Unsupervised feature extraction techniques of hyperspectral images (HSIs) have recently drawn significant attention for their excellent performance and efficiency in classification. In some existing methods, the denoising process that reduces the influence of inherent noise is ignored, and the nonlinear edge characteristics and multi-scale features that help to classify still need to be fully considered. To solve these issues, we employ a multi-scale nonlinear edge-based unsupervised three-phase model (UTPM) for hyperspectral feature extraction. Specifically, in the initial phase, a noise-adjusted principal components technique is adopted to lower the noise to improve the performance of the proposed model. Then, a neighbor band grouping technique is designed to reduce redundancy and computational cost with information entropy. Because the information entropy can concretely reflect the importance of different bands in the same group, the inner structure can be maximally preserved. Finally, we utilize a multi-scale feature fusion on kernel low-rank entropic analysis to extract nonlinear edge features and combine it with a convolution algorithm to fuse the elements of multiple scales to improve the classification performance. Compared with several other classical or progressive unsupervised hyperspectral feature extraction algorithms, the classification results on three public HSI datasets validate the effectiveness of UTPM.