A method includes decomposing a training set to obtain a principal component matrix having a plurality of principal component vectors. The method also includes variably rejecting portions of a sample spectrum vector that do not correspond to a selected one of the plurality of principal component vectors by incrementally providing a coefficient indicative of the weighting of the selected principal component vector for selected sub-regions. A corrected spectrum vector can be obtained by excluding certain sub-regions of the sample spectrum vector and corresponding principal component vector, multiplying the sample spectrum vector with the principal component matrix for non-excluded sub-regions, providing a predicted interference vector, and subtracting the predicted interference vector from the sample spectrum vector.