Using neural networks to map Africa's land cover with Landsat ETM+ SLC-off imagery

Landsat SLC-off imagery has been downloaded for the whole of Africa, with the imagery being acquired by the satellite during the period September 2007 to May 2008. This imagery is intended to be used for the production of a land cover map of Africa at Landsat ETM+ image resolution. The quantity of data (>1000 scenes, each containing ~40 million pixels) means that automated image analysis is required in order to achieve this. Manual identification of land cover classes has been carried out and a classification system developed that is based on the FAO's LCCS. This classification system is designed to provide a level of detail that will be useful to land managers, farmers and environmentalists alike. Prior to mapping the entire continent it is necessary to determine whether or not the method selected will produce sufficiently accurate maps. Here a test of the classification method, which uses neural network classification followed by nearest-neighbour interpolation, is described. Results show that the mapping accuracy is greater than 85% for all classes where the pixels were available, and that interpolation of missing pixels post-classification gives accuracy greater than 80% for all classes (the number of classes varied between scenes, but was generally between 5 and 10). In addition, key requirements for the development of a continental land cover map of Africa have been identified during this work.