Mason Archival Repository Service

Hyperspectral Signature Detection of Low Abundance Intimate Mixtures in Microscenes Using Convolutional Neural Networks

Show simple item record

dc.contributor.advisor Croitoru, Arie
dc.contributor.author Christiansen, Kevin P
dc.creator Christiansen, Kevin P
dc.date 2019-07-31
dc.date.accessioned 2019-10-11T19:18:09Z
dc.date.available 2019-10-11T19:18:09Z
dc.identifier.uri https://hdl.handle.net/1920/11616
dc.description.abstract The high-confidence detection and identification of very low abundance, subpixel quantities of solid materials in nonlinear (intimate) mixtures are still significant challenges for HSI data analysis. Machine learning with convolutional neural networks (CNN) has proven to be an accurate means of identifying and typing features for various forms of data including estimating nonlinear functions and detecting features in one-dimensional data series. The application of CNNs to low abundance intimate mixtures, could improve minimum detectable quantities (MDQs) compared to current algorithms by processing spectra such that subtle features are enhanced and more discriminable. To test this, microscenes of three different intimate mixtures at varying abundance ratios (weight percent) were generated and measured using a benchtop shortwave infrared (SWIR) hyperspectral imager. A microscene is a hyperspectral image of a small-scale human-generated landscape measured in a laboratory. The 3 mixtures measured consisted of silicate sand + Nd2O3, silicate sand + powdered sugar, and soil + aspartame. Several hundred thousand labeled spectra are easily and rapidly generated in one HSI cube of a microscene for neural network training and testing. The mixture detection abilities of several processing methods were compared including deep learning (DL) CNNs in TensorFlow, shallow (non-convolutional) neural networks in both TensorFlow and MATLAB, and ENVI’s matched filter (MF), support vector machine (SVM), and linear spectral unmixing (LSU) functions. The CNN models for each mixture have >98 % average validation accuracies for detecting mixtures of varying abundances, including the lowest abundances measured. Both the DL/CNN and shallow neural networks tested showed increased detection capabilities compared to some more traditional ENVI methods. The networks were able to identify the lowest weight percent of sand + Nd2O3, but the MF and LSU methods did not yield results that would be considered reliable detections. The results establish confidence in using CNNs as a means of detecting low abundance intimate mixtures in real-world scenarios.
dc.language.iso en en_US
dc.subject hyperspectral en_US
dc.subject microscene en_US
dc.subject convolutional neural networks en_US
dc.subject signature detection en_US
dc.subject low abundance en_US
dc.subject mixtures en_US
dc.title Hyperspectral Signature Detection of Low Abundance Intimate Mixtures in Microscenes Using Convolutional Neural Networks en_US
dc.type Thesis en_US
thesis.degree.name Master of Science in Geoinformatics and Geospatial Intelligence en_US
thesis.degree.level Master's en_US
thesis.degree.discipline Geoinformatics and Geospatial Intelligence en_US
thesis.degree.grantor George Mason University en_US


Files in this item

This item appears in the following Collection(s)

Show simple item record

Search MARS


Browse

My Account

Statistics