Efficient evolutionary multi-scale spectral-spatial attention fusion network for hyperspectral image classification

Mengxuan Zhang, Zhikun Lei, Long Liu, Kun Ma, Ronghua Shang, Licheng Jiao

Abstract: Most deep neural networks for hyperspectral image classification are designed manually, which may bring biases, architecture redundancy and some other unexpected human influences. Although neural architecture search can automatically design the networks, the dilemma of search efficiency and optimal effectiveness has become the major obstacle in current neural architecture search methods. This study proposes a surrogate-assisted evolutionary neural architecture search for hyperspectral image classification. Multi-scale spectral-spatial attention fusion networks are encoded as the individuals in the evolutionary neural architecture search. The surrogate model based on logistic regression can decrease the high time cost of a large number of architecture evaluations in the evolution search. To improve the predictive ability of the surrogate model and reduce the initialisation time, a novel surrogate-assisted evolution scheme is designed to take the dominance relationships between chromosomes as the prediction objective. Furthermore, the surrogate model is extended according to the differences in the influences of the base types in the chromosomes on the phenotypes. In the evolution, a global and local alternating optimisation scheme is adopted to improve the search performance for network architectures with high precision. Experiments on different hyperspectral image datasets show that this work can design multi-scale spectral-spatial attention fusion networks with high classification performance automatically within a shorten search time.