Hyperspectral image analysis using genetic programming
Introduction
Remote sensing using aircraft and satellite photography is well-established technology. The use of hyperspectral imagery, however, is relatively new. Hyperspectral images are capable of precisely capturing narrow bands of spectra through a wide range of wavelengths. Since many organic and inorganic materials exhibit unique absorption and reflection characteristics at particular bandwidths, these spectra are useful for remotely identifying various materials and phenomena of interest. This is an important area of work, since hyperspectral data permits the discovery of valuable natural resources in areas largely inaccessible by foot. Literally any area of the Earth can be mapped by hyperspectral imagery, be it with aircraft or satellites.
One complication in using this technology is the time and expertise required to interpret the data. Hyperspectral imaging systems such as the NASA/JPL AVIRIS1 sensor can capture over 200 bandwidths for a single geographic location [8]. This is denoted by a hyperspectral cube, which takes the form of many hundreds of mega-bytes of information. Interpreting this massive amount of data is difficult, especially considering that the spectra obtained represent mixed spectral signatures of a variety of materials. Moreover, noise and other unwanted effects must be considered. Deciphering this enormous volume of cryptic data is therefore next to impossible for humans to do manually.
This paper uses genetic programming (GP) to evolve mineral classifiers for use on hyperspectral images. Separate mineral classifiers are evolved for three specific minerals—buddingtonite, alunite, and kaolinite. The classifiers take the form of programs which, when given a vector of data from a particular pixel location on a hyperspectral cube, determine whether the mineral of interest resides there or not. Evolution proceeds by evaluating the performance of classifiers on positive and negative training sets. In addition, given the effects of noise at low reflectance levels, separate threshold stages are examined. This is done to investigate whether more accurate classification arises at higher reflectance levels, where there are more intense mineral concentrations.
Section 2 reviews concepts in hyperspectral imaging. The experimental design is outlined in Section 3. Section 4 presents the results of the experiments. A discussion and comparison to related research concludes the paper in Section 5.
Section snippets
Background
The AVIRIS data used in this study was taken over Cuprite, Nevada on June 12, 1996 (19:31UT). The sensor acquires data in the wavelength region from 0.38 to 2.50 μm, with a ground sampling interval of 16.2 m across track (horizontal) and 18.1 m along track (vertical). At-sensor radiance data were converted to surface reflectance via an atmospheric correction using the MODTRAN3 radiative transfer (RT) code, as implemented in the imaging spectrometer data analysis system (ISDAS) [19]. This removes
Hyperspectral data preparation
The reflectance data from Cuprite derived from the AVIRIS hyperspectral data set was analyzed by [14]. The mineral fraction maps which resulted from their work are used as the training solution for this study. From the full AVIRIS bandset available, we started with data at 0.428 μm and eliminated bands near 1.4 and 1.9, where strong absorption in the atmosphere occurs due to water vapour. This left 168 bands of data as input for our GP experiment.
Training set sampling
The general goal is to evolve a separate
Results
Table 4 shows the training and testing performances for the GP runs. Every mineral and threshold experiment combines the results for 30 runs (three target languages, 10 runs per language). Cross-validation is performed on the testing set, which is the remainder of the input data excluding the training pixels. The testing “avg overall” value denotes the overall classification performance of the single best solution from each run, averaged for all 30 runs. The performance of the single best
Conclusion
The hyperspectral mineral identifiers evolved by GP work quite differently from conventional approaches. With least-squares spectra fitting, signature spectra for materials of interest are fitted to the hyperspectral values at each pixel on the map [4]. Identification entails exaggerating the signature differences between materials, and looking for such fluctuations in the hyperspectral data. Our system evolves classifiers that use spectral characteristics directly from the hyperspectral data
Acknowledgement
Thanks to anonymous referees for their constructive comments. This research is supported through NSERC Operating Grants 138467 and 0046210, and an NSERC USRA grant.
References (21)
- et al.
Imaging spectrometry and the airborne visible/infrared imaging spectrometer (AVIRIS)
Remote Sens. Environ.
(1998) - D.P.K. Aguilar, D.P.M. Cobo, D.R.P. Utrero, D.M.A.H. Nieves, Abundance extractions from AVIRIS image using a...
- S.P. Brumby, J. Theiler, S. Perkins, N.R. Harvey, J.J. Szymanski, Genetic programming approach to extracting features...
Artificial neural networks for land-cover classification and mapping
Int. J. Geogr. Inform. Syst.
(1993)- et al.
Mapping minerals, amorphous materials, environmental materials, vegetation, water, ice, snow, other materials: the USGS tricorder algorithm
- et al.
Algorithm discovery using the genetic programming paradigm: extracting low-contrast curvilinear features from SAR images of arctic ice
Advances in Genetic Programming II
(1996) Classification of land cover using optimized neural nets on SPOT data
Photogr. Eng. Remote Sens.
(1993)- et al.
An evaluation of some factors affecting the accuracy of classification by an artificial neural network
Int. J. Remote Sens.
(1997) - et al.
Parallel evolution of image processing tools for multispectral imagery
- et al.
A staged genetic programming strategy for image analysis