Abstract: Scene classification is not an easy task owing to the variability, ambiguity, and the wide range of illumination and scale conditions the scenes may apply. Since feature extraction and scene representation play important roles in classification tasks, this paper presents an approach for unsupervised feature learning based on independent subspace analysis. The proposed method could automatically learn structural feature bases organized in a grouped fashion from randomly sampled natural image patches in independent subspaces. Optimization process of feature bases is implemented under an incremental learning framework to cope with the learning difficulty with large or dynamic samples. Patch-based image descriptors are computed over regularly divided grids using nonlinear combination coefficients of the learned feature bases. These descriptors are then taken into the spatial pyramid matching model, which incorporates spatial layout information and global geometric correspondence for recognizing scene categories, to build hierarchical scene representations. Experiment reveals how the related parameters influence objective optimization process and the final classification performance. Compared with several typical models in classification task on OT scene dataset, the proposed method could form low-dimensional but efficient image patch descriptors and achieve high classification accuracy with stability.