Presenter Information

Scout Jarman, Utah State University

Class

Article

College

College of Science

Department

Mathematics and Statistics Department

Faculty Mentor

Kevin Moon

Presentation Type

Oral Presentation

Abstract

Compared to color images, hyperspectral images are high dimensional, containing hundreds of channels of information. To distill this information, and capture spatial information, image segmentation is used to group similar pixels together. A popular image segmentation algorithm is the marker-based Watershed Transform. One difficulty with this algorithm is choosing the markers, or locations, that seed the algorithm. There are various approaches for automatic marker placement depending on the application, with little consensus on the most general method for hyperspectral images. We propose using an ensemble of random segmentations. Specifically, we investigate a simple, unbiased random marker placement strategy to generate an ensemble of diverse segmentations. Our methods were tested on four standard hyperspectral images, Indian Pines and Salinas (rural), and Pavia University and Center (urban), looking at overall accuracy to compare performance. We find that the ensemble methods perform equivalently or better than single segmentation and is more stable than generating a single segmentation.

Location

Logan, UT

Start Date

4-11-2023 10:30 AM

End Date

4-11-2023 11:30 AM

Share

COinS
 
Apr 11th, 10:30 AM Apr 11th, 11:30 AM

Introduction to Ensemble Watershed Segmentation

Logan, UT

Compared to color images, hyperspectral images are high dimensional, containing hundreds of channels of information. To distill this information, and capture spatial information, image segmentation is used to group similar pixels together. A popular image segmentation algorithm is the marker-based Watershed Transform. One difficulty with this algorithm is choosing the markers, or locations, that seed the algorithm. There are various approaches for automatic marker placement depending on the application, with little consensus on the most general method for hyperspectral images. We propose using an ensemble of random segmentations. Specifically, we investigate a simple, unbiased random marker placement strategy to generate an ensemble of diverse segmentations. Our methods were tested on four standard hyperspectral images, Indian Pines and Salinas (rural), and Pavia University and Center (urban), looking at overall accuracy to compare performance. We find that the ensemble methods perform equivalently or better than single segmentation and is more stable than generating a single segmentation.