Fuzzy graph convolutional network for hyperspectral image classification
Jindong Xu,Kang Li,Ziyi Li,Qianpeng Chong,Haihua Xing,Qianguo Xing,Mengying Ni
DOI: https://doi.org/10.1016/j.engappai.2023.107280
IF: 8
2023-10-14
Engineering Applications of Artificial Intelligence
Abstract:—Graph convolutional network (GCN) has attracted much attention in the field of hyperspectral image classification for its excellent feature representation and convolution on arbitrarily structured non-Euclidean data. However, most state-of-the-art methods build a graph utilize the distance measure, which makes it challenging to fully characterize the complex relationship of hyperspectral remote sensing data . Moreover, the hyperspectral image usually has uncertainty introduced by the problems of the spectral variability and noise interference. This article uses fuzzy theory to optimize the GCN and thus solve the uncertainty problem in hyperspectral images, and presents a novel fuzzy graph convolutional network (F-GCN) for hyperspectral image classification. By calculating the fuzzy similarity of samples, a robust graph is first built rather than using the traditional Euclidean distance method, which allows a better representation of the complex relationship between hyperspectral remote sensing data. Furthermore, the proposed network introduces fuzzy layers into the model to cope with the ambiguity of the hyperspectral image. Finally, the classification results for three real-world hyperspectral data sets to show its feasibility and effectiveness in hyperspectral image classification.
automation & control systems,computer science, artificial intelligence,engineering, electrical & electronic, multidisciplinary