DEALB: A Post-classification Framework for Regionalizing Local Climate Zones in the Urban Environment

Abstract

Local climate zone (LCZ) map, an outcome of a supervised classification procedure using satellite imagery, can be generated at different landscape resolutions. Because of the large spatial extent, huge-sized satellite imagery, and fine granularity, it is difficult to analyze supervised LCZ (pixel-classified satellite image) outcome creating a scope for some post-classification tasks. In this paper, we have proposed an entropy-based directional edge algorithm for locating LCZ boundaries, named as DEALB, which creates homogeneous LCZ regions and delineates their boundaries. In DEALB, an image is initially partitioned into superpixels using directional edges considered at different angles (0°, 90°, 45°, and 135°) within a specified spatial scale. Next, similar but spatially cohesive superpixels are clustered to form large homogeneous regions. Spatial cohesiveness, which is a crucial characteristic to be considered in landscape clustering, is implemented by using the breadth-first search and deque data structure. Further, to validate the correctness and pureness of boundaries in the absence of any ground truth image, we have proposed the concept of boundary purity index focusing on spatial contrast inside and outside of LCZ regions. We have demonstrated the algorithm on LCZ classified results for heterogeneous landscape of the city Nagpur in India that has been found useful by domain experts.