Ensemble Network-Based Distillation for Hyperspectral Image Classification in the Presence of Label Noise

Abstract

Deep learning has made remarkable strides in hyperspectral image (HSI) classification, significantly improving classification performance. However, the challenge of obtaining accurately labeled training samples persists, primarily due to the subjectivity of human annotators and their limited domain knowledge. This often results in erroneous labels, commonly referred to as label noise. Such noisy labels can critically impair the performance of deep learning models, making it essential to address this issue. While previous studies focused on label noise filtering and label correction, these approaches often require estimating noise rates and may inadvertently propagate noisy labels to clean labels, especially in scenarios with high noise levels. In this study, we introduce an ensemble network-based distillation (END) method specifically designed to address the challenges posed by label noise in HSI classification. The core idea is to leverage multiple base neural networks to generate an estimated label distribution from the training data. This estimated distribution is then used alongside the ground-truth labels to train the target network effectively. Moreover, we propose a parameter-adaptive loss function that balances the impact of both the estimated and ground-truth label distributions during the training process. Our approach not only simplifies architectural requirements but also integrates seamlessly into existing deep learning frameworks. Comparative experiments on four hyperspectral datasets demonstrate the effectiveness of our method, highlighting its competitive performance in the presence of label noise.

Keywords

• ensemble network
• distillation
• hyperspectral image
• classification