3910 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 7, JULY 2013 Building a Forward-Mode Three-Dimensional Reflectance Model for Topographic Normalization of High-Resolution (1–5 m) Imagery: Validation Phase in a Forested Environment Stéphane Couturier, Jean-Philippe Gastellu-Etchegorry, Emmanuel Martin, and Pavka Patiño Abstract—The aim of the topographic normalization of re- motely sensed imagery (TNRSI) is to reduce reflectance variability caused by steep terrain and, subsequently, to improve land-cover classification. Recently, multiple-forward-mode (FM) (MFM) reflectance models for topographic normalizations of medium- resolution (20–30 m) satellite imagery have improved the clas- sification of forested covers with respect to more conventional topographic corrections. We propose an FM 3-D reflectance (FM3DR) model, based on the Discrete Anisotropic Radiative Transfer simulator, for the topographic normalization of high- resolution (1–5 m) imagery. The feasibility of this approach was first verified on real IKONOS imagery for three forest types within major biomes (oak, pine, and high tropical forest) in Mexico. Next, we formalized the topographic normalization performance index and variability as relevant criteria to test TNRSI across incident angles in terms of maximum likelihood classification effective- ness. The FM3DR model outperformed five previously published topographic corrections (cosine, Minnaert, sun–canopy–sensor (SCS), Civco two-stage, and slope matching corrections), and image-based statistical strategies (Civco two-stage and slope matching corrections) tended to perform better than more analyt- ical strategies (cosine, Minnaert, and SCS corrections). An asset of this approach versus former models is the realistic account of terrain-related variation of understory and crown cover within a cover type. On top of that, once validated across forest types, the model is sufficient for the application of a full MFM 3-D reflectance-based topographic normalization without additional field measurement. Manuscript received January 27, 2012; revised July 22, 2012; accepted September 7, 2012. Date of publication February 13, 2013; date of current version June 20, 2013. This work was supported by the National Council of Science and Technology (CONACYT) in Mexico under Projects 38965T and CB-152747. S. Couturier is with the Laboratorio de Análisis Geoespacial, Instituto de Geografía, Universidad Nacional Autónoma de México (UNAM), 04510 Mexico, Mexico (e-mail: andres@igg.unam.mx). J.-P. Gastellu-Etchegorry is with the Centre d’Etudes Spatiale de la Biosphère, Centre National d’Etudes Spatiales/Centre National de la Recherche Scientifique/University of Toulouse III (Paul Sabatier University)/Institut de Recherche pour le Développement, 31401 Toulouse, France, and also with the University of Toulouse III (Paul Sabatier University), 31062 Toulouse, France (e-mail: gastellu@cesbio.cnes.fr). E. Martin is with Magellium, 31520 Ramonville Saint Agne, France (e-mail: emmanuel.martin@magellium.fr). P. Patiño is with the Centro Universitario de la Costa Sur (CUCSUR), Universidad de Guadalajara (UdG), 48900 Autlán de Navarro, Jalisco, Mexico (e-mail: pav_k@oikos.unam.mx). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2012.2226593 Index Terms—DART, forest classification, remote sensing, rough topography, topographic correction, tropical forest. I. I NTRODUCTION T HE CLASSIFICATION of high-resolution (1–5 m) re- mote sensing images is increasingly useful for detailed land-cover and habitat mapping as well as for the assessment of biodiversity and landscape dynamics. Case studies of auto- matic classification have already proved encouraging in distin- guishing forested land covers [35], [37] and even individual tree species [6] when focused at crown scale. However, like most case studies at crown scale, these deal with forest cover on relatively flat terrain. Nevertheless, biodiversity assessment and habitat description are important on rugged terrain, where access is more difficult and forest patches remain in spite of deforestation trends. Correction methods to reduce the topographic effect of rugged terrain in visible–near-infrared (NIR) (VISNIR) re- motely sensed imagery include the cosine method [30], [34], the Minnaert correction [16], [23], [36], the two-stage normal- ization method [5], and the sun–canopy–sensor (SCS) correc- tion particularly designed for forested land cover [39]. Other empirical image-based topographic corrections were applied to high-resolution imagery and tested on aerial photography (e.g., [33]) and IKONOS imagery (slope matching [24]). All these studies report enhanced classification accuracy with respect to other methods on a specific site, but cannot provide information on the upper limit of the classification accuracy which might characterize these topographic corrections, as a function of environmental and image acquisition conditions. Physical modeling of the remote sensing signal allows ex- ploration of a wide range of environmental settings, including terrain conditions. At scales well above tree crown, the average signal above a forested scene is dominated by the macroscopic properties of illuminated and shadowed crowns and ground components, modeled as subpixel components by canopy scat- tering simulators (e.g., [8], [9], and [21]). Soenen et al. [31] used the multiple forward mode (FM) (MFM) of the Li–Strahler geometric optical mutual shadowing (GOMS) reflectance model to derive the forest pixel signal on flat terrain on 30-m resolution imagery. At this scale, the topographic normaliza- tion based on the MFM 3-D reflectance (MFM3DR) model outperformed empirical topographic corrections. No such 0196-2892/$31.00 © 2013 IEEE Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on March 15,2021 at 21:32:36 UTC from IEEE Xplore. Restrictions apply.