UNDECLARED CONSTRUCTIONS: A GOVERNMENT’S SUPPORT DEEP LEARNING SOLUTION FOR AUTOMATIC CHANGE DETECTION

摘要

In our cities, in particular those with high demographic density, the proliferation of buildings goes so fast that it is not possible or -in the best scenarios- very difficult to be handled by the government departments regulating the legitimacy -in term of safety and taxes- of those constructions.In this paper, we propose a deep learning tool for computer vision trained with a corpus of satellite images provided by SpaceNet to detect changes in cities, showing the most recent constructions automatically, which allows different municipal officers to check if they have been -or haven't been- declared. To achieve this, we implemented the layered architecture of the SpaceNet Challenge Round 2 winning solution, and decided to improve it with an output comparison which gives us a high value final result for the end-user in the detection of changes, giving him the possibility to appreciate in the graphic user interface how many new buildings and square meters were detected.