Discovering the Way: Automated Machine Learning Improvement of Path Network Data

Paths are a dominant feature of the English countryside; whilst the health and social benefits of their use is well studied, in the literature, discussion of their spatio-temporal drivers and morphologies is limited. In exploring this research gap, this study investigates how paths’ morphologies are linked to their spatial and temporal contexts (research question one). The study is divided into two sections; firstly, paths’ spatial and historical contexts are explored (research objective one). Through understanding these contexts, this study explores approaches to replicating existing paths (research objective two). Analyses are undertaken using Python and QGIS. Whilst waiting for permission to access the Ordnance Survey (OS) Maps app GPS paths dataset, UK Rights of Way (RoW) are used to represent paths.

This study identifies the following when exploring paths’ spatial and historical contexts. Churches and roads guide path direction with paths being noted as having either functional or leisure-based purposes. In exploring paths’ spatial contexts, improved grassland and arable land-cover types are most prevalent near paths, whilst fen, heather, and inland rock land-cover types are least prevalent. Additionally, waterways and amenity OpenStreetMap (OSM) points of interest (POI) were found to play a strong role in forming paths’ spatial contexts. These analyses highlighted the importance of not assuming homogeneity of paths’ contexts across study sites. This study additionally defines and explores natural POI, establishing a presence score method which allows OS POIs to be classified based on their high/low presence and percentage scores. Understanding paths’ spatial contexts allows for the identification of datasets to use as spatial context variables in the path prediction element of this study.

Using variables identified in the first part of the study as being valuable in forming paths’ spatial contexts, this study uses logistic regression, k-nearest neighbours, and random forests-based machine learning, and shortest-paths and least-cost-paths-based analyses, to replicate existing paths. The approaches taken did not accurately replicate paths, however, as model parameters were fine-tuned, the machine learning based predictions’ quality marginally improved. The shortest and least cost paths methods produced low quality predictions. Limitations with the data and approaches taken note areas for improvement in future studies. This study identifies that physical spatial context variables alone do not provide enough information to successfully replicate paths. Additional contextual information is required, with literature recommendations noting the importance of representing historical and social drivers of paths in data in improving prediction quality (Supernant, 2017). Further exploration of historical and social drivers of paths has potential to improve prediction quality. Understanding paths spatial contexts can help organisations with decision making and in developing tools to improve spatial planning and meeting SDGs. Through the identification of new paths, updating maps for consumers and outdoor accessibility can be improved, positively affecting people’s wellbeing and physical activity. This research has potential for development and application in future studies and in industry.