Both supervised and semisupervised algorithms for hyperspectral data analysis typically assume that all unlabeled data belong to the same set of land-cover classes that is represented by labeled data. This is not true in general, however, since there may be new classes in the unexplored regions within an image or in areas that are geographically near but topographically distinct. This problem is more likely to occur when one attempts to build classifiers that cover wider areas; such classifiers also need to address spatial variations in acquired spectral signatures if they are to be accurate and robust. This paper presents a semisupervised spatially adaptive mixture model (SESSAMM) to identify land covers from hyperspectral images in the presence of previously unknown land-cover classes and spatial variation of spectral responses. SESSAMM uses a nonparametric Bayesian framework to apply spatially adaptive mechanisms to the mixture model with (potentially) infinitely many components. In this method, each component in the mixture has spatially adapted parameters estimated by Gaussian process regression, and spatial correlations between indicator variables are also considered. The proposed SESSAMM algorithm is applied to hyperspectral data from Botswana and from the DC Mall, where some classes are present only in the unlabeled data. SESSAMM successfully differentiates unlabeled instances of previously known classes from unknown classes and provides better results than the standard Dirichlet process mixture model and other alternatives.