Global-local Feature Adaptive Fusion Мethod for Small Sample Classification of Нyperspectral Images

Acquisition of labeled samples for hyperspectral image classification is usually a time- and labor-consuming task. How to effectively improve the classification accuracy using a small number of samples is one of the challenges in the field of hyperspectral image classification. Most of existing classification methods for hyperspectral images lack sufficient multi-scale information mining, which leads to unsatisfactory classification performance due to small sample numbers. To address the aforementioned issue, this paper designed an adaptive fusion method by integrating global and local features for hyperspectral image classification with small sample numbers. Based on the dynamic graph convolutional network and the depth wise separable convolutional network, a two- branch network structure was constructed to mine the potential information of hyperspectral images from the global and local scales, which realizes the effective usage of labeled samples. Furthermore, the polarization self-attention mechanism was introduced to further improve the expression of intermediate features in the network while cutting down the loss of feature information, and the adaptive feature fusion mechanism was adopted to carry out adaptive fusion of global and local features. Finally, the fusion features flow into the full- connection layer and are manipulated by softmax to obtain prediction labels for each pixel of the hyperspectral image. In order to verify the effectiveness of the proposed method, classification experiments were carried out on four hyperspectral image benchmark data sets including University of Pavia, Salinas, WHU-Hi-LongKou, and WHU-Hi-HanChuan. We discussed and analyzed the influence of model parameters and different modules on the classification accuracy. Subsequently, a comprehensive comparison with seven existing advanced classification methods was conducted in terms of classification visualization, classification accuracy, number of labeled samples, and execution efficiency. The experimental results show that the dynamic graph convolutional network, depth wise separable convolutional network, the polarization self- attention mechanism, and the adaptive feature fusion mechanism all contributed to the improvement of hyperspectral image classification accuracy. In addition, compared with traditional classifiers and advanced deep learning models, the proposed method considered both execution efficiency and classification accuracy, and can achieve better classification performance under the condition of small sample numbers. Specifically, on these four data sets (i.e., University of Pavia, Salinas, WHU- Hi- LongKou and WHU- Hi- HanChuan). The overall classification accuracy was 99.01%, 99.42%, 99.18% and 95.84%, respectively; the average classification accuracy was 99.31%, 99.65%, 98.89% and 95.49%, respectively; and the Kappa coefficient was 98.69%, 99.35%, 98.93% and 95.14%, respectively. © 2023 Journal of Geo-Information Science. All rights reserved.