Spectrum Selection and Deep Feature Fusion based Hyperspectral Image Natural Scene Classification Network

Hyperspectral image classification plays an important role in many remote sensing applications. However, the high-dimensional characteristics of hyperspectral images and the appropriate feature representations leave it with great challenges. In this article, these difficulties are addressed by developing a Spectrum Selection and Deep Feature Fusion based method. The proposed method has the following contributions: 1) reducing redundant infounation caused by high-dimension through spectrum selection which is just needed in training phase. 2) extracting the joint spectral-spatial features by deep feature fusion, which effectively improves the accuracy of scene classification. 3) increasing the network attention to scene classes of small number by the Class-Balanced loss function and overcome the influence of unbalanced distribution of experimental data. Experiments results in the Tiangong-1 natural scene images dataset (TG1-NSCD) demonstrate that the effectiveness of our algorithm and the OA is 17% higher than the baseline.