HyperMLL: Toward Robust Hyperspectral Image Classification With Multisource Label Learning

Recent advancements in hyperspectral image (HSI) classification rely on deep learning, demanding high-quality labels. Traditional methods of label collection are costly and expertise-intensive. Alternative approaches, such as crowdsourcing, large-scale models, and others, have emerged as cost-effective solutions for label acquisition. However, labels obtained from these sources often exhibit noise. Currently, there is a research gap in addressing the challenges posed by multisource noisy labels in HSI interpretation. To bridge this gap, we introduce a novel method for effectively interpreting HSI with multisource label learning (HyperMLL). First, we develop a novel multisource label noise simulation approach for HSI based on spectral features. Then, we propose the spectral- and spatial-related confusion matrices (CMs) to learn the distribution of noisy labels, aiming at mitigating the impact of multisource noisy labels. In addition, these learned CMs are leveraged to assess the overall credibility and class-specific credibility of different label sources, laying the foundation for future work to ensure label quality while reducing labeling costs. We generate multisource noisy labels at different proficiency levels and conduct detailed experiments on three publicly available datasets. The experimental results demonstrate that our framework outperforms state-of-the-art HSI noise learning methods by a large margin and can accurately evaluate the sourcewise label qualities.