Methods were explored for the classification of heterogeneous powder mixtures using Fourier transform infrared (FT-IR) hyperspectral image data. The images collected were non-congruent, meaning that samples of the same mixture do not have the same spatial arrangements of their components in their images. In order to classify such images on a one-image-per-object basis, dimension reduction was carried out so as to produce a score or feature vector for each image that preserved information about the heterogeneity of the sample. These feature vectors were then classified using discriminant analysis (DA) or soft independent modeling of class analogy (SIMCA). The most successful approach was the use of a median-interquartile range "super-spectrum" as the feature vector representing each image; using principal component analysis (PCA) DA classification, 87.5% of training samples were correctly classified using leave-one-out cross-validation, and 100% of a test set were correctly classified. This compares with 52.5% and 72%, respectively, when single-point spectra were used to classify the samples.