Faculty Mentor(s)
Dr. Brian Yurk, Mathematics
Document Type
Poster
Event Date
4-17-2020
Abstract
Landslide formation is a significant contributor to montane rainforest biodiversity, opening gaps in the tree canopy and leading to the germination of pioneer plants. In order to study the spatio-temporal patterns of landslide formation, we developed a technique using a machine learning algorithm (random forest) to automatically identify landslides in high resolution satellite imagery. Using 4-band Planetscope and 5-band RapidEye satellite imagery, we identified landslides associated with major rain events over the Monteverde Cloud Forest Reserve in Costa Rica. The classifier uses reflectance values along with texture measures and topographic slope to sort pixels into different landscape classes. Slope was calculated using digital elevation models. The overall accuracy of the resulting classifier was 99% when tested against a validation set consisting of landforms that were visually classified using the imagery. Ground-truthing was also performed in July 2019, and several features that were classified in the imagery were visited in the field. This work confirmed the accuracy of the classifier for large features, though some small features observed in the field were not identified by the algorithm. By applying the classifier to imagery collected before and after major rain events, we have been able to see the development and subsequent recovery of landslides associated with the events.
Recommended Citation
Repository citation: Leu, Eric, "Identifying Cloud Forest Landslides in Satellite Imagery: A Machine Learning Approach" (2020). 19th Annual Celebration of Undergraduate Research and Creative Activity (2020). Paper 18.
https://digitalcommons.hope.edu/curca_19/18
April 17, 2020. Copyright © 2020 Hope College, Holland, Michigan.
Comments
This work was partially funded by Michigan Space Grant Consortium, NASA grant #NNX15AJ20H and the Frank and Dorothy Sherburne Mathematics Summer Research Fund.