Abstract
A method for classification of remotely sensed data having incomplete supervised data is proposed. To consider the time difference between supervised data and remote sensing data, the causalities based on Markov Random Field are utilized. By including the temporal class dependencies, the reliable classification is undertaken. It expands to a method evaluating the spatial class dependencies between neighboring pixels and the post processing. The performance of the method for land cover classification is investigated using LANDSAT TM covering Aichi Prefecture, Japan, and compared with statistical data. The results show well coincidence with verification data. When post processing is carried out, the accuracy improves to RMSE of about 1%.
