Mapping urban problems with AI: Published Research
This thesis was completed towards an MSc in Geoinformation Technology & Cartography at the University of Glasgow in 2019 and is about improving urban monitoring towards representation of cities as living and breathing entities. It was inspired by the difficulties pedestrians are faced with in Athens and other big cities as they go about their everyday lives.
This is the project presentation at the 5th International Conference on Urban e-Planning and project abstract. The project received the best thesis award for the class.
Abstract
A good urban quality of life, including urban mobility and public space walkability, is becoming harder to attain in growing contemporary cities facing diverse and complex problems. Pedestrians globally face obstructions, pollution and risk of accident. There are numerous demonstrable benefits in improving the effectiveness of urban management, which is widely under budgetary pressure and often ends up reactive and inconsistent. New standardised urban monitoring methods are needed with a focus on sustainable planning.
This project shows that the combination of automatic object detection in remotely sensed imagery with existing geospatial data can support the solution of urban problems. Remotely sensed imagery provides prompt, uniform data with city-wide coverage, while the recent advances in computer vision enable fast and accessible automatic interpretation, and established cartographic methods can generate and visualise analytical insight. Urban detection promotes cost effective, evidence-based, transparent and participatory planning, and contributes to the United Nations 2030 Agenda goals.
In this proof of concept a workflow was developed for deriving insight on the restriction of pavement accessibility by inconsiderately parked vehicles in target areas. The semi-automatic production was achieved of large scale urban map reports of 200 by 200 m areas, featuring imagery-derived car detections, urban topography and useful geospatial metrics. The maps are usable by urban planners and field surveyors.
Two examples were provided for Paisley and Birmingham. Complementary analysis involved the calculation of pavement ‘load’, i.e. overlap with adjacent car detections, and its temporal change. The presented workflow uses orthorectified aerial imagery of 0.25 m spatial resolution and can be replicated using open software. The detection algorithm achieved a workable F-score of 70.72% but could not be adapted for processing 0.46 m satellite imagery. Detection accuracy was found to depend on input image histogram and deteriorate rapidly with loss in image quality. Standard graphics equipment would render a two miles wide urban area in under one hour.
Further research should examine the determinants of object detection accuracy in remotely sensed imagery, and assess other opportunities for cartographic insight. Future implementation should focus on satellite detection, addressing temporal inflexibility and the quantification of urban mobility via walkability indices.