Publications

Koukal, Tatjana; Atzberger, Clement; Schneider, Werner (2014). Evaluation of semi-empirical BRDF models inverted against multi-angle data from a digital airborne frame camera for enhancing forest type classification. REMOTE SENSING OF ENVIRONMENT, 151, 27-43.

Abstract
Forest mapping based on remote sensing data usually relies on purely spectral information ignoring that the observed signal may be substantially influenced by angular effects. On the other hand, it is known that for forest canopies the variation of reflectance with the sun-view-geometry is significant. The study examines different approaches to extract spectro-directional information from airborne imagery and explores its use in forest type classification. The images were acquired in a standard aerial survey. They were taken with a common forward and side overlap with the result that each point on the ground is observed from several view directions.To obtain directional information that might be useful in forest type classification, two widely used semi-empirical models of the bidirectional reflectance distribution function (BRDF) were tested, i.e., the Rahman-Pinty-Verstraete model (RPV; with 3 and 4 parameters) (Rahman et al., 1993) and the RossThick-LiSparse model (RTLS; with 3 and 5 parameters) (Wanner et al., 1995). For each sample plot, observations from at least 10 view directions were available. The models were inverted against the observations applying a look-up table approach. The estimated model parameters were used 1) directly as explanatory variables in the classification, and 2) to estimate the bidirectional reflectance factor (BRF) for selected sun-view-geometries, referred to as modeled BRFs. The BRFs also served as explanatory variables in the classification. For the classification the Random Forests classifier was used.Both models proved to provide information that can be successfully used in forest type classification outperforming conventional (single-angle) multi-spectral data. The 3-parameter version of the RPV model was found most effective. Compared to the conventional dataset, the overall classification accuracy increased from 72% (kappa: 0.64) to 85% (kappa: 0.81), when the parameters of this model were used as explanatory variables. The 4-parameter version of RPV and the 5-parameter version of RTLS fitted the observed data more accurately than the corresponding 3-parameter versions. However, the classification accuracies obtained with the estimated model parameters were significantly lower than those obtained with the 3-parameter models. All tested models yielded higher classification accuracies with the modeled BRFs than with the model parameters. Again, the highest classification accuracy was obtained with the RPV-3P model (overall accuracy: 92%; kappa: 0.89). The classification accuracy varied with the sun-view-geometry. These differences could partly be explained by the angular sampling scheme and partly by the models' capability to estimate the reflectance at angles that were not observed. (C) 2014 Elsevier Inc. All rights reserved.

DOI:
10.1016/j.rse.2013.12.014

ISSN:
0034-4257