Improving Object-Based Land Use/Cover Classification from Medium Resolution Imagery by Markov Chain Geostatistical Post-Classification

Land use/land cover maps derived from remotely sensed imagery are often insufficient in quality for some quantitative application purposes due to a variety of reasons such as spectral confusion. Although object-based classification has some advantages over pixel-based classification in identifying relatively homogeneous land use/cover areas from medium resolution remotely sensed images, the classification accuracy is usually still relatively low. In this study, we aimed to test whether the recently proposed Markov chain random field (MCRF) post-classification method, that is, the spectral similarity-enhanced MCRF co-simulation (SS-coMCRF) model, can effectively improve object-based land use/cover classifications on different landscapes. Four study areas (Cixi, Yinchuan and Maanshan in China and Hartford in USA) with different landscapes and classification schemes were chosen for case studies. Expert-interpreted sample data (0.087% to 0.258% of total pixels) were obtained for each study area from the original Landsat images used in object-based pre-classification and other sources (e.g., Google satellite imagery). Post-classification results showed that the overall classification accuracies of the four cases were obviously improved over the corresponding pre-classification results by 14.1% for Cixi, 5% for Yinchuan, 11.8% for Maanshan and 5.6% for Hartford, respectively. At the meantime, SS-coMCRF also reduced the noise and minor patches contained in pre-classifications. This means that the Markov chain geostatistical post-classification method is capable of improving the accuracy and quality of object-based land use/cover classification from medium resolution remotely sensed imagery in various landscape situations.