Double Attention Based Multilevel One-Dimensional Convolution Neural Network for Hyperspectral Image Classification

Abstract

The large spectral variability and nonlinearity of hyperspectral images (HSIs) make classification a challenging task. Hence, the powerful capacities for feature extraction and nonlinear representation provided by deep learning have made it a widely-used tool for HSI classification. However, most of these methods involve complex spectral-spatial networks with a mass of parameters to achieve high-performance results, which are limited by the small number of training samples and large computational costs in practice. Simple networks harnessing spectral information alone effectively alleviate these problems, but these approaches generally underutilize channel correlations and hierarchical complementarity, which degrades discriminability and lowers performance. To solve these problems, a novel end-to-end double attention based multilevel one-dimensional convolution neural network (DAMN) is proposed, that fully exploits the rich spectral information contained in HSIs to boost an accurate and efficient classification with a limited number of samples. An effective multilevel one-dimensional convolution neural network is constructed to hierarchically mine local-channel correlations and comprehensively utilize shallow and deep features. Then, a multilayer perceptron attention module is utilized to extract informative features from raw spectral signatures as network feeds, with an advanced ultra-lightweight subspace attention module employed to learn complementary cross-channel dependencies to further enhance the discriminability of features. The effectiveness of DAMN was validated on three frequently-used hyperspectral datasets, with the results showing that DAMN outperforms the other state-of-the-art spectral classifiers and can be a competitive tool in practical applications.