Hierarchical Recognition for Urban Villages Fusing Multiview Feature Information

Abstract

Urban village (UV) renovation is crucial for urban renewal, with effective UV recognition serving as a prerequisite. While existing studies on UV recognition predominantly rely on high-resolution remote sensing images (RSI), and few integrate street view images (SVI), which could cause confusion in regions with similar planar features, such as old residential area and industrial parks. This article proposed a hierarchical framework for UV recognition which integrated multiview images. The spectral, textural, and structural features were extracted from Google RSI by machine-learning classifiers for each segmented block. The deep-learning method was applied to SVI to capture the architectural feature at each viewpoint. The rule-constrained fusion was conducted to combine the block-level and point-level UV recognition results. Taking a typical high-density megacity Nanjing as the study area, a high recognition overall accuracy (OA) and Kappa of 95.04% and 0.860 were achieved, identifying 172 UVs covering an area of 27.93 km² by 2020. The results demonstrated an “urban village ring” pattern in the city, with central urban areas showing a “multicenter and multicluster” spatial distribution, while suburban areas exhibited “large and concentrated” characteristics. Compared with results from single-view of RSI, the complementarity with SVI for multiview features increased the OA and Kappa by 3.34% and 0.079, which could effectively distinguish the old industrial parks. We believe that our proposed hierarchical framework is essential to the scientific and accurate UV recognition, which can guide the urban management and high-quality development.