435Crambe moisture, protein and oil percentages were predicted by fixed-filter near-infrared reflectance with standard errors of prediction (SEP) of 0.26, 0.59 and 0.86 percentage points, respectively. Crambe had a large range of protein and oil percentages, 17.4%-25.0% and 17.7%-36.4% respectively. Calibration samples were selected on the basis of relative spectral data, with no advance knowledge of protein and oil content. This procedure selected samples representing the full range of constituent values, and resulted in calibrations that had lower SEP's than standard errors of calibration.is regarded as a promising new industrial oil seed. Crambe oil is about 55-60% erucic acid (cis-13docosenoic) (1). When treated with ozone, erucic acid gives brassylic and pelargonic acids. Brassylic acid can be used for polyesters, plasticizers, alkyd resins, lubricants, rubber additives and surface-active agents. Pelargonic acid is used for plasticizers, alkyd resins, vinyl stabilizers, hydrotropic salts, pharmaceuticals, synthetic flavors and odors, flotation agents and insect repellents (2).The oil-free, hull-free meal contains 46%-58% crude protein with a good amino-acid balance for feed. The meal will contain thioglucosides, which places some restrictions on its use as a feedstuff.Agronomic research in crambe breeding gave rise to a need for nutrient analysis of crambe seed. Previous success with soybeans (3) and other oilseeds (4) suggested that near-infrared reflectance (NIR) could be used in lieu of wet chemical methods. However, crambe seed is light and small, conditions certain to complicate grinding.The multiple-linear-regression calibration of fixedwavelength near-infrared reflectance analyzers has been reported extensively in the literature. Little has been written about the actual selection process of calibration and prediction samples. Typically, a number of samples are collected and randomly divided into two groups--one for calibration and one for prediction (5,6), or samples are compared against those of previously known chemical composition (4,7). Barton and Cavanagh {8) used two stepwise linear-regression analysis programs, which allowed every i th sample to be saved for prediction verification.Selection of calibration samples by spectra alone has been reported for near-infrared monochramotors (9) but not for fixed-filter units.The range of protein and oil values for crambe was unknown. Instead of running wet chemistry on all our available samples, we wanted to select a spectrally representative calibration set by using only reflectance *To whom correspondence should be addressed.