You are here

ECE MS Thesis Defense: "Reflectance Retrieval for Hyperspectral Imagery Collected Over Urban Environments," Ian Fletcher

24
May

442 Dana

May 24, 2018 10:00 am
May 24, 2018 10:00 am

Abstract:

Hyperspectral Imaging (HSI) aims to classify targets based upon the extracted spectral reflectance signatures of objects within an image. This reflectance retrieval process estimates and compensates for the atmospheric effects on measured radiance data. Traditional reflectance retrieval methods rely upon simplifying assumptions about the target scene geometry. 

In particular, these methods require the target to exist in an open environment, in which all scene-incident solar and atmospheric illumination reach the target. When the open environment assumption is invalid, such as in urban environments where targets exist in complex lighting conditions, traditional reflectance retrieval methods will fail.

This thesis builds upon recent research into reflectance retrieval methods that do not rely on the open-environment assumption. These approaches fuse hyperspectral imagery with a model of target scene geometry to estimate the irradiance incident to targets in urban settings. In this thesis, we discuss an improvement to this method that fully models how light propagates through an environment before it reaches the target. In theory, this new irradiance estimate will fully account for the complex illumination conditions found in urban environments, thereby enabling accurate reflectance retrieval for targets in these settings. This model can be adapted to solve the forward problem of radiance estimation or the inverse problem of reflectance retrieval.

In this thesis, we discuss this new framework for estimating light propagation through urban scenes. Our main contribution to the problem of urban reflectance retrieval is the performance analysis of this method. For our analysis, we use this method to estimate the radiance that was measured by an HSI sensor during a collection campaign over two representative urban scenes. We then evaluate the areas in which this model performs well and those in which the model is unable to estimate this measured radiance. We trace some errors to simplifying assumptions made by the model and identify multiple possible sources of error for other observed modeling discrepancies. In all cases, we recommend modeling improvements to address these errors and experiments to test our presented hypotheses. Finally, we use this model to retrieve the reflectance signatures of the targets in our urban scenes. In doing so, we show that this model greatly improves the accuracy of retrieved reflectance over that retrieved by traditional models for targets in complex illumination conditions.

  • Professor Vinay Ingle (Advisor)
  • Professor Bahram Shafai
  • Dr. Steven Golowich
  • Dr. Dimitris Manolakis