Fine-grained building attributes mapping based on deep learning and a satellite-to-street view matching method

Street view images (SVIs) are a kind of data with rich semantic information, which have unique advantages in the fine-grained recognition of building land use. Compared with seamless dense remote sensing image data, SVIs are sparse and unevenly distributed in space, which brings many challenges to the application of SVIs for urban mapping. To solve this problem, this study proposes a satellite-to-street data matching method between SVIs and building footprint data. This method first performed dense sampling on the nodes of building footprint vectors, then designed a constraint based on the spatial relationship of cross-view data to match the buildings recognized in SVIs with their corresponding building footprints. Based on the matching results, large-scale building scale land use mapping was conducted in the validation area. The experimental results show that the accuracy of matching can reach more than 80%. The building land use classification in the mapping result reaches an accuracy of 62.15%, 56.41%, and 0.535 for overall accuracy, F1-score, and Kappa coefficient, respectively. This study provides a new technical means for fine-grained urban land use recognition and mapping, which can effectively improve the efficiency of acquiring fine-grained attribute information of urban buildings.