Towards better exploiting object-based image analysis paradigm for local climate zones mapping

Abstract

Local Climate Zones (LCZs) have demonstrated significant promise in urban climate research. Although an LCZ consists of diverse geographic objects, most object-based LCZ mapping research encounters difficulty and seems to have no advantage due to the huge intra-class differences of objects within an LCZ. To address this problem, we propose a new object-based LCZ mapping framework based on multi-level features, which consist of information from pixel and region, bridging adjacent objects through region-level features while preserving original image features. As only using remote sensing imagery makes it hard to fully depict urban form, we modified urban morphological parameters (UMPs) to help with mapping, including sky view factor (SVF), building surface fraction (BSF), and permeable surface fraction (PSF). In addition, a refined sampling strategy is proposed for object-based LCZ mapping to avoid unrepresentative objects. Experiments were conducted in three cities (Hong Kong, Nanjing, and Nanchang), and the results showed that the accuracy of built LCZ types (OAbu) with modified UMPs and region-level features performed best, since they performed about 20% better than those without UMPs and region-level features. Furthermore, compared with patch-based methods, it is demonstrated that the object-based strategy can more appropriately depict the boundary of an LCZ. With further analysis of results at different scales and with features extracted from different region sizes, we found that built types benefit more than natural types at fine scales, and the region size of 17 × 17 is satisfied. In fact, we expect this study would directly reverse the unfavorable situation of object-based image analysis (OBIA) paradigm in LCZ mapping.