A new spatially-explicit and object-conditional land cover (LC) data set based on the FAO Land Cover Classification System (LCCS) for the GLOWA catchment area

Abstract

Within the GLOWA river basin hydrology and LC project in West Africa, a unique and new LC mapping approach was implemented, using 26 Landsat Enhanced Thematic Mapper (ETM+) data tiles from the year 2000 and corresponding 26 Landsat Thematic Mapper (TM) data tiles from the period 1990. An area of 400.000 squared kilometres was mapped, in both time frames, using an object-orientated conditional approach and a standard FAO LCCS legend. On the first hierarchical level permanent and seasonally wet flooded areas were extracted using thermal emission data, and wetness and greenness indices as a function of landscape position from 90-meter topographic shuttle (SRTM) data. Linear classes were flagged using a fuzzy convolution statistical neighbourhood filter and non-linear classes were consequently flagged and mapped using an unsupervised neural network classifier. A preliminary statistical accuracy assessment was performed using independently collected low to moderate resolution LC data from the same observation period. The neural network classifier was effectively used to discriminate woody from herbaceous cover and shrub from other woody classes, especially in very open savanna areas. Most errors of commissions/confusion were found between grass and very open woody, as well as between managed woody and woody closed areas.