Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
4973014 | ISPRS Journal of Photogrammetry and Remote Sensing | 2016 | 17 Pages |
Abstract
We present a novel semi-supervised algorithm for classification of hyperspectral data from remote sensors. Our method is inspired by the Tracking-Learning-Detection (TLD) framework, originally applied for tracking objects in a video stream. TLD introduced the co-training approach called P-N learning, making use of two independent 'experts' (or learners) that scored samples in different feature spaces. In a similar fashion, we formulated the hyperspectral classification task as a co-training problem, that can be solved with the P-N learning scheme. Our method uses both spatial and spectral features of data, extending a small set of initial labelled samples during the process of region growing. We show that this approach is stable and achieves very good accuracy even for small training sets. We analyse the algorithm's performance on several publicly available hyperspectral data sets.
Keywords
Related Topics
Physical Sciences and Engineering
Computer Science
Information Systems
Authors
MichaÅ Romaszewski, PrzemysÅaw GÅomb, MichaÅ Cholewa,