Estimating the density of urban trees in 1890s Leeds and Edinburgh using object detection on historical maps

We present a new end-to-end methodology for extracting symbols from historical maps and demonstrate an application of the method to extract details of the urban forests of Leeds and Edinburgh in the UK using Ordnance Survey maps from the 1890s. The methods presented allow tree symbols on 1:500 scale maps to be efficiently extracted, with our object detection model achieving an F 1 -score of 0.945. The results for each city are presented on the National Library of Scotland website and have been used to generate an estimate of 37 ± 1 tree symbols per hectare for Leeds in 1888–90 and 40 ± 1 tree symbols per hectare for Edinburgh in 1893–94. This is the first time that quantitative data has been obtained for historical urban tree counts in these two cities. The method presented can be expanded to other UK towns and cities and is a valuable tool for learning about the past, and changes to both the natural and built environment over time, aiding decisions on future tree planting. We discuss the process used to automate the generation of training data and to train a machine learning model to extract the symbols, comparing it with other possible models. This discussion provides context on how best to tackle similar problems of symbol extraction from historical maps and the issues that may arise in such automated analysis, alongside factors that must be considered when using historical maps as a data source. • We know very little about urban trees from the 1890s, but they can help us understand our current urban forests. • An object detection machine learning algorithm is used to extract tree symbols from historical Ordnance Survey maps. • Tree symbol locations, sizes and species are detected on digitised maps of 1890s Leeds and Edinburgh. • Edinburgh has a higher tree symbol density per hectare and a higher proportion of conifer tree symbols compared to Leeds.