Tree-shaped multiobjective evolutionary CNN for hyperspectral image classification

Abstract

Convolutional neural networks (CNNs) have achieved significant performances in hyperspectral image (HSI) classification in recent years. However, designing a high-performance CNN depends on human expertise heavily, which usually takes considerable time and labor. With regard to reducing the burden of designing the networks, neural architecture search (NAS) has attracted increasing attention. A typical NAS approach aims to optimize the network architectures in a predefined search space with a suitable search algorithm automatically. However, the existing NAS work does not fully consider the spatial resolution and the spectral noise interference of HSIs. Furthermore, most NAS approaches use sequential blocks or cells to construct the networks, which are unsuitable for extracting multiscale features of HSIs and result in degraded performance. Considering the above challenges, we propose a tree-shaped multiobjective evolutionary CNN (TMOE-CNN) for HSI classification. An expanded search space is designed, which includes the image patch size and the channel number of the input image patches. A multibranch supernetwork structure is proposed, which resembles a tree as the fundamental architecture for the network block. The image patch size and the denoising strength of the input image patches can be established adaptively throughout the evolutionary search process. The tree-shaped networks can fuse multiscale features to enhance the capacity of the network for feature extraction. Additionally, we consider both the classification accuracy and the floating-point computational complexity in the environmental selection. It is helpful to find the networks with simple structure and low complexity while ensuring classification accuracy. Experiments on different HSI datasets show that TMOE-CNN can search CNNs with high accuracies and simple structures automatically.