RGB2HSI Spectral Reconstruction on NEON Tree Canopy Data NEON Aǧaç Kanopi Verileri Üzerinde RGB'den HSG'ye Spektral Geri Çatim
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636673
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: deep learning, GMSR-Net, hyperspectral imaging, NEON dataset, RGB2HSI, spectral reconstruction
- Hacettepe Üniversitesi Adresli: Evet
Özet
Spectral reconstruction aims to generate dense-band hyperspectral data from low-cost RGB images. While existing methods are mostly evaluated on controlled datasets, preserving spectral signatures under high natural texture diversity remains a fundamental challenge. This study systematically investigates spectral reconstruction performance and the resolution mismatch problem using the NEON dataset, which features diverse forest and plant covers. Various methods are compared, and the state-of-the-art GMSR-Net state-space model is proposed with an L1 and SAM-based hybrid loss function for reconstruction. Evaluated via PSNR, RMSE, and SAM metrics, the proposed approach ensures consistent spectral signature preservation, with generalization validated via an external dataset. The findings demonstrate that RGB-based reconstruction offers a practical, cost-effective approximate solution when hyperspectral data is limited.