Infrared Single-Pixel Hyperspectral Imaging via Spatial-Temporal Multiplexing

Technical Summary: Infrared Single-Pixel Hyperspectral Imaging via Spatial-Temporal Multiplexing

Problem Statement
Near-infrared (NIR) hyperspectral imaging is a critical tool for revealing morphological and chemical information in fields ranging from biomedical diagnostics to material characterization. However, conventional spectral imagers face significant limitations. They typically rely on costly focal plane arrays (FPAs) and complex spectral-resolved configurations (such as spatial dispersion devices or spectral filtering), which introduce system redundancy, reduce optical throughput, and impose a fundamental trade-off between spectral bandwidth and resolution. Furthermore, existing single-pixel spectral imaging approaches often depend on mechanical moving parts (introducing inertial errors) or require multi-pixel detectors that limit acquisition rates for dynamic scenes. There is a need for a system that offers high-throughput, high-speed, and cost-effective NIR hyperspectral imaging without the complexity of traditional FPA-based systems.

Methodology
The authors propose and implement a single-pixel NIR hyperspectral imaging system that integrates time-stretch spectroscopy with spatial-encoding compressive measurements. The system architecture operates through the following core mechanisms:

1. Spectrum-to-Time Mapping: A broadband NIR ultrashort pulse (from an Er-doped mode-locked fiber laser) is propagated through a 10-km single-mode fiber (SMF). This dispersive medium imposes a large group velocity dispersion (GVD), mapping the optical spectrum of the pulse directly into a temporal waveform. This creates a "time-stretched" optical source where wavelength is linearly encoded into time.
2. Spatial Encoding: The time-stretched pulse illuminates a sample. The reflected or transmitted light is then modulated by a programmable Digital Micromirror Device (DMD) positioned at the image plane. The DMD imposes structured spatial patterns (using a Walsh-ordered Hadamard matrix) onto the temporally dispersed signal.
3. Single-Pixel Detection: Instead of a multi-pixel array, the entire modulated signal is collected by a single high-bandwidth (5 GHz) InGaAs photodetector. The detector captures the intensity of the light as a function of time, which inherently contains both spatial (via the DMD pattern) and spectral (via the time-stretch mapping) information.
4. Data Acquisition and Reconstruction: The temporal waveforms are recorded by a high-speed oscilloscope. By correlating the recorded time-domain sequences with the known projected spatial patterns, the system reconstructs the hyperspectral datacube. The authors employ compressive sensing (specifically the Total Variation Augmented Lagrangian, TVAL3, algorithm) to reconstruct the 3D datacube from sub-Nyquist sampling rates, leveraging the intrinsic redundancy of hyperspectral data.

Key Contributions

• Novel Architecture: The paper presents a novel imaging modality that multiplexes 3D spatial-spectral information into a 1D temporal sequence without requiring complex spectroscopic devices or costly NIR FPAs.
• High-Fidelity Mapping: The system achieves a precise wavelength-to-time mapping with a slope of 0.173 ns/nm, enabling the reconstruction of 50 spectral bands across the 1550–1600 nm range.
• Efficient Acquisition: By utilizing the time-stretch technique, the system captures spectral information within the duration of a single spatial pattern switch, avoiding additional temporal overhead. This allows for high-speed imaging where increasing the number of spectral channels does not compromise the frame rate.
• Dynamic Monitoring: The integration of compressive sensing enables real-time monitoring of dynamic processes. The authors demonstrated a datacube refreshing rate of 12 Hz for dynamic liquid injection scenarios under sub-Nyquist sampling (0.4 sampling ratio).

Experimental Results
The system was validated through static and dynamic experiments:

• Static Imaging: The system successfully reconstructed a 64×64 spatially resolved hyperspectral datacube. It resolved a USAF-1951 target with a spatial resolution of 177 µm. The reconstructed spectra of bandpass filters (centered at 1550 nm and 1600 nm) showed excellent agreement with reference measurements from an Optical Spectrum Analyzer (OSA).
• Dynamic Imaging: The system monitored the injection of carbon tetrachloride (CCl₄) into ethanolamine (C₂H₇NO). The reconstructed transmission spectra matched FTIR references. The system visualized the density-driven sinking of CCl₄ droplets, distinguishing the two liquids based on their spectral signatures at specific channels (n=5 and n=25).
• Performance Metrics: The system achieved a spectral resolution of approximately 1 nm (4 cm⁻¹) and demonstrated the ability to reconstruct high-quality images with a sampling ratio significantly below the Nyquist limit.

Significance and Claims
The authors claim that this architecture offers a single-pixel simplicity combined with high optical throughput and efficient data acquisition. By eliminating the need for expensive FPAs and mechanical scanning components, the system provides a cost-effective solution for NIR spectral imaging. The work demonstrates that time-stretch spectroscopy can be successfully extended to the field of hyperspectral imaging, enabling rapid, high-resolution, and high-throughput imaging. The authors suggest that this approach paves a novel way for applications in biomedical diagnostics and material characterization, particularly in scenarios requiring lightweight, compact, and adaptive imaging systems. They also note that the architecture could potentially be extended to mid-/far-infrared or terahertz wavelengths through frequency upconversion, addressing current detector limitations in those spectral regions.