Hyperspectral Image Analysis & Unmixing

A hyperspectral sensor sees hundreds of narrow spectral bands where a camera sees three — enough to identify what materials compose a scene, if you can decompose each pixel’s spectrum into its constituent signatures. That decomposition, spectral unmixing, has been an AICIP signature problem for two decades.

The lab’s minimum-volume-constrained NMF method (MVC-NMF, IEEE TGRS 2007) became one of the field’s standard baselines and earned the IEEE Geoscience and Remote Sensing Society’s Highest Impact Paper Award in 2012. A best-paper-winning line at WHISPERS 2015 introduced autoencoder cascades for blind unmixing, followed by deep approaches to the problem’s harder variants: uDAS (denoising-autoencoder unmixing), uSDN (unsupervised sparse Dirichlet-Net for hyperspectral super-resolution, CVPR 2018 line), and mutual-Dirichlet networks for joint resolution enhancement.

The thread continues: recent work (arXiv 2026) uses cycle-consistent generative adversarial networks to model nonlinear mixing — the physically realistic case where light scatters between materials before reaching the sensor — moving beyond the linear mixing assumptions that limited two decades of classical methods.