ATMOSPHERIC CORRECTIONS OF AVIRIS IMAGES WITH A PROCEDURE BASED ON THE INVERSION OF THE 5S MODEL

An algorithm for atmospheric correction was developed for correcting A VIRIS (Airborne Visible and Infrared Imaging Spectrometer) images that were acquired during the 1991 Mac Europe campaign of NASA/JPL over th e 'Landes' (south-west France). The methodology is based on the invers ion of the 5S atmospheric model, through an iterative procedure that u ses the Gauss Seidel principle. The environmental effect is fully take n into account, on a pixel per pixel basis, by the use of circular nei ghbourhoods the radii of which are variable with wavelength. Here, inp ut parameters, i.e. optical characteristics of the atmosphere, are est imated with in situ atmospheric profile and visibility measurements co mbined with the 5S model. The visual analysis of AVIRIS spectral bands in the blue region clearly showed a heterogeneous spatial distributio n of aerosol effects. Consequently, a procedure was developed which co mputes the aerosol optical depth directly from the image. The only ass umption is the presence of dense dark vegetation with spectral reflect ances lying in narrow intervals the bounds of which, unknown at first, are iteratively determined. Spectral bands centred at 459.8 nm (band 7), 489.4 nm (band 10) and 607.9 nm (band 22) were the most efficient sensor's bands for that approach. The spatial variation of the aerosol optical depth was [0.16-0.26], [0.15-0.23] and [0.10-0.18] in the 459 .8 nm, 489.4 nm and 607.9 nm bands, respectively, with a mean 1.4 Angs trom exponent. Spectral aerosol optical depth maps were computed and u sed as input parameters in the atmospheric correction procedure. This converged after five iterations, for all AVIRIS spectral bands. This c orrection procedure was conducted with radii of circular neighbourhood ranging from 0 to 100 pixels, which allowed us to compare this approa ch with those procedures that do not take into account adjacency effec t or assume that the latter can be derived from the pixel value. Moreo ver, these computations allowed us to determine the minimal sizes of t he circular neighbourhoods, for each spectral band, thus ensuring a go od approximation of the adjacency effect; e.g., for the 459.8 nm band, a radius of 600 m (i.e., 30 pixels) was necessary for obtaining corre cted reflectances with an accuracy of 0.5 per cent.