top of page
Katherine Mistick
Graduate Research Projects
2020-2022
As an MS student in the University of Utah's Geography department, I undertook many research projects as part of my coursework and thesis hours. While many projects related back to my interest in wildland firefighter safety, others utilized remotely sensed or geospatial data to solve a variety of environmental questions.
Projects highlighted here were completed for courses focusing on geospatial data visualization, optical remote sensing, and machine learning. These projects improved my skills in R, python, and ENVI.
Geovisualization: Building a Utah Avalanche Dashboard using Shiny in R
The Shiny package in R allows users to build powerful web applications that allow users to interact with your data and analysis. Using data from the Utah Avalanche Center, I built a web application to display avalanche incidents in an interactive and educational capacity. The goal of this project was to create a compelling, interactive visualization that allows users to better understand historical avalanche incidents in Utah.
Follow this link to interact with the web app.
Screenshot of my Utah Avalanche Dashboard, showing avalanche incidents color-coded by incident type.
Spatial Modeling and Geocomputation: Machine learning for methane plume delineation, a feasibility study
The output of a random forest classifier for an AVIRIS-NG scene with a methane plume. White indicates areas of the plume that were correctly classified, while pink/red shows areas that were false negatives.
In a collaborative project, my team assessed the feasibility of multiple machine learning algorithms for identifying the shape and location of methane plumes in hyperspectral AVIRIS-NG imagery. My contribution focused on using a Random Forest model to classify imagery as methane/non-methane.
Overall, I found that a random forest classifier was able to identify methane plumes but severely under-classified images, resulting in many false negatives. High methane concentrations were correctly identified, but the model likely needed a much larger training dataset to be useful. This project was continued by other members of my research group, exploring the use of convolutional neural networks to delineate methane plumes from imagery.
Advanced Optical Remote Sensing: Determining the best wavelength range for Identifying CO2 Plumes using Matched Filtering
Matched filtering is a type of spectral transformation used to enhance target features in imagery. For this project I compared different input wavelength ranges to a matched filtering algorithm designed to identify significant concentrations of carbon dioxide in hyperspectral imagery. This work was done in python and using ENVI for image analysis.
Initially designed by Foote et. al 2020, the matched filter uses a default wavelength range of 2122-2488 nm to enhance methane or carbon dioxide plumes. I found that restricting this wavelength range to the shortwave infrared (SWIR) improved plume detection by minimizing false positives. Further, this wavelength restriction improved the calculation of the carbon dioxide flux when compared to EPA-reported fluxes for the study areas. These findings are logical considering carbon dioxide shows strong spectral absorption in the SIWR.
Top left: red concentrations indicate strong false positives from the default wavelength range of a matched filter image transformation. Bottom left: the same extent as the top left, indicating a dramatic improvement in false positives by restricting the wavelength to just incorporate shortwave infrared reflectance. Right: by minimizing false positives we can more accurately indicate CO2 plumes in imagery.
bottom of page