Uncertainty in spatio-temporal data is described as the discrepancy between a measured value of an object and the true value of that object. Common causes of uncertainty in data can be identified as errors of precision in the data measurement devices, inadequate domain knowledge of the data collector, absence of gatekeepers, etc., known in this dissertation as inherent or source uncertainties. These inherent uncertainties further vary depending on the type of data (e.g., geotagged text or image data), as well as the explicit and implicit nature of the spatial dimension in the data. Static and dynamic visualization methods have been used to communicate uncertainties. However, a gap we see in such uncertainty visualizations is that users have little to no leeway of controlling the system outcomes (e.g., by weighing in their domain expertise, controlling to what extent uncertainty plays a role in the analysis, or reducing uncertainty in the data). Visual analytics help to fill this gap by allowing the user to steer the analysis process through interaction. The challenge of uncertainty analysis with visual analytics is that we not only have to encounter the inherent data uncertainties, but also the uncertainties that keep propagating through every component in a visual analytics system (the data, data models, data visualizations, and model-visualization couplings), and through every interaction from the user. To address this challenge, this dissertation introduces a framework that defines the role of uncertainty throughout the visual analytics knowledge generation process. At each component of the visual analytics system, guidelines in terms of methods are specified for assessing the uncertainties. Following this framework, four novel visual analytics approaches are introduced that enable a user to explore, assess, and mitigate context-specific uncertainties in heterogeneous data types: image data, text data, location data, and numerical data. By enabling a strong interaction between the user and the system, uncertainties are mitigated and trustworthy knowledge is extracted, thereby bridging the gap identified in static and dynamic uncertainty visualizations. The approaches developed are evaluated against anecdotal evidence and a usability experiment.