Hyperspectral Image Classification Based on Convolutional Neural Network Embedded with Attention Mechanism and Shadow Enhancement by Dynamic Stochastic Resonance

Abstract

Hyperspectral images (HSIs) classification based on deep learning networks which can effectively extract the feature information in the image promotes the exploitation of rich information contained in remote sensing image. However, all features are considered to be equally important in most classification networks and the value of some key features that affect the classification accuracy greatly has not been fully utilized. Attention mechanism can select the information that is more critical to the current task goal from a large number of information. In this paper, a three-dimensional convolutional neural network (3D CNN) embedded with convolutional block attention module(CBAM) has been proposed to the classification of HSIs, meanwhile dynamic stochastic resonance (DSR) has been introduced to enhance the shadow in HSIs to further improve the extraction of implicit feature information in shadow. Firstly, DSR is used to increase the target information in the shadow area. Next, a 3D CNN model is constructed to extract the spatial and spectral features. Then, to realize the potential values of different features, CBAM is embeded into the 3D CNN to recalibrate the extracted features by giving different weights due to their importance to the classification. Finally, the proposed method can help to classify the enhanced HSIs. Comparative studies have been carried out on a real-world HSI. The experimental results show that the proposed approach has promising prospects in the field of HSI classification.