Dual-channel hyperspectral single-pixel imaging with visible and near-infrared bands and a physics-informed neural network

Broadband hyperspectral imaging is a powerful technique that simultaneously captures both spatial and spectral information of objects, with versatile applications such as material discrimination and agricultural monitoring. Conventional broadband hyperspectral imaging systems typically employ multi-channel architectures and spatial scanning methods that result in low light throughput. In this study, we propose a dual-channel hyperspectral single-pixel imaging (DHSSPI) system that achieves high-throughput acquisition across both visible (VIS) and near-infrared (NIR) bands. The system utilizes a digital micromirror device for both spatial modulation and beam separation, enabling a compact optical configuration and simultaneous acquisition across the two spectral channels. Through both numerical simulation and experiments, the proposed DHSSPI system provides improved imaging quality with higher light throughput, compared with whiskbroom hyperspectral imaging systems. The DHSSPI system achieves hyperspectral imaging with the spectral range covering 450–750 nm and 1050–1650 nm, at a spatial resolution of 64 × 64 pixels. By using a monochromatic light source, we confirm that the system achieves the spectral resolutions of 1.9 nm and 10 nm for the VIS and NIR channels, respectively. We also introduce a physics-informed neural network (PINN) for image reconstruction without pre-training to improve reconstruction quality at low sampling rates. Furthermore, we demonstrate that the proposed method can be used for effective material discrimination of real and artificial leaves. The proposed method offers a cost-effective approach for practical applications in agricultural monitoring and material inspection.