DeepLab-Based Spatial Feature Extraction for Hyperspectral Image Classification

Abstract

Recently, deep learning has been used for hyperspectral image classification (HSIC) due to its powerful feature learning and classification ability. In this letter, a novel deep learning-based framework based on DeepLab is proposed for HSIC. Inspired by the excellent performance of DeepLab in semantic segmentation, the proposed framework applies DeepLab to excavate spatial features of the hyperspectral image (HSI) pixel to pixel. It breaks through the limitation of patch-wise feature learning in the most of existing deep learning methods used in HSIC. More importantly, it can extract features at multiple scales and effectively avoid the reduction of spatial resolution. Furthermore, to improve the HSIC performance, the spatial features extracted by DeepLab and the spectral features are fused by a weighted fusion method, then the fused features are input into support vector machine for final classification. Experimental results on two public HSI data sets demonstrate that the proposed framework outperformed the traditional methods and the existing deep learning-based methods, especially for small-scale classes.