A query-driven twin network framework with optimization-based meta-learning for few-shot hyperspectral image classification

Deep learning has achieved remarkable results in hyperspectral image (HSI) classification due to its powerful deep feature extraction and nonlinear relationship processing capabilities. However, the success of deep learning methods is largely dependent on extensive labeled samples, which is both time-consuming and labor-intensive. To address this issue, a novel query-driven meta-learning twin network (QMTN) framework is proposed for HSI few-shot learning. QMTN uses two meta-learning channels, allowing for the comprehensive learning of meta-knowledge across diverse meta-tasks and enhancing learning efficiency. Within the QMTN framework, a lightweight spectral-spatial attention residual network is proposed for extraction of HSI features. The network incorporates a residual mechanism in both spectral and spatial feature extraction processes and includes an attention block to improve network performance by focusing on key locations in the spatial features. To maximize the use of the limited samples for constructing diverse meta-tasks, two meta-task generation approaches are employed, with and without simulated noise. Experiments on three public HSI datasets demonstrate that the QMTN framework effectively reduces the dependence on labeled samples in a single scene and significantly improves the classification performance and convergence of the internal network. The meta-task generation method with simulated noise can improve the classification performance of the QMTN.