Detection of individual trees in urban alignment from airborne data and contextual information: A marked point process approach

With the current expansion of cities, urban trees have an important role for preserving the health of its inhabitants. With their evapotranspiration, they reduce the urban heat island phenomenon, by trapping CO2 emission, improve air quality. In particular, street trees or alignment trees, create shade on the road network, are structuring elements of the cities and decorate the roads. Street trees are also subject to specific conditions as they have little space for growth, are pruned and can be affected by the spread of diseases in single-species plantations. Thus, their detection, identification and monitoring are necessary. In this study, an approach is proposed for mapping these trees that are characteristic of the urban environment. Three areas of the city of Toulouse in the south of France are studied. Airborne hyperspectral data and a Digital Surface Model (DSM) for high vegetation detection are used. Then, contextual information is used to identify the street trees. Indeed, Geographic Information System (GIS) data are considered to detect the vegetation canopies close to the streets. Afterwards, individual street tree crown delineation is carried out by modeling the discriminative contextual features of individual street trees (hypotheses of small angle between the trees and similar heights) based on Marked Point Process (MPP). Compared to a baseline individual tree crown delineation method based on region growing, our method logically provides the best results with F-score values of 91%, 75% and 85% against 70%, 41% and 20% for the three studied areas respectively. Our approach mainly succeeds in identifying the street trees. In addition, the contribution of the angle, the height and the GIS data in the street tree mapping has been studied. The results encourage the use of the angle, the height and the GIS data together. However, with only the angle and the height, the results are similar to those obtained with the inclusion of the GIS data for the first and the second study cases with F-score values of 88%, 79% and 62% against 91%, 75% and 85% for the three study cases respectively. Finally, it is shown that the GIS data only is not sufficient.