A new deep learning approach for classification of hyperspectral images: feature and decision level fusion of spectral and spatial features in multiscale CNN

Abstract

Classification is the main field of hyperspectral data processing. To date, many methods are introduced to increase the accuracy of image classification. In recent years, various convolutional neural network models are proposed for hyperspectral image classification. This study puts forward a multiscale structure of convolutional neural networks that use several patches of different sizes to extract complex spatial features. Due to spatial features' effectiveness in improving the classification accuracy of hyperspectral images, the proposed framework integrates spatial features of three methods; morphological profiles, Gabor filter, and local binary pattern with spectral features at both the feature-level and decision-level. The experiments on three hyperspectral images, Indian Pine, Pavia University, and NCALM demonstrate the proposed method's efficiency. The final results show that the proposed method's overall classification accuracy is 6% higher than some other recent techniques.