Development of a deep learning assisted webgis framework for updating building databases

The diversity of the methods for geodata collection has increased significantly in recent years. The need for automated approaches for updating geodatabases has also increased in parallel to this development. In addition to the novel machine learning (ML) methods, the contributions of non-professionals and volunteers are immensely required to achieve this goal. Geographical Information Systems (GIS) on the web and on mobile devices provide the required tools and methods for utilizing geoinformation (GI) contributed by individuals of all backgrounds. Thanks to the increasing attention on the volunteered geographic information (VGI) approaches and the Citizen Science (CitSci) projects, the contributions of non-professionals to participatory GIS efforts can be utilized gradually. On the other hand,to improve the accuracy of data obtained by volunteers, new algorithms and platforms can help with semi-automatic GI extraction. By incorporating novel artificial intelligence (AI) methods, such as the deep learning (DL) algorithms, into WebGIS, the semi-automatic GI extraction task can be facilitated. As a result, volunteers with limited experience on GI collection and in particular image interpretation can be assisted in performing proper processing and making informed decisions with AI guidance. In this thesis, a DL-assisted WebGIS framework was developed to detect buildings, to delineate their rooftop boundaries, and to compare with existing building vectors to serve for change detection, and accordingly database updating. The input aerial and satellite images provided by the users can be processed by using pre-trained DL models for building detection. The detected rooftops can be vectorised in the proposed system, and a change detection component reveals the alterations between the existing and the detected vector data. Thus, the framework supports vector modification and drawing, and final products are stored in a spatial database management system (DBMS). The framework is adaptable and various DL methods can be integrated into the framework for different image segmentation problems. It is expected that with the availability of such systems, more users can support the geodata collection, updating and analysis processes, which are crucial for different applications such as environmental monitoring, spatial planning, digital twin creation for land management and simulations, etc.