A Multi-Scale Residential Areas Matching Method Using Relevance Vector Machine and Active Learning

Multi-scale object matching is the key technology for upgrading feature cascade and integrating multi-source spatial data. Considering the distinctiveness of data at different scales, the present study selects residential areas in a multi-scale database as research objects and focuses on characteristic similarities. This study adopts the method of merging with no simplification, clarifies all the matching pairs that lack one-to-one relationships and places them into one-to-one matching pairs, and conducts similarity measurements on five characteristics (i.e., position, area, shape, orientation, and surroundings). The relevance vector machine (RVM) algorithm is introduced, and the method of RVM-based spatial entity matching is designed, thus avoiding the needs of weighing feature similarity and selecting matching thresholds. Moreover, the study utilizes the active learning approach to select the most effective sample for classification, which reduces the manual work of labeling samples. By means of 1:5000 and 1:25,000 residential areas matching experiments, it is shown that the RVM method could achieve high matching precision, which can be used to accurately recognize 1:1, 1:m, and m:n matching relations, thus improving automation and the intelligence level of geographical spatial data management.