Neural Architecture Search-Based Few-Shot Learning for Hyperspectral Image Classification

Few-shot learning (FSL) has achieved promising performance in hyperspectral image classification (HSIC) with few labeled samples by designing a proper embedding feature extractor. However, the performance of embedding feature extractors relies on the design of efficient deep convolutional neural network (CNN) architectures, which heavily depends on expertise knowledge. Particularly, FSL requires extracting discriminative features effectively across different domains, which makes the construction even more challenging. In this article, we propose a novel neural architecture search-based FSL model for HSI classification (HCFSL-NAS). Three novel strategies are proposed in this work. First, a neural architecture search (NAS)-based embedding feature extractor is developed to the FSL in HSIC, whose search space includes a group of proposed multiscale convolutions with channel attention. Second, a multisource learning framework is employed to aggregate abundant heterogeneous and homogeneous source data, which enables the powerful generalization of the network to the HSIC with only few labeled samples. Finally, the pointwise-based cross-entropy (CE) loss and the pairwise-based adaptive sparse loss are jointly optimized to maximize interclass distance and minimize the distance within a class simultaneously. Experimental results on four publicly hyperspectral datasets demonstrate that HCFSL-NAS outperforms both the exiting FSL methods and supervised learning methods for HSI classification with only few labeled samples. The code is available at https://github.com/xh-captain/HCFSL-NAS.

I. Introduction

Hyperspectral images (HSIs) contain rich spatial and spectral information, which provides discriminative features for ground objects and has been widely used in geological exploration, target detection, and military reconnaissance [1], [2], [3], [4]. HSI classification (HSIC) is a fundamental task aiming to assign a specific category to each pixel [5], [6], [7]. Over the past decades, deep learning (DL) has shown a remarkable ability to extract effective and hierarchical features for different visual tasks. Models such as stacked autoencoders (SAEs) [8], deep belief networks (DBNs) [9], and CNNs [10], [11], [12], [13], [14] have demonstrated satisfactory performance on HSIC. It is well known that the performance of DL-based methods heavily relies on the sufficient number of labeled samples for each class [15], [16], while obtaining labeled samples of hyperspectral image data is expensive and hard.