Evaluation of feature extraction and reduction methods for hyperspectral images

摘要

The number of hyperspectral sensors used in remote sensing is rapidly increasing. Although the availability of hyperspectral images is widespread, there is still a lack of efficient algorithms to properly handle the data. The major problem when trying to apply traditional image classification procedures to hyperspectral data is the high dimensionality, which besides increasing computational burden can impair classification due to the Hughes phenomenon. One possible solution to this problem is the dimensionality reduction, or feature reduction, prior to the application of other image processing procedures, such as image classification. It often happens that what was initially a hyperspectral image with several tens or even hundreds of bands is reduced to a 'standard' multi-spectral image with a small number of bands.The process of feature reduction/selection is nevertheless a delicate task, as ideally one would like to preserve as much information as possible from the original dataset.In this work some of the most commonly used feature selection and reduction algorithms are tested. The Jeffries-Matusita distance is used to determine the best subset of features and some spectral transformations such as Segmented Principal Component Analysis, Segmented Canonical Analysis, Independent Component Analysis, and Minimum Mean Squared Error are carried out to reduce the data dimensionality. These methods were compared using four test images from the AVIRIS, Hymap and Hysens sensors. Although no straightforward method exists to perform an evaluation, the consistency between the various methods is discussed.