Optical Implementation of Learned Spectral Band Selection for Material Classification

Spectral signatures uniquely characterize materials based on their interaction with light across different wave-lengths, enabling an accurate classification. However, acquiring full-spectrum data is time-consuming and storage-intensive. To address this problem, we present an optical whiskbroom system that emulates band selection during acquisition, selecting only the most informative spectral bands. Unlike traditional post-processing methods, our approach integrates band selection into the sensing stage. Spectral signatures with 478 bands were acquired for 15 materials and demonstrated that a classification accuracy of 98% can be achieved using only 4 selected bands. This highlights the potential to reduce system complexity, acquisition time, and data redundancy while maintaining high classification performance.