Evolving feature extraction algorithms for hyperspectral and fused imagery

Abstract

Hyperspectral imagery with moderate spatial resolution (/spl sim/30 m) presents an interesting challenge to feature extraction algorithm developers, as both spatial and spectral signatures may be required to identify the feature of interest. We describe a genetic programming software system, called GENIE, which augments the human scientist/analyst by evolving customized spatio-spectral feature extraction pipelines from training data provided via an intuitive, point-and-click interface. We describe recent work exploring geospatial feature extraction from hyperspectral imagery, and from a multi-instrument fused dataset. For hyperspectral imagery, we demonstrate our system on NASA Earth Observer 1 (EO-1) Hyperion imagery, applied to agricultural crop detection. We present an evolved pipeline, and discuss its operation. We also discuss work with multi-spectral imagery (DOE/NNSA Multispectral Thermal Imager) fused with USGS digital elevation model (DEM) data, with the application of detecting mixed conifer forest.