Comparison of Segmentation Algorithms for an Object-Based Anomaly Detection in Hyperspectral Images
Martinez de Agirre, Alex; Rodriguez-Cuenca, Borja; Alonso, Concepción; Del Val, Alberto
Alcala University, SPAIN

Anomaly detection in aerial and satellite images is important in location, reconnaissance and surveillance tasks. In remote sensing, the images which best achieve these objectives are hyperspectral imagery. These images provide a virtually continuous information of the electromagnetic spectrum, resulting in a high correlation between the bands coming in that spectrum. For this reason, one of the challenges in hyperspectral image analysis is develop an efficient method for dimensionality reduction. In this work a segmented principal component transformation for a proper band reduction is proposed.
For anomaly detection, we suggest such an analysis from an object-based perspective. This paper presents a comparison of different segmentation algorithms such as fuzzy k-means and mean-shift, in order to reduce the image to a number of regions (much smaller than the number of pixels) whose pixels exhibit a uniform spectral response. In this way, we compare the anomalous regions detected by the different segmentation algorithms.