Research & Publications

Spectral Channel Selection for Hyperspectral Remote Sensing Using Learned Gaussian Filters

Research Assistant · Auburn University at Montgomery · Jan 2026 – Present

Advisor: Dr. Olcay Kursun

Manuscript in preparation · Atmospheric Measurement Techniques (AMT) · Target: July 2026

Key Results

R² = 0.979 ± 0.015

Photon flux prediction (6 channels)

R² = 0.999 ± 0.0006

Radiance transfer (frozen channels)

6 of 462

Channels sufficient for full accuracy

High-dimensional hyperspectral sensors capture 400+ spectral channels, most of which are redundant for specific prediction tasks. The goal was to determine the minimum fixed-filter camera configuration that retains full predictive power — enabling cheaper, lighter remote sensing hardware.

Trained an end-to-end model with a Gaussian channel-selection head and a compact MLP predictor. Learned channel centers converged consistently across 10 random seeds regardless of initialization, suggesting the solution is structurally determined by the data.

A follow-up radiance transfer experiment froze the learned 6-channel configuration and retraining only the MLP head for a second target variable, achieving R² = 0.999 ± 0.0006. This confirmed a single optimized 6-filter camera design can support multiple measurement objectives simultaneously.