Using Random Forest Machine Learning Methods to Identify Spatiotemporal Patterns of Cheatgrass Invasion through Landsat Land Cover Classification in the Great Basin from 1984 - 2011

Murphy, Dylan

Undergraduate Honors Thesis

Using Random Forest Machine Learning Methods to Identify Spatiotemporal Patterns of Cheatgrass Invasion through Landsat Land Cover Classification in the Great Basin from 1984 - 2011 Public Deposited

Analytics

Citations

Citeable URL: https://scholar.colorado.edu/concern/undergraduate_honors_theses/0k225b56f

Abstract

In the Great Basin region of the Western United States, the establishment of cheatgrass (Bromus tectorum), an invasive annual grass, has dramatically affected native sagebrush steppe ecosystems. Cheatgrass, and the grass-fire cycle associated with the species, have led to extensive sagebrush community degradation, fragmentation, and eventually, conversion to near-monoculture exotic cheatgrass grasslands. This process has changed regional fire regimes, placing people, property, and native fauna at risk. Despite the known impacts of cheatgrass invasion in the Great Basin, attempts to quantify regional scale changes in ecosystem type have been limited. In this study, a methodology is developed which utilizes Random Forest machine learning to perform a regional-scale, high resolution land cover classification for an annual time series between 1984 and 2011. I discuss several challenges faced during the development of the data-intensive land cover classification methodology and highlight the strengths and weaknesses of this approach. Additionally, I identify preliminary quantitative land cover trends shown in the resulting time series which support the hypothesis that cheatgrass grassland area is increasing and sagebrush steppe is being lost as a consequence. The uncertainty surrounding these predictions is also quantified and potential next steps to improve the accuracy of the time series results are emphasized. The impacts of several features of the land cover classification process on data continuity and model accuracy are also discussed. This research provides a framework for future large-scale, community-level land-cover classifications performed using Random Forest and provides one of the first multi-decade annual time series documentations of cheatgrass’ invasion into the Great Basin’s sagebrush steppe.

Creator