A euclidean distance transformation for improved anomaly detection in spectral imagery

Remotely sensed spectral imagery is used in many disciplines, including environmental monitoring, agricultural health, defense and security applications, astronomy, medical imaging, and food quality assessment. The basic tasks performed within any of these fields are target or anomaly detection, classification or clustering, change detection, and physical parameter estimation. Hyperspectral image (HSI) analysis in any field often involves mathematically transforming the raw data into a new space using Principal Components Analysis (PCA) or similar techniques. The dimensionality of this new space is usually smaller than the collected space and as a result reduces computation time of subsequent algorithms. Additionally and more importantly, the results of standard algorithms may perform better in this new, uncor-related space. Many of the currently used transformations in HSI analysis are statistical in nature and therefore place Gaussian or similar assumptions on the data distribution. These assumptions work well with remote sensing imagery with low spectral and/or spatial resolution. In low resolution imagery, each pixel is a mixture of many materials and the data distribution is often sufficiently represented by statistical distributions. However, as the current generation sensors typically have higher spatial and/or spectral resolution, the complexity of the data collected is increasing and these assumptions are no longer adequate. As a result, algorithms based on these statistical transformations do not necessarily provide improved results on modern datasets. A new, data driven, mathematical transformation is presented as a preprocessing step for HSI analysis. Termed the Nearest Neighbor Transformation, this new transformation does not rely on placing assumptions on the data and may improve analytical results from standard HSI anomaly detection algorithms.