Generation of a global forest/non forest map from TanDEM-X interferometric data

Abstract

Generation of a Global Forest/Non Forest Map from TanDEM-X Interferometric Data Christopher Wecklich, Michele Martone, Paola Rizzoli, Gerhard Krieger The remote sensing data that the TanDEM-X system is providing represent a highly valuable source for land classification purposes. A precise and up-to-date knowledge of the land cover information is of great importance for a wide range of scientific and commercial purposes. In particular, the identification and the monitoring of vegetated areas is critical for a variety of applications, such as agriculture, cartography, geology, forestry, global change research, and for regional planning. This paper presents an overview of the forest cover map being generated from the global TanDEM-X quicklook dataset, which are multilooked and re-quantized versions (50 m x 50 m) of the full resolution data. A detailed description of the generation method of a forest/non-forest classification map from individual TanDEM-X interferometric SAR scenes, by exploiting the volume decorrelation contribution, which characterizes forested areas, will be presented. Volume decorrelation represents the interferometric coherence loss which is caused by multiple backscattering within a volume and is predominantly affected by the presence of vegetation. Given a coherence estimate, it is possible to estimate the volume decorrelation contribution and this characteristic can be used to distinguish vegetated areas from non-vegetated ones. A clustering approach based on fuzzy logic is utilized for partitioning each pixel into two classes: forest or non-forest, by associating to it a membership value to each single class. Furthermore, the method for the mosaicking of the overlapping scenes to generate a global forest map will be presented. To accomplish the final mosaic, multiple coverages over the same area are properly combined together by performing a weighted average which takes into account quality indicators of their reliability, such as the dependency of volume decorrelation on the height of ambiguity and the SNR. The inclusion of the SNR component in the weighting logic reduces the influence of the pixels on the edge of each scene where the SNR is reduced, due to a lower antenna gain. Various additional layers are derived and applied on top of the forest/non-forest map, for example to remove urban areas and water pixels, before the pixels are finally mosaicked together. The final product will be a forest/non-forest map with a 50 m x 50 m resolution and global coverage. To conclude the paper, the validation and performance measurement approach using ground truth maps and independent sources and preliminary results will be described and discussed.