Abstract
Actual land cover maps are a very good source of information on present human activities. It increases value of actual spatial databases and it is a key element for decision makers. Therefore, it is important to develop fast and cheap algorithms and procedures of spatial data updating. Every day, satellite remote sensing deliver vast amount of new data, which can be semi-automatically classified.
The paper presents a method of land cover classification based on a fuzzy artificial neural network simulator and Landsat TM satellite images. The latest CORINE Land Cover 2012 polygons were used as reference data. Three satellite images acquired 21 April 2011, 5 June 2010, 27 August 2011 over Warsaw and surrounding areas were processed. As an outcome of classification procedure, the maps, error matrices and a set of overall, producer and user accuracies and a kappa coefficient were achieved. The classification accuracy oscillates around 76% and confirms that artificial neural networks can be successfully used for forest, urban fabric, arable land, pastures, inland waters and permanent crops mapping. Low accuracies were obtained in case of heterogenic land cover units.