Morphological characterization of landscape using context-rich geometrical features extracted from path centre lines

Understanding the morphology of connective networks can reveal hidden aspects of our cities when correlated to geographical data. A strong body of work exists on examining the relationship between space syntax analysis and socioeconomic indicators, in some cases complemented with information on urban blocks and typologies. Other authors study street morphology either as the basis of generative algorithms or detecting global groups of urban patterns across many cities. However, other aspects of urban and landscape design, such as style or character, while widely theorised and used within generative algorithms, have not been subjected to systematic geospatial exploration. This study presents an analytical method that can adequately capture the geometrical nature of urban landscape networks and correlate it with its historical style or character. In our experiment, we extracted geometric features from OSM cleaned path centre lines for 40 parks in London. This was followed by a historical survey carried out for each park, labelling lines into six distinct historical-stylistic categories. Eight classifiers were trained and evaluated through a machine learning process predicting the historical categories of lines from their geometrical features. This was repeated in three consecutive tests with a growing degree of contextual features derived from neighbouring lines. For the last test, a bespoke technique including spatial and non-spatial clustering was used to identify morphologically coherent pieces of fabric which are fed into the classifier. Classifiers accuracy range from 55% in the first test to above 90% using context-aware geometric features. This suggests that the historical development of the urban or landscape fabric leaves a footprint in the map that can be unravelled by analysing its morphology. In future studies, these methods have a great potential to be expanded further towards investigating these findings in relationship with other more sophisticated, historically dependent data and be applied to different types of networks.