A unique state-of-the-art process for exploiting hyperspectral satellite imagery, based on evolutionary computing methods, has been developed and a proof-of-concept demonstration has been conducted. This development is projected to lead to several important commercial products, including a fully integrated, high-payoff, user-friendly software package — the Integrated Hyperspectral Imagery Analysis Toolbox. This software would be capable of end-to-end processing of industrial and governmental hyperspectral satellite image data with extensions to several popular commercial software products like ENVI from Research Systems, Inc., and ESRI's ArcView Geographical Information System. The development of the end product will focus on accurate detection and identification of natural and artificial materials and objects, the use of large libraries of laboratory reference data, and ease of use. Potential commercial applications include assessment of crops (including estimation of crop yields), exploration for minerals and oil, planning of military missions and automated identification of military targets, urban planning, environmental assessment, and search-and-rescue operations.
The process effected by the proof-of-concept version of the software is the following: A spectral band signature is extracted from a hyperspectral satellite image and filtered to remove noise. It is then normalized to remove global gain differences. Next, an artificial neural network identifies those categories of objects and materials that correlate with the sensed data. The categories are expressed as orthonormal feature vectors derived from training signatures that were preprocessed in a manner similar to that of the sensed data. Finally, an evolutionary algorithm processes the output of the artificial neural network, detects the relevant materials, and estimates the amounts of the materials.
A computer program that performs an end-to-end computational simulation of the process has been developed. This program is capable of accepting real hyperspectral image data, performing the requisite processing on the data, and providing a graphical display of the results. The ENVI software, which provides a computational environment for rapid prototyping of other software, was used to facilitate the development of the simulation program. In the simulation, high performance in detecting and identifying materials in the terrain, including materials with spectrally mixed signatures, was demonstrated. Terrain-classification maps were generated, illustrating how signature variations can be handled, given terrain variations. Also demonstrated was an excellent capability to discriminate between strongly similar, but different, types of vegetation.
Further efforts are planned to accomplish the following:
- Development and execution of a data-collection plan (which would include the collection of ground-truth data) involving commercial and governmental sensors.
- Extension of public-domain atmospheric-compensation methods to include terrain topography.
- Updating of algorithms for improved performance against mixed spectral signatures, including signatures of unknown materials.