulted in a cost redu ction of 89% in reference value analysis. Guidelines for using this pr ocedure are presented.