Segmentation and Object-Based Land Cover Classification of Airborne Images in Kraliky County

Old airborne images still represent a challenge for effective classification of land cover due to single-band acquisition and missing the ground true. The land cover of the Kraliky county (NE of Czechia) captured by 55 orthophotos in 1953 and 2016 was classified to evaluate the long-term LC development of this peripheral region. The comparison of manual digitization, per-pixel and object-oriented classification demonstrated benefits of the last approach. The multiresolution segmentation was tuned separately for images with and without built-up areas. The object-oriented classification was focused to distinguish 5 basic classes – forest, grassland, cropland, water and built-up. To improve accuracy, the last class required a visual inspection and part reclassification. Linear features such as roads and railways were classified differently based on ancillary vector data, i.e. its visual inspection in images and modifications. The LC development of Kraliky county shows 21% increased forested area and the same level of decrease for grasslands. Built-up areas are larger by 8%, and the area of cropland remains the same despite collectivization in the 1950s. The segmentation and object-oriented classification of airborne images enabled quick statistical assessment of the long-term LC changes. Results indicate that the object-oriented classification is much more effective than manual digitization despite the possible inclusion of manual parts such as partial visual inspection and modification after object-oriented classification, and that the processing time can be reduced to half the average.