Level set segmentation of hyperspectral images using joint spectral edge and signature information

This paper describes a new method for segmenting hyperspectral imagery (HSI) using dynamic curves. We are concerned about challenging HSI target segmentation/detection use cases where the scene includes confusers exhibiting a spectral return similar to the desired signature and in close proximity of the object of interest. Our method is based on a level sets approach. It fuses all available spectral bands and incorporates spectral as well as spatial information to obtain a finer target segmentation. The proposed method applies level set segmentation to HSI by defining an expansion force field that combines both hyperspectral gradient information as well as the desired spectral signature. We carry out experiments on HSI datacubes obtained from a sensor spanning visible and near IR and show improved results when compared to direct spectral matching in challenging close range scenes including significant level of nearby confusers.

1 Introduction

Recent advances in hyperspectral sensors with high spectral and spatial resolution have led to increased interest in exploiting spectral imagery for target detection, image segmentation, and object classification [1] [4] [6] [7] [13]. HSI provides spectral information for each pixel in the image, resulting in a hyperspectral datacube. This spectrum often spans frequencies beyond the visible range including near-, mid-wave, and long-wave infrared. One advantage of the added spectral diversity is that it provides rich information that may be exploited to yield a better segmentation of the scene.