HyperLUCID: Zero-label hyperspectral change detection hits 97.9% accuracy, 100x faster

Hyperspectral change detection (HCD) typically relies on semi-supervised learning with labeled pixels, which breaks down when new images arrive onboard a satellite or drone — labels simply aren't available for freshly acquired data. In this paper, Chia-Hsiang Lin, Shih-Min Hsu, Ching-Yun Liang, Jocelyn Chanussot, and Jhih-Yan Chen introduce HyperLUCID, a fully unsupervised algorithm that eliminates the labeling requirement entirely. The method works by iteratively augmenting a training set with confidently unchanged pixel samples, using a safe pseudo-labeling scheme to avoid noise accumulation. It then learns a spectrum calibration function that compensates for variability in acquisition conditions across bitemporal images, making changed pixels stand out clearly in the calibrated spectral domain.

On multiple real benchmark HCD datasets, HyperLUCID delivers state-of-the-art accuracy of 93.6% to 97.9% overall, while being 10 to 100 times faster than most benchmark HCD methods. This computational efficiency comes from a lightweight model architecture designed for edge deployment. The work is accepted at IEEE Transactions on Image Processing and published on arXiv (2608.06028), with source code publicly available. For professionals in remote sensing, environmental monitoring, and defense, HyperLUCID addresses a critical bottleneck: real-time change detection in resource-constrained, label-starved environments. It opens the door to fully autonomous onboard analysis of hyperspectral imagery, enabling immediate detection of land cover changes, illegal construction, or environmental hazards without waiting for ground-truth labeling.