Improved accuracy in the retrieval of crop biophysical variables is needed to enable a greater contribution of hyperspectral remote sensing data to site-specific crop management. One season of maize and two seasons of wheat field experiments were used to explore the potential of multivariate data analysis for retrieving crop biophysical variables from spectroscopic data. Canopy spectral data at 350-2500 nm were collected during the experiments, in which various seeding densities, fertilization, and irrigation treatments were applied to generate dry biomass, water-content and nitrogen-content variability. Partial least squares (PLS) models that considered the reflectance derivatives (1st and 2nd) and that used only significant wavelengths were evaluated. As the measurements were conducted throughout the whole season, a wide variability was observed, which was critical for obtaining a good calibration model. The use of derivatives and selection of the most significant wavelengths were found to be the best pre-processing methodologies to increase prediction accuracy. Significant predictive power was achieved for the validation dataset, especially for the wheat dry biomass (R2 ~0.75), for which similar results were obtained, even with data from a different season (R2 ~0.70). PLS-predicted wheat water content had a correlation of R2 ~0.60 with the measured values. An advantage was found in the use of PLS models, compared to common vegetation indices. Based on ground spectral measurements, this study confirms the potential of multivariate-data-analysis procedures for the interpretation of hyperspectral remote sensing data.