We have developed an architecture for spectral classification in spectral imaging. Adaptive kernels are designed using probabilistically-weighted principal component analysis. Simulation predicts orders-of-magnitude reduction in error rates, and a prototype system is currently under construction. © 2012 OSA.