Using Satellite Imagery and Computer Vision to Improve Underwriting and Risk Selection in Churches

Balkerne used Computer Vision techniques to achieve a data set of the presence of lead on individual Church, the amount of lead and the location of the lead on roofs for an individual Diocese of the Church of England. The amount of lead was categorised by physical weight, financial scrap value and financial replacement value. The location of the lead was categorised by named areas of a Church.

The core hypothesis behind this project, was the belief, that is was possible to achieve a validated, scalable, data led approach to the identification of key factors in the vulnerability of Churches to the theft of lead from their roofs.

Balkerne used web scraping techniques to create a data set of 250 known Church lead theft incidents in England for the period 2014 to 2019. A desk-based review of incidents was conducted, together with interviews with Church Wardens of a select number of Churches which had been subject to roof lead theft. Balkerne was also fortunate to receive the benefit of the extensive professional knowledge and experience of investigating acquisitive organised crime from Mr R Edwards, a former Police Chief Inspector.

This initial research phase refined the hypothesis on what data points had to be collected. A sub-hypothesis that the activities of the criminals were heavily shaped by key geospatial factors both in the vicinity of a Church and in the immediate area of a Church gained early substantial validity. This informed the selection of data points to be collected at each Church and integrated with the existing lead roof data.

Scaling the conduct of consistent data gathering across 750 Churches became a crucial requirement. Balkerne returned to the use of a web mining service and made amendments to enable the extraction of all building and infrastructure polygons within a set radius of the centre mass of an individual Church. These area co-ordinates were then used through means of an API, to extract satellite images of individual Churches and the area in the immediate vicinity. From the integrated polygon and satellite area data, a key data point for each Church could automatically be extracted and structured.

Local open source geospatial and socio-economic data was also identified, extracted and structured to associate with an individual Church. Efforts to develop a method of using Computer Vision to identify the presence of a particular feature in the immediate vicinity of a Church were unsuccessful and Balkerne was forced to implement a manual image mark-up and geospatial data capture process for each Church. Although time consuming, the data created was reliable and accurate. This manual approach also enabled the validated structuring of all data points to each individual Church.

Concurrent to the collection and preparation of the data, Balkerne engaged with Dr Hongsheng Dai of the Department of Mathematical Sciences of the University of Essex. Dr Dai advised on the options of the application of different statistical models and methods of conducting the modelling. Assessments of different models by Balkerne and Dr Dai, resulted in the decision to use a logistic regression model.

Stepwise selection was conducted, requiring analysis at each step to determine the contribution of the individual predictor variables. This enabled understanding of the contribution of the previous variables and the identification of variables to retain or delete, based on their statistical contribution. The most insignificant variables were removed as analysis was re-run which identified the most important covariates for determining the vulnerability of a Church to theft of roof lead. The model was then run multiple times and all outputs reviewed and consolidated. These outputs were then compared to known Church Roof Lead incidents in the Diocese over the period 2014 to 2019.

The modelling demonstrated that the likelihood of the Church being subject to the theft of lead substantially increases when lead is present on a Porch and notably increases when lead is present on Side Aisles. Furthermore, the existence of two specific geographic features in the immediate vicinity of a Church also contribute to an increase likelihood of lead theft.
Critical limitations of the project were that Balkerne did not have access to any data on the existence of Forensic Marking, Roof Alarms, CCTV, Security Lighting, or other risk mitigation measures present at individual Churches.

A culminating review of the project highlighted the potential for further research on the geospatial-temporal patterns of Church lead theft and also relationships between the conduct of criminal activity and both hyper local weather and night time illumination levels.
The knowledge of Church vulnerability to Roof Lead theft at both the individual Church and Diocese level can be used to inform, encourage and assist procurement decisions of risk mitigation measures and also direct behavioural actions and procedures to be followed by relevant personnel.