An Adaptive Noise Reduction Technique for Improving the Utility of Hyperspectral Data

Although hyperspectral data can contribute to the discrimination between spectrally similar objects in the landscape (like tree species) the noisy nature of data from airborne scanners like the Airborne Visible Infrared/Imaging Spectrometer (AVIRIS) limits the utility of this data. The minimum noise fraction (MNF) is a data transformation that is commonly used in hyperspectral image processing to align the data along axes of decreasing signal to noise ratio (SNR). Although the higher order components of the MNF have low SNRs, meaning the signal is highly degraded by noise, there is meaningful data in these MNF bands that contribute to applications such as classification. This paper proposes using a median filter with a kernel size that is inversely related to the SNR ratios of the MNF bands in order to more aggressively filter noise in bands with low SNRs. After transforming an AVIRIS dataset obtained over the Appomattox Buckingham State Forest in Virginia to MNF coordinates and applying this progressive filter, a stepwise discriminant analysis was used to select bands useful in identifying three pine species (loblolly, shortleaf, and Virginia). Using this progressive nonlinear filter resulted in a classification accuracy increase from 86% using the MNF dataset without filtration to 99% with the progressive filtration. Furthermore, the bands selected by the discriminant analysis for the filtered MNF data were generally much higher in number than the corresponding bands selected for the unfiltered MNF data.