Font Size:
High Dimensional Model Representation Based Hyperspectral Anomaly Detection
Last modified: 2020-01-22
Abstract
Anomaly Detection (AD) is an important problem in Hyperspectral Imagery(HSI). However AD can be an intricate problem to solve since there is noa priori information to separate anomaly pixels from HSI background.Considering this, accuracy is the most important target in AD to avoid falsealarms. This study proposes an accurate AD algorithm using HighDimensional Model Representation (HDMR) which is a divide-and-conqueralgorithm for multivariate functions and data sets.
HDMR decomposes multiway data sets which are known as tensors. In thiscase HDMR provides working with vectors or matrices other than tensor typedata sets. For detection problem a two-stage algorithm is designedby using this facility of HDMR. The first stage is enhancing HSI withHDMR to increase distinctiveness of anomaly pixels from the background pixels. The second stage is using low-variate components of HDMR with sliding window localisation. HDMR scores of each pixel through the local window is stored to distinguish anomaly pixels from background. Algorithm has been tested both on synthetic data set and real world HSI data set by comparison with other art-of-state detectors which also use local windows.
HDMR decomposes multiway data sets which are known as tensors. In thiscase HDMR provides working with vectors or matrices other than tensor typedata sets. For detection problem a two-stage algorithm is designedby using this facility of HDMR. The first stage is enhancing HSI withHDMR to increase distinctiveness of anomaly pixels from the background pixels. The second stage is using low-variate components of HDMR with sliding window localisation. HDMR scores of each pixel through the local window is stored to distinguish anomaly pixels from background. Algorithm has been tested both on synthetic data set and real world HSI data set by comparison with other art-of-state detectors which also use local windows.