Enhanced discrimination of boreal forest covers with directional reflectances from the airborne polarization and directionality of Earth reflectances (POLDER) instrument

During the Boreal Ecosystem-Atmosphere Study (BOREAS), directional and spectral reflectance measurements were acquired from May to July 1994 with the polarization and directionality of Earth reflectances (POLDER) instrument on board a NASA C-130 aircraft. The instrument has a wide field-of-view optics, a two-dimensional CCD array, and a rotating wheel carrying filters in the visible and near infrared. Measurements were obtained (1) over coniferous forests at the young and old jack pine and old black spruce sites, (2) over a deciduous forest at the old aspen site, and (3) over a fen at the fen site. A prominent hot spot feature was apparent at each site, with an additional strong peak in the specular direction for the fen site. Strong variations of the bidirectional reflectance distribution function (BRDF) with sun zenith angle were observed. For a constant sun zenith angle, the variation of the BRDF of conifer stands between May and July was relatively weak. A key objective of this paper is to quantify the improvement of discrimination of various forest covers when remotely sensed directional signatures are added to the more conventional spectral signatures. The experimental protocol consisted of the following steps. First, 150 pixels pertaining to five different classes of forest covers were selected on land cover maps available in the BOREAS Information System (BORIS) data base. Second, the BRDF measurements acquired by POLDER at each pixel were adjusted against a three-parameter semiempirical BRDF model and processed to retrieve the reflectance seen in three different viewing directions. Third, the results of supervised classifications were compared on all selected pixels, using as input, either the reflectances in only one direction (this simulates the case of conventional spectral signatures), or reflectances acquired in three directions (this simulates the case of spectral + directional signatures). The results showed that when only one spectral band was used, the proportion of correctly classified pixels increased from 36-59% with one viewing direction to 64-84% with three viewing directions. When three spectral bands were considered, this proportion improved from 72-87% to 83-97%. These results demonstrate that the account of directional information enhances the ability to discriminate forest covers by remote sensing.