Computational methods for scene inference from thermal hyperspectral measurements
Date
2025
DOI
Authors
Version
OA Version
Citation
Abstract
Conventional imaging systems relying on the visible spectrum face significant limitations in darkness, such as nighttime environments without illumination. Longwave infrared (LWIR) imaging provides a compelling solution by harnessing ambient thermal radiation naturally emitted by objects that is present both day and night. Extending LWIR imaging to a hyperspectral approach, which captures detailed spectral information across multiple wavelengths, enhances our ability to perceive, identify, and reconstruct scenes, even in highly challenging environments. This thesis develops computational methods for performing scene inference from ambient thermal hyperspectral measurements. Its central focus is on absorption-based ranging, a method that estimates object distances by analyzing spectral signatures introduced by atmospheric absorption. Absorption-based ranging in natural environments is a challenging problem due to two main factors: (i) weak thermal emissions from the object, and (ii) influence from surrounding sources of environmental thermal radiation, which can obscure or distort the object-emitted signal. To assess feasibility and inform practical sensor design, we perform a comprehensive trade-space analysis using Fisher information and the Cramér-Rao bound. This analysis systematically evaluates ranging performance under varying naturalconditions and sensor parameters. Our results provide valuable insights into feasible scenarios and help guide the development of optimized sensors designed to enhance passive ranging performance. We propose inversion algorithms and validate their effectiveness through demonstrations with both simulated data and experimental measurements acquired with contemporary LWIR hyperspectral sensors under realistic conditions characterized by minimal temperature variations in natural scenes. The results demonstrate promising distance estimation accuracy for highly emissive objects. However, for highly reflective surfaces, reflections of downwelling radiance (incoming radiation from sky) introduce biases, leading to inaccuracies. To address this, we propose methods that exploit ozone absorption features in downwelling radiance to separate reflections from object emissions, significantly improving ranging accuracy for reflective materials. These methods are validated with experimental data. We analyze a stereo hyperspectral configuration that combines disparity-based and absorption-based ranging methods. These approaches rely on distinct scene cues, with disparity sensitive to spatial structure and absorption relying on spectral signatures. Using Fisher information and the Cramér-Rao bound, we show that disparity dominates at short ranges, while absorption cues are more informative at longer distances. This analysis underscores the potential of integrating both methods to improve ranging accuracy across diverse scenarios. We extend our thermal radiation models to estimate spatially resolved air temperature along the propagation path by formulating an inverse tomography problem using spectrally resolved measurements. We analyze the feasibility and conditioning of this problem using Fisher information and the Cram´er-Rao bound, revealing that temperature estimation accuracy strongly depends on atmospheric absorption profiles and measurement configurations. Through numerical simulations, we demonstrate that our inversion methods successfully reconstruct air temperature profiles, accurately capturing localized temperature variations.
Description
2025
License
Attribution 4.0 International