CADR-BL: Class-Adaptive Dictionary Reconstruction with Broad Learning for Few-Shot Hyperspectral Image Classification

Hyperspectral image (HSI) classification in few-shot scenarios faces two core challenges. Limited samples and high spectral similarity lead to insufficient inter-class feature discriminability, and commonly used deep models suffer from the risk of overfitting. To address these problems, this paper proposes a Class-Adaptive Dictionary Reconstruction with Broad Learning (CADR-BL) method. Specifically, the method constructs an exclusive adaptive dictionary for each category and adopts an alternating minimization strategy to achieve sparse reconstruction of intra-class pixels, thereby enhancing intra-class spectral consistency and suppressing inter-class interference. On this basis, an improved Hyperspectral Broad Learning (HS-BL) model is introduced to efficiently classify the reconstructed features. Random feature mapping and closed-form solutions of output weights are utilized to alleviate overfitting in few-shot learning. Experiments conducted on three benchmark datasets, namely Indian Pines, Salinas, and WHU-Hi-HanChuan, show that CADR-BL outperforms several mainstream few-shot classification methods in terms of overall accuracy, average accuracy, and Kappa coefficient. Notably, CADR-BL maintains robust performance even with extremely limited training samples, and is less sensitive to variations in sample size than other comparative methods, demonstrating strong generalization ability. The proposed method provides a reliable technical reference for few-shot HSI classification in applications such as precision agriculture, environmental monitoring, and resource exploration.