On the basis of the characteristics of multi-band, nonlinear and spatial correlation of hyperspectral remote sensing data, a new feature extraction algorithm based on spatial-spectral deep learning (SSDL) is proposed. This algorithm uses a multiple layers deep learning model, which is the stacked automatic encoder to study high spectral data layer by layer and explore the deep nonlinear characteristics of the image. Based on the spatial neighbor information of each feature pixel, the spatial-spectral combination of sample depth feature and spatial information is used to increase the compactness of homogeneous data and the separability of non-homogeneous data, and improve the performance of subsequent classification. The ground objects classification experiments are performed on Pavia University and Salinas Valley hyperspectral remote sensing datasets. When sample proportion is 1%, the ground objects overall classification accuracy reaches 91.05% and 94.16%. When sample proportion is 5%, the ground objects overall classification accuracy reaches 97.38% and 97.50%. The results show that the SSDL feature extraction algorithm fuses the deep nonlinear characteristics and spatial information of data. It can effectively extract the discriminant features, and obtain higher classification accuracy than other algorithms.