Automatic change detection in digital maps using aerial images and point clouds

Abstract

In many countries digital maps are created and provided by the national cadastres: Usually they consist of multiple polygons, each with an exact location and shape, describing which kind of surface can be found at the position of the polygon (e. g. building, street, vegetation). They must be accurate and well maintained, as they are used by companies or authorities for purposes like urban planning or demographic statistics. However, especially cities are in a constant change. Old buildings are torn down, new buildings are built, and complete streets and neighbourhoods are changed. Monitoring these changes is difficult and identifying and updating the virtual maps is still done mostly manually today. A method is developed to detect changes on the ground and identify changes for the virtual maps automatically using machine learning approaches. As input data the virtual map, their corresponding aerial images and point clouds from different years are needed. As a case study for this thesis, this method is developed and applied to the BGT, the Dutch virtual map with a resolution of 20cm. The research area is the city of Haarlem for 2017 and 2018. High resolution aerial images are used in combination with point clouds created by Photogrammetry. The output is again a digital map of the area where every polygon has a probability score of how likely its category (for example building, street, etc..) changed. This can support the manual updating process eminently, as a minor percentage of polygons (for which the algorithm was unsure) must be checked manually. The research question of this thesis is to check whether this change detection is feasible even for highly heterogeneous structures like cities. Many visual changes in the aerial images are happening that are not relevant for the virtual map. In the one year, a street can be full of colourful cars, in the other year, the street is empty and completely grey. Many scenes are easy for humans to distinguish but are challenging for an algorithm. The goal is to detect a high amount of true changes while keeping the number of false positives low to reduce the manual work as much as possible. To achieve good change detection and answer the research questions, the machine learning library of XGBoost is used. It provides a gradient boosting framework for many different environments, including Python. Many weak learners, each classifying a change only with a very low detection rate (for example minimum height of all points inside the polygon), are combined to get a strong learner. This learner should be able to classify polygons with a high accuracy into polygons that change and polygons that do not change. With this method it is possible to detect a high amount of changes. 80% of all changes can be found within a reasonable number of False positives. Especially for buildings almost all changes can be identified. It is furthermore possible to localize the changes in larger polygons. However, not all changes can be identified, so that this approach should be seen as an aid for manual change detection and not to replace it.