Automatic object recognition and change detection of urban trees

Monitoring of tree objects is relevant in many current policy issues and relate to the quality of the public space, municipal urban green management, management fees for green areas or Kyoto protocol reporting and all have one thing in common: the need for an up to date tree database. This study, part of the Tree and Image research project, developed a database-driven approach for object recognition and change detection using optical imagery including contextual data from urban tree databases and topographic data. Trees are 3D objects and vary in shape throughout the season. The tree is modelled in a 2D aerial image using the point location and species information. The tree model consists of a projected crown and shadow. These projections, polygons, are used to recognise the object with NDVI and texture parameters related to the ground surface derived from the topographic map and its neighbouring segments. This resulted in classifications of trees ‘still present’ or ‘disappeared’ with an overall accuracy of 85%. However, the errors of commission and omission were quite high due to the use of an early image with no full-grown crowns, resulting in difficult recognisable trees. Detected changes can be used for further human verification or directly serve as input for database management and decision making