The traditional Dimensionality Reduction (DR) methods consider the spectral features but ignores useful spatial information in HSI. To overcome this problem, this paper proposed a new dimensionality reduction method called Multi-Feature Manifold Discriminant Embedding (MFDE). First, the MFDE method extracted the features of the local binary pattern from HSI data. Next, the with-class and between-class graphs were constructed using sample labels to exploit the local manifold structure. Then, an optimal object function was designed to learn the combined spatial-spectral features by compacting the intra-class samples and simultaneously separating the inter-class samples. Thus, the discriminative ability of embedding features was improved. Experimental results in the Indian Pines and Heihe hyperspectral data sets show that the proposed MFDE method performs better than some state-of-the-art DR methods in most cases and achieves an overall classification accuracy of 95.05% and 96.20%, respectively. Its advantage is more significant for less training samples, making it more conducive to practical applications.