The thesis describes the development of a rapid and convenient method for the determination of the total antioxidant capacity in green tea, using near-infrared (NIR) spectroscopy and multivariate calibration. NIR is an indirect method and a calibration model therefore needs to be constructed by relating the NIR spectrum of green tea samples with the corresponding antioxidant capacity values, which are measured by a reference method called "Trolox Equivalent Antioxidant Capacity assay". The thesis focuses on two problems that are often encountered in modelling. One is outlier detection. A new method called "Robust principal components regression based on principal sensitivity vectors" is proposed and its outlier detection ability is investigated. The other problem is the selection of an adequate number of components to be included in the partial least square model. A model with a too small number of components may lead to bad prediction since some relevant information is not included whereas a model with a too large number of components may lead to the perfect fitting of the calibration samples but poor prediction for new samples (overfitting). Two novel methods, "Averaged partial least squares" and "Boosting partial least squares", are proposed to solve this problem.