Hyperspectral imaging (HSI) generally ignores the polarimetric information inherent in electromagnetic waves. However, by incorporating such polarimetric data, Polarimetric HSI (P-HSI) offers the chance to increase the separation between classes of objects and thereby improve general anomaly detection accuracy. However, P-HSI further expands on the multitude of data inherent in HSI due there now being multiple images of the same scene, albeit at different polarization angles. Thus, selecting appropriate and useful subsets of this data for anomaly detection is of interest. This paper considers this issue in P-HSI by considering fusion/ensemble methodologies at the data and feature level for effective combination of the available P-HSI data with an emphasis on anomaly detection. Data fusion is considered by combining multiple Stokes Vectors into a larger image cube, from which features are extracted. Feature fusion is considered by computing features from individual Stokes Vectors and then combining retained features for anomaly detection performance. Three anomaly detection algorithms and two experimentally collected P-HSI images are considered; full factorial experiments of all combinations of data/features are considered for evaluation. The results show that the appropriate selection of data/features is critical for anomaly detection performance in P-HSI with both the level of fusion and subset being critical to accuracy.
Data and Feature Fusion Approaches for Anomaly Detection in Polarimetric Hyperspectral Imagery
2021-08-16
1280658 byte
Conference paper
Electronic Resource
English
6.0105 Using Self-Organizing Maps for Anomaly Detection in Hyperspectral Imagery
British Library Conference Proceedings | 2002
|