Using vegetation reflectance variability for species level classification of hyperspectral data. 6q2a2q

Título	Using vegetation reflectance variability for species level classification of hyperspectral data.
Autor	Mark A. Cochrane
Ano de publicação	2000
DOI	https://doi.org/10.1080/01431160050021303

Cochrane, M. A. (2000). Using vegetation reflectance variability for species level classification of hyperspectral data. International Journal of Remote Sensing, 21(10), 2075–2087.

Resumo 5vp42

Raw hyperspectral data that were acquired over Israel in 1989 by the GER 63-channel scanner were processed to provide surface reflectances seven years after the flight. Because no ground data measurements were available for the time of the flight, four atmospheric correction methods were applied: Atmospheric REMoval program (ATREM), Internal Average Relative Reflectance (IARR), Flat Field (FF) and Empirical Line (EL). Neither the ATREM program, which is an atmospheric model-based method, nor the IARR or the FF techniques, which are scene-dependent methods, were able to provide reasonable results. Whereas the failure of the ATREM program was probably because of a sensor’s radiometric problem in the visible (VIS) region, the IARR and FF methods failed because of the relative complexity of the landscape. Of the three EL combinations examined, only one was able to convert the raw digital data into reasonable apparent reflectance information. Processing the data with this combination resulted in a good match between the spectra of selected targets taken from the image and their associated laboratory spectra. Samples, which were collected seven years after the actual flight, were used to assess the ability of each EL correction method to remove atmospheric attenuation. It was concluded that when working with such so-called ‘hopeless’ data, in order to obtain reasonable results, several combinations of the EL method need to be applied. The results of each method should then be judged, from both a spectral and a spatial perspective, against a separate set of samples, which were not a part of the correction procedure. In this case study, the spectral examination involved 20 samples, and the spatial examination involved two irrigated cotton plots. In the spectral examination, good agreement was obtained between the corrected spectra and the laboratory spectra. During the spatial examination it was possible to distinguish between two cotton plots having different soil water statuses by applying the Spectra Angle Mapper (SAM) classifier. It is felt that careful selection of samples is a prerequisite for achieving reasonable results using the EL correction technique. The samples should consist of albedo information representative of the study area, should have only minor changes related to the age of time, and should be precisely identified on both the image and the ground. It was concluded that even a ‘hopeless’ raw data set, such as the current GER data, can be processed to yield reasonable physical information. Assuming that future hyperspectral data taken from orbit will not often be followed by simultaneous ground measurements, the results of this paper are promising.