Large-scale vegetation maps derived from the combined L-band GRFM and C-band CAMP wide area radar mosaics of Central Africa

A new dataset has been compiled by combining the wide area Synthetic Aperture Radar ( SAR) mosaics over Central Africa generated in the context of the NASDA Global Rain Forest Mapping (GRFM) and the ESA/EC Central Africa Mosaic Projects (CAMP). The CAMP mosaic consists of more than 700 SAR scenes acquired over the Central Africa region (6degrees S-8degrees N and 5degrees E-26degrees E) by the ESA ERS satellites; the acquisitions were performed in 1994 ( July, August) and in 1996 ( January, February) in two different seasonal conditions. The GRFM Africa mosaic consists of some 3900 JERS-1 images acquired over the region (10degrees S-10degrees N, 14degrees W and 42degrees E) at two dates (January-March 1996 and October-November 1996). In this paper the methods used for combining the two wide area radar mosaics are at first presented. The GRFM Africa mosaic was processed using a block adjustment algorithm with the inclusion of external observations derived from high precision maps along the coastline, which assures an absolute geolocation residual mean squared error of 240 m with respect to ground control points. On the other hand, the CAMP mosaic was compiled taking into account only the internal relative geometric accuracy. Therefore the GRFM dataset was taken as the reference system and the C-band ERS layer composed by rectifying each ERS frame, after down-sampling at 100 m pixel spacing, to the reference mosaic. The rectification procedure uses a set of tie-points measured automatically between each ERS frame and the homologous subset in the JERS mosaic. Due to the different characteristics of the two sensors (microwave centre frequency, viewing geometry, polarization) and the different acquisition dates, each mosaic presents a different window over the same ecosystem. This fact suggests that a new dimension in terms of thematic information content can be added by the fusion of the two datasets. In support of this statement, the complementary characteristics of the two sensors are first discussed with respect to observations related to the vegetation cover in the Congo River floodplain. The potential of the combined dataset for vegetation mapping at the regional scale is further demonstrated by a classification pursuit of the main vegetation types in the central part of the Congo Basin. The main land-cover classes are: lowland rain forest, permanently flooded forest, periodically flooded forest, swamp grassland, and savannah. The classification map is validated using a compilation of national vegetation maps derived from other high resolution remote sensing data or by ground surveys. This first thematic result already confirms that the combined contributions from the L-band and the C-band sensors improve the information extraction capability. Indeed, the radar-derived vegetation map contains better spatial detail than any existing map, especially with respect to the extent of flooded formations.