Land use mapping based on composite regions in aerial images

Abstract

ABSTRACTLand use data are widely applied for the analysis and management of social, economic, and ecological conditions and are in great demand due to the rapid changes of landscapes. Remote sensing images with high spatial resolutions, especially the aerial ones, are popularly utilized as data sources for large-scale, highly accurate land use mapping. However, unlike simple land cover classes, land use areas are compound with miscellaneous terrestrial objects, which presents the automatic land use mapping with new challenges. In this study, we analysed the characteristics of land uses in high-resolution aerial images and found that the highly complex composition of land use areas causes an individual terrestrial object alone insufficient to determine its land use type. To address this issue, we proposed the concept of Composite Regions as regions formed by diverse objects sharing the same land use type. We defined Geographic Context which is the key determinant for land use mapping. Additionally, three Geographic Contextual features describing different aspects of the Composite Regions were developed: (1) Class Index Area Ratio (CIAR) for spectral distribution, (2) Density Level Area Ratio (DLAR) for scale range, and (3) Class Boundary Length Ratio (CBLR) for semantic affinity. These features take advantage of Earth’s Mover Distance as the proper metric function. The accompanying mapping strategy is formulated by applying CIAR, DLAR, and CBLR in an ordered way, and it involves Fuzzy Object Merging to ensure convincing and sufficient neighbourhood information for objects before final identification. The experiments were performed on aerial images acquired in Zhejiang province, China. The landscapes in the study areas consist of five major land use types and exhibits typical complex interiors which were used to be recognized manually. The mapping results successfully extracted land use areas that were highly consistent with visual interpreted data, despite the complex heterogeneous composition. Assessment analysis also showed high accuracies which validated the effectiveness of the methodology.