Remote Sensing of Environment, 2015,
171, p. 245-260 ISSN 0034-4257
Quantifying and mapping dense tropical forest structure at region to country level have become pressing needs, notably but not exclusively for assessing carbon stocks as part of the Reducing Emission from Deforestation and forest Degradation (REDD +) process. Fourier texture features from very high spatial resolution passive optical data have shown good potential as non-saturating proxies for stand parameters, including above-ground biomass, within required standards of precision and accuracy. These proxies are, however, sensitive to acquisition geometry (sun view angles), even for acquisition geometries usually in use in VHR sensors, hampering regional or multi-temporal studies combining multiple acquisitions. Our aim was to improve the understanding of this variation formalized in the bidirectional texture function (BTF), and find ways to mitigate it. We used simulated stands and the Discrete Anisotropic Radiative Transfer (DART) model, as well as a collection of Ikonos images over a forest site near Santarem (Para, Brazil). BTF proved dependent on forest structure and displayed strong anisotropy with respect to forward vs. backward scattering modes. But it remained approximately constant over a large range of angular configurations in forward mode, thereby allowing operational use without any correction. This range could be broadened by correcting bias using empirical BTF fitting or (more practically) by inter-calibrating Fourier spectra when some overlap area is available between images. Prediction of a forest structure parameter (D-max, the estimated maximum trunk diameter class) using images in varying configurations then remained unbiased and below 15% relative RMSE except in the vicinity (+/- 10 degrees in the principal bidirectional plane) of the hotspot direction. Near hotspot directions need to be proscribed, as the absence of visible shadows impedes textural description. These results, and the increasing availability of large swath VHR sensor constellations (e.g. SPOT 6-7), open the way to operational broad scale applications for forest characterization, above-ground biomass mapping and multitemporal degradation monitoring.
Plan de classement
Etudes, transformation, conservation du milieu naturel [082]
;
Télédétection [126]
