Based on the characteristics of multi-band, inter-band information redundancy and spatial information correlation of hyperspectral images, a spatially-correlated and semi-supervised feature extraction (SCSSFE) algorithm with locality preserving projection (LPP) is proposed. This algorithm defines a new pixel weight calculation function for the different spectral characteristics with the same objects and the different objects with the same spectral characteristics to preserve the spatial distance and the spectral similarity of hyperspectral image by means of the neighbor structure in image space and the intra-class and inter-class discriminant weights. Then, the features of hyperspectral images are extracted by the weight function combined with LPP. Thus the similarity among the same objects and the discrepancy among the different objects are maximized. The proposed LPP-SCSSFE algorithm is verified through the hyperspectral image classification experiments on the two datasets of Indian Pines and Pavia University. The highest overall classification accuracies of the LPP-SCSSFE algorithm reach 87.50% and 91.29% for the respective datasets, better than those of the existing feature extraction algorithms. These results indicate that the spatial correlation and the spectral similarity of hyperspectral images are fully taken into account in the proposed algorithm, and thus the more representative features are extracted and the classification accuracy is enhanced.