Hyperspectral image classification based on spectral and spatial information using ResNet with channel attention

Classification of hyperspectral image (HSI) is widely used for the study of remotely sensed images. Convolutional Neural Networks (CNNs) are one of the most commonly chosen deep learning algorithms for visual data analysis. The HSI classification framework based on the CNN is presented in this paper. Since the imbalance between the high dimension of HSI input data and the limited amount of labeled training data would induce overfitting, current convolutional networks are fairly superficial for HSI classification. To stop the limited efficiency of feature learning, a new HSI classification network called Residual Spectral Spatial-Channel Attention Network (RSS-CAN) is proposed. By utilizing the shortcut connection framework, RSS-CAN can use deeper layers to extract more succinct and efficient features. Furthermore, attention mechanism is used to emphasize meaningful features. In addition, we revised an HSI dataset called Shandong Feicheng. The resolution and pixel quantity of this dataset are significantly greater. In order to check its variety, it has been contrasted with state-of-the-art approaches. Experimental results with widely used hyperspectral image datasets demonstrate that, our proposed method has achieved better performance in comparison with state-of-the-art classifiers and conventional deep learning-based classifiers.

Automated summary

The paper introduces a HSI classification framework based on CNN, which addresses the imbalance between limited labeled training data and high dimension input data by proposing the RSS-CAN network to enhance efficiency in feature learning through shortcut connections and attention mechanism. The modified HSI dataset called Shandong Feicheng was compared with state-of-the-art approaches, and experimental results showed that our proposed method outperformed other classifiers in widely used hyperspectral image datasets.