Date of Award

Spring 5-15-2019

Author's School

School of Engineering & Applied Science

Author's Department

Electrical & Systems Engineering

Degree Name

Doctor of Philosophy (PhD)

Degree Type

Dissertation

Abstract

Hyperspectral data are important for remote applications such as mineralogy, geology, agriculture and surveillance sensing. A general pipeline converting measured hyperspectral radiance to the surface reflectance image can provide planetary scientists with clean, robust and repeatable products to work on.

In this dissertation, the surface single scattering albedos (SSAs), the ratios of scattering eciency to scattering plus absorption eciences of a single particle, are selected to describe the reflectance. Moreover, the IOF, the ratio of measured spectral radiance (in the unit of watts per squared-meter and micrometer) to the solar spectral radiance (in the unit of watts per squared-meter and micrometer) at the observed time, is used to indicate the measurements.

This pipeline includes two main parts: retrieving SSAs from IOF and reconstructing the SSA images from the SSA cube. The Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) on the Mars Reconnaissance Orbiter (MRO) helps scientists identify locations on Mars that may have exhibit hydrated mineral phases. This dissertation mainly focuses on developing the pipeline for CRISM data. One should notice that pipelines for other hyperspectral spectrometers can also be developed based on almost the same idea. Retrieving surface kinetic temperatures and SSA values from IOF data is challenging because the problem is under-determined and ill-posed, including modulating effects of atmospheric aerosols and gases, and surface scattering and emission properties. We introduce a general framework called STANN (Separating Temperature and Albedo using Neural Networks) to solve this kind of problem. STANN takes the hyperspectral IOF cube as inputs and outputs the retrieved temperature mapping and the corresponding SSA cube. Our STANN is derived using the Discrete Ordinates Radiative Transfer function to describe the forward model from SSA and temperature to IOF. In the STANN, we have a generator to generate more training samples based on limited library spectra and a neural network to approximate the inverse function based on enough generated training samples. This framework has been implemented for the Compact Imaging Spectrometer for Mars in a detailed manner. SSA can be computed from IOF directly by modeling the thermal and solar reflectance together, based on retrieved temperatures. Because accurate retrieved temperature directly leads to accurate SSA, we compare the accuracy of retrieved temperatures from STANN.

The retrieved temperature has only 4 K error by one point validation (242 K) using the Curiosity Rover's thermal radiometer data. Our STANN temperature map is compared with a temperature map generated independently from a theoretical thermal model. The theoretical thermal model describes the relationship between Lambert albedo at the wavelength 1.0 µm, thermal inertia and the surface temperature. However, because the thermal inertia has pixel size larger than 100 m/pixel, the generated temperature also has the same pixel size. Our STANN temperature is projected into the same pixel size (100 m/pixel) by the classic projection method. The two temperature maps have consistent global patterns.

Retrieved from an IOF cube, a noisy hyperspectral SSA cube needs to be denoised and reconstructed onto the Mars surface. We propose a new algorithm, hypothesis-based estimation with regularization (HyBER), to reconstruct and denoise hyperspectral image data without extra statistical assumptions. The hypothesis test selects the best statistical model approximating measurements based on the data only. Gaussian and Poisson distributions are common respectively for continuous and integer random variables, due to the law of large numbers. Hyperspectral IOF data result from converting discrete photon counting data to continuous electrical signals after calibration. Thus, so far, Gaussian and Poisson are candidate distributions for our hypothesis tests. A regularized maximum log-likelihood estimation method is derived based on the selected model. A spatially dependent weighting on the regularization penalty is presented, substantially eliminating row artifacts that are due to non-uniform sampling. A new spectral weighting penalty is introduced to suppress varying detector-related noise. HyBER generates reconstructions with sharpened images and spectra in which the noise is suppressed, whereas fine-scale mineral absorptions are preserved. The performance is quantitatively analyzed for simulations with relative error 0.002%, which is better than the traditional non-statistical methods (baselines) and statistical methods with improper assumptions. When applied to the Mars Reconnaissance Orbiter's Compact Reconnaissance Imaging Spectrometer for Mars data, the spatial resolution and contrast are about 2 times better as compared to map projecting data without the use of HyBER. So far, part of our results has enabled planetary scientists to identify minerals and understand the forming history of Mars craters. Some of these findings are verified by the Opportunity Rover's measurements. In the future, results from this pipeline for CRISM are promising to play more and more critical roles in the planetary science.

Language

English (en)

Chair

Joseph A. O'Sullivan

Committee Members

R. M. Arthur, Raymond E. Arvidson, Matthew D. Lew, Douglas A. Wiens,

Comments

Permanent URL: https://doi.org/7936/j2wp-x475

Share

COinS