Textural and contextual land-cover classification using single and multiple classifier systems.

摘要

The objective of this study was to improve the quality of the digital land-cover and land-use classification when using high-resolution (10 to 30 m) remote sensing data. Three classification techniques were compared, which can be divided into two groups: single classifiers (a five-nearest neighbour and the C4.5 decision tree classifier) and multiple classifier systems (BAGFS). Textural and contextual features (roads, hydrology, relief, etc.) were introduced during the classification process. Elevenland-cover categories, in a Belgian varied landscape, were analysed and classified using Landsat Thematic Mapper data. The eleven classes includes continuous urban fabric, discontinuous urban fabric, industrial or commercial units, road networks, rail networks, arable land: cultivated soil, arable land: without vegetation, pastures, broadleaved forest, coniferous forest and water bodies. The accuracy assessment increased with the introduction of textural features and contextual data, between 0.60 and 0.82 for the Kappa coefficient. The best kappa value was achieved using numerous textural and contextual features with the multiple classifier system (BAGFS).