Object Context Information for Advanced Forest Change Classification Strategies

This paper describes a contextual classification approach to classify complex forest change classes. Multi temporal forest change classifications are performed comparing the 1989 baseline for the Kyoto definition of ARD (Afforestation, Reforestation and Deforestation) with the status in 2000 for test territories in Siberia with extensive ground truth information from forest inventory. The differentiation of human induced forest cover change (e.g. logging and clear cutting) and changes by fires (that are potentially a nonhuman induced change) is usually very much complicated by similar spectral signatures. Tests with object based change detection approaches showed that object shape can increase the classification accuracy of these change classes. This work concentrates on contextual information that is used as a secondary information type for change class differentiation. Logging is usually only possible with infrastructure, road or path networks to transport larger amounts of timber. The existence of linear road objects can therefore be used as a prerequisite for the classification of specific change classes using neighbourhood relationships and distance measures in different scales. An analysis of the results of a direct two-date change detection and post classification procedure shows the effect for the classification of deforestation classes in Siberia with atmosphere corrected multitemporal Landsat data and forest inventory information. The presented work is part of the Siberia-II project that was finished in 2005. * Corresponding author.