Generation and cross-validation of large-area forest stem volume maps for northeast and southeast China using ERS-1/2 tandem coherence

Abstract

The creation of large-area forest stem volume maps of Northeast (~1,5 Million km2) and Southeast China (~3 Million km2), based on ERS-1/2 tandem coherence data, was the aim of the FOREST DRAGON 1 project. The accuracy assessment of the map products is one of the objectives of the ongoing FOREST DRAGON 2 project. The ERS-1/2 tandem datasets consisted of 223 and 407 image pairs for Northeast and Southeast China respectively and were acquired in all seasons between 1995 and 1998. ERS-1/2 tandem coherence has been shown to provide accurate estimates of forest stem volume but is also known to depend strongly on the meteorological and environmental conditions at image acquisition. Existing algorithms for large-area mapping have so far not been able to classify forest stem volume based on a multi-seasonal dataset. For this reason a new classification approach, based on synergy between the optical remote sensing product MODIS Vegetation Continuous Fields and ERS-1/2 tandem coherence has been developed for automatic and seasonal-adaptive retrieval of forest stem volume. The procedure integrates the semi-empirical Interferometric Water Cloud Model and discriminates between four stem volume classes (0-20, 20-50, 50-80 and >80 m3/ha). For the evaluation of the large-area forest stem volume maps, a special cross-comparison design based on other Earth Observation products had to be developed due to the unavailability of large-scale datasets of in-situ measurements. A multi-scale comparative assessment design with existing land cover products such as GLC2000, GlobCover and MODIS VCF product has been applied. The sampling design, based on the FAO FRA2010 Sample Design and the Degree Confluence Project, uses a 1 degree sampling grid with 10 x 10 km sample plots and is completely transferable to other large-area investigation areas. A reasonable agreement above 70% between the forest stem volume map and the land cover datasets in terms of forest/ non-forest could be achieved.