SPATIAL PATTERN ANALYSIS THROUGH DISTRIBUTION METRICS

Abstract. Moving from the controversial results on the link between urban structure and performance aspect, this article wants to encourage the development of the independent research on urban structure, and more generally on spatial patterns, at different scales to enable future further correlations with a wider set of performance aspects (environmental, social, economic, medical). The work also exploits the potential of several unsupervised learning algorithms, whose performance and power are increasingly promising and whose use is becoming more widespread in different fields; but for which there are still many challenges concerning the correct application in urban areas and the interpretability of the results. We propose an approach for the creation of new spatial attributes and metrics (features) aiming to quantitatively describe the qualitative distribution of objects (e.g., buildings) in a 2D space. It explores an incremental bottom-up process for the creation of groups of objects (e.g., urban patches) and the evaluation of their physical properties alone and in respect with a sample area at each iteration. The process consists of 7 phases: data preparation, data processing, parameters collection, feature calculation, feature selection, clustering, results comparison. The results can be mainly divided in two. First, the feature selection allowed to extract a minimum set of non-redundant, valid, and consistent features that can explain qualitative distribution aspects of spatial patterns. Second, the comparison between feature-based and neural network clustering, gave useful insights for a preliminary understanding of unsupervised learning techniques internal mechanisms.