Semi-supervised Learning on Hyperspectral Imagery
Abstract
Classification of hyperspectral imagery (HSI) is an important research topic for both military and commercial applications. Significant research into deep learning techniques have been concentrated on this task. However, classifying high dimensional HSI data with a limited number of training samples remains an open issue. Semi-supervised learning offers a solution as it requires labeling only a small subset of the pixels and leverages a large number of unlabeled data during its training. In this paper we investigate a variety of both supervised and semi-supervised methods on a new, large HSI dataset called AeroRIT. We determine that the results previously reported in the semi-supervised literature for smaller, older HSI datasets did not generalize to the AeroRIT dataset. Overall, the semi-supervised methods did not provide a significant improvement in performance relative to supervised learning in the limited labeled pixel regime.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 10, 2021
- Accession Number
- AD1124677
Entities
People
- Colin Olson
- Krish Ganotra
- Kristen Nock
- Leslie N. Smith
- Mark Snyder
Organizations
- United States Naval Research Laboratory