L2L-Pix: A Training-Free Pixel Replacement Method to Mitigate Unlabeled Noise in Patch-Based Hyperspectral Image Classification

The main contributions of this paper are summarized as follows:

1. We identify the problem of unlabeled-pixel noise within cropped HSI test patches, an issue largely overlooked in prior patch-based HSIC literature.

2. We propose L2L-Pix, a lightweight, training-free algorithm that replaces unlabeled pixels within a patch with spectrally consistent labeled pixels and integrates directly into any patch-based classification pipeline, requiring no additional learnable parameters and negligible computational overhead.

3. We conduct extensive experiments showing that incorporating L2L-Pix consistently improves classification accuracy across multiple datasets and backbone networks.