Deep Learning Integrated with Multiscale Pixel and Object Features for Hyperspectral Image Classification

The spectral resolution and spatial resolution of hyperspectral images are continuously improving, providing rich information for interpreting remote sensing image. How to improve the image classification accuracy has become the focus of many studies. Recently, Deep learning is capable to extract discriminating high-level abstract features for image classification task, and some interesting results have been acquired in image processing. However, when deep learning is applied to the classification of hyperspectral remote sensing images, the spectral-based classification method is short of spatial and scale information; the image patch-based classification method ignores the rich spectral information provided by hyperspectral images. In this study, a multi-scale feature fusion hyperspectral image classification method based on deep learning was proposed. Firstly, multiscale features were obtained by multi-scale segmentation. Then multiscale features were input into the convolution neural network to extract high-level features. Finally, the high-level features were used for classification. Experimental results show that the classification results of the fusion multi-scale features are better than the single-scale features and regional feature classification results.