Adaptive majority filtering for contextual classification of remote sensing data

This Letter presents an adaptive contextual filter, developed by the addition of a heterogeneity rule and a confidence rule to the conventional majority filter. Experiments have been carried out using a Landsat Thematic Mapper (TM) image. Results show that the adaptive majority filter has a capability of reducing the classification errors due to spectrally mixed pixels and preserves the connection of thin features. The proposed filter needs only a moderate increase in processing time compared with the conventional majority filter.