Fusion of Remote Sensing and Citizen Science Information through Machine Learning for Geospatial Analysis

Heterogeneous geospatial big data, from multi-modal Earth Observation (EO) data to geo-social media data, has become more and more accessible in recent years. This provides a potential data source for automatically extracting and mapping key geographical characteristics, hence mitigating the global mapping problem using current data mining methods. These automated geographic feature mapping techniques, especially for man-made infrastructure, are crucial to a lot of our socio-economic existence. Machine learning techniques, among many other data mining methodologies, have demonstrated better performance across a wide range of academic domains, most notably natural language processing and computer vision. In recent times, there has been a growing interest in research on ML-based Geospatial Artificial Intelligence (GeoAI), particularly in its ability to support autonomous mapping with heterogeneous geographical data. Though the potential is high and obvious, it remains a major challenge to handle inherent heterogeneity and empower data synergy when building robust and scalable GeoAI models for large-scale automated mapping purposes. For geospatial analysis, citizen science initiative is seen to be the most effective. This is a result of the fast development of Web 2.0 and crowdsourcing/Volunteered Geographic Information (VGI) technologies. These technologies enable even regular users or volunteer mappers to develop, gather, and distribute geospatial data using a variety of digital devices (such as desktop computers, mobile tablets, and smartphones). The technological obstacles to digital mapping have been significantly reduced by ongoing crowdsourcing and VGI efforts. In the real world, though, problems with global mapping have persisted for a considerable amount of time even in higher-income nations. Intelligent automated mapping techniques for geospatial analysis are desperately needed in this situation since they may effectively and efficiently close significant data gaps across nations. The research effort reported in this dissertation explores the possibilities of using citizen science or VGI to conduct geospatial analysis of different man-made infrastructures using ML from diverse geospatial data sources (e.g., multi-modal EO data, OSM, and GIS data). Three main research questions (RQs), derived from data-driven, method-driven, and application-driven research perspectives, are established to better address the issue of geospatial analysis with remote sensing and citizen science. The thesis especially goes in this direction by i) investigating the data-driven issue that combines ML for segmentation tasks; ii) creating strategies to deal with VGI data noises; and iii) using the created strategies in various mapping tasks. This creates even more intriguing possibilities for related works in the future.