Research On Land Use Extraction Based On Geographic Context In High Resolution Images

Land use is one of the most important fundamental information for national management on land resources,and it plays a vital role in strategic support in many areas such as social development,economy and environment protection.With the rapid evolution of the modernization in China,the requirement for highly accurate and swiftly updated land use data is severe.Nowadays,high spatial resolution(HR)remote sensing image is the most convenient means for wide coverage,large scale and highly accurate land cover data.The current manual interpretation to achieve land use information consumes too much labor and time to keep up with the vast remote sensing image data and great demand.However,HR images are characterized of the abundant detailed information and complex structures,which brings new challenges for automatic land use mapping.The research began with the observation of the miscellaneous appearances behaviors of land uses in HR images.It turns out that they differ from those on lower spatial resolution images in the way that even a single land use region tends to be composed of diverse terrestrial objects,making it heterogeneous and complex.This phenomenon causes serious ambiguity in land use mapping,which disabled traditional feature descriptors.Inspired from the analysis and analogue of the visual understanding process,this paper proposed the core idea for land use extraction in HR images,namely the Geographic Context.It was based on the spatial relationship,and deemed as the effective way to fully represent the complexity of the land use areas.There are abundant details on HR images,which is prune to disturb the land use mapping.Thus,the spatial relationship generalization on pixel level is first performed in an edge-preserving way.The homogeneous segments generated would replace the pixels as the basic processing units.According to the specific behaviors of Geographic Context,this paper proposed several Geographic Contextual features,namely Class Area Ratio(CAR)and Class Boundary Length Ratio(CBLR)in spectral space,the Density Level Area Ratio(DLAR)for scale representation,and the linearity for geometric properties.These proposed features stand out in the way that they are dedicated for the overall spatial distributional characteristics of land uses,not just restricted to simplex land covers.Geographic Contextual features hold great flexibility,diverse forms and direct connection to visual understanding.All these require a sophisticated metric function for similarity evaluation.This paper proposed EMD(Earth's Mover Distance)based metric function to handle the problems,and it was proved to suit the Geographic Contextual features in its adaptability and consistency with visual intuition.With the benefits of Geographic Contextual features and EMD based metric function,the research then formulated the classification strategy for highly accurate results.After analyzing the relationship between the Geographic Contextual features and visual interpretation,it was clear that the proposed features should follow the procedure which are established from simple to complex,as well as from fuzzy to precise.To be specific,it means that the combination usage of Geographic Contextual features are adopted in a stratified workflow.Meanwhile,since inappropriate neighborhood information has negative effects of disturbance,this paper brought up the idea of Ambiguous Set of samples which are used for the fuzzy homogeneous criterion of adjacent object merging.During the classification,the Ambiguous Set is employed in the neighborhood growing,and it is the key part dealing with the disturbing adjacency while guaranteeing enough neighborhood for the Geographic Context where features are extracted.Experiments were conducted on the aerial remote sensing images with 0.5m resolution acquired in Zhejiang province where the landscape is complex and divers.The results were largely consistent with the manual thematic maps,which proved that the proposed automatic methodology is effective for the heterogeneity and complexity of land uses on HR remote sensing images.