Tuesday, November 19 at 11:00am
VanderWerf Hall, 102
27 Graves Place, Holland, MI 49423-3617
“Remote Sensing and Machine Learning: A Model for the Digital Mapping of Cloud Forest Landslides” by Eric Leu
Large features in satellite imagery can be digitally mapped using machine learning algorithms. A natural application of this field is the identification of landslides across remotely sensed images of high elevation cloud forests before and after major rainstorms, producing models for studying the patterns of tree gap openings and the creation of habitat patches stimulating pioneer plant germination. The goal of our research was to create a machine learning model for the automatic classification of landslides and other geological features over large sets of satellite imagery, enabling the study of spatio-temporal patterns of landslide activity in the Monteverde Cloud Forest Reserve of Costa Rica. Using imagery from the 4-band Planetscope and 5-band RapidEye satellite constellations, we have created a Random Forest classifier capable of sorting individual pixels into distinct topographic classes, combining pixel-based spectral information along with digitally calculated image texture and elevation models. Our classifier achieved an overall accuracy of 99% when tested against validation data generated through visually sampled features in the imagery. Field validation conducted in July 2019 allowed the correction of Type I and Type II errors in the model, and images preceding and subsequent to major rain events were used to generate a time series emphasizing the effect of the storms.
“Dunes and Drones: A Machine Learning Approach to Mapping Vegetation With Ground-Based and Aerial Imagery” by Jack Krebsbach
Active coastal dune complexes are dynamic environments. While patches of dune activity encourage ecological diversity, mobile dunes may also negatively impact human structures in coastal regions. Sediment transport, which is primarily driven by wind flow and mitigated by vegetation, affects coastal dunes’ topographical features. Multispectral imagery acquired by a small unmanned aerial system (sUAS) can be used to create high resolution vegetation density maps to study factors of dune mobility and for the management of dune complexes.
Our method starts with close up ground-based photography to estimate vegetation density. First, textural analysis is calculated on the ground imagery and the photographs are transformed into different color spaces (HSV,LAB.etc) to create several feature images. Next, the machine learning algorithm Random Forest is applied to these feature images to classify all the pixels into a variety of classes (dead vegetation, sand, etc.) to estimate vegetation density. These estimates are used for calibration to process multispectral sUAS imagery and calculate the vegetation density across an entire orthomosaic map. The orthomosaic is an aerial image geometrically corrected and made by stitching together hundreds of sUAS obtained images.