Classifying heterogeneous urban form into local climate zones using supervised learning and greedy clustering incorporating Landsat dataset

Abstract

The concept of Local climate zone (LCZ) is used to generate LCZ maps worldwide as a part of World Urban Database and Access Portal Tools (WUDAPT) project utilizing Landsat imagery. Though WUDAPT has become the standard for LCZ mapping, it has certain limitations: 1) its outcome for heterogeneous cities is less accurate. 2) Due to fine granularity and huge-sized imagery, validating WUDAPT outcome becomes difficult for an urban expert. To overcome these, we have presented the LCZ classification and greedy clustering framework for heterogeneous cities (Nagpur (India)). Three classification schemes based on random forest, deep neural network, 1D-convolutional neural network and, subset ensemble are proposed. The overall accuracies are respectively 88%, 86%, 82%, 88%. Contextual information (mean, minimum, maximum, median, 25th, and 75th quantile values) at 90 m resolution is adopted while classification to handle spatial heterogeneity. Adding contextual information increases the overall accuracy by 4–6%, urban and mixed classes' accuracy by 4–7% and, 3–4% respectively. Further, clustering groups the spatially connected but similar pixels (classified Landsat) using breadth-first search concept and deque data structure creating LCZ map. Such LCZ map is helpful for an urban expert to validate correctness, select suitable level of granularity, visualize heterogeneity and suggest improvements.