Adaptive Spatial-Spectral Feature Learning for Hyperspectral Image Classification

Abstract

The combination of spectral and spatial information provides an effective way to improve the hyperspectral image (HSI) classification. However, the local spatial contexture changes with different neighborhood regions over the HSI image plane, and methods with fixed weights to integrate spatial information for all neighborhood regions could result in inaccurate spatial features, leading to adverse effects on classification performance. To address this issue, a novel adaptive spatial-spectral feature learning network (ASSFL) has been proposed to reflect spatial contexture changes and learn robust adaptive features in this paper. In the implementation of the proposed method, a convolution neural network (CNN) is first applied to learn weight features for each pixel within a hyperspectral patch and adaptive weights can be obtained based on a softmax normalization. Then, the shallow joint adaptive features can be acquired according to these weights. After that, a stacked auto-encoder (SAE) is proposed to further extract deeper hierarchical features for the final classification. The experimental results on four benchmark HSI data sets demonstrate that the proposed method can achieve competitive classification results compared with other existing classifiers.