Shape complexity analysis of building features with shape indices and ensemble learning classification algorithms

Abstract

Shape analysis is used to characterize spatial phenomena/objects and reveal spatial patterns in various fields such as computer vision, geographic information science, cartography, remote sensing, urban morphology, land management and ecology. In this context, shape indices quantitatively describe the morphological properties such as complexity and similarity usually through the metric properties of the geometries of spatial features and/or the auxiliary geometries derived from them. However, shape indices measure different shape characteristics of spatial features. Therefore, the use of a single shape index is not always sufficient when characterizing a feature morphologically. In addition, it is also important to use appropriate classification methods for this purpose. In this study, using circularity, convexity and rectangularity shape indices together with random forest and gradient boosted ensemble learning algorithms, 300 building features were classified as simple, moderate and complex by their shape complexity. Compared to the benchmark data generated based on visual perception, the random forest algorithm produced 93.33% overall accuracy (kappa = 0.900) while the gradient boosting algorithm produced 92.33% overall accuracy (kappa = 0.885). These findings showed that the shape complexity levels of building features can be categorized with quite high accuracy if various shape indices are used in conjunction with widely used ensemble learning classification algorithms.