Generative Design of Geospatial Interventions With Automated Machine Learning and Bayesian Optimization

Geo-interventions are actions implemented in geographic space that attempt to alter targeted outcomes in the real world. Many common Geographic Information Systems (GIS) modelling approaches, such as clustering, regression, multiple criteria decision analysis, and cellular automata, do not focus on the concept of geo-interventions, albeit eventually using these approaches to inform decision-making that lead to the implementation of geo-interventions. This dissertation developed a user-guided approach to design and generate geo-interventions with vector geospatial data by defining a conceptual framework, designing a logical architecture, developing a web prototype, and conducting a case study for geo-interventions based on Automated Machine Learning (AutoML) and Hyperparameter Optimization (HPO) techniques. A literature review was conducted on the importance of geo-interventions, and to gather key literature for defining the conceptual framework with ideas from integrating GIS modelling, AutoML, and HPO. Next, the concept of geo-interventions was formulated into a geodata-driven optimization problem to structure methods for the conceptual framework, logical architecture, web prototype, and case study. The conceptual framework was defined by using conceptual components from the literature review, while the logical architecture used this framework to specify components and subsystems for software development. This architecture was applied to develop a web prototype for demonstrating the functionality of a geo-intervention design and generation system. A case study on reducing traffic collisions in Toronto, Canada was conducted to test the web prototype on a real-world problem. Considerations, benefits, limitations, and research implications on the conceptual framework, logical architecture, web prototype, case study, and their methods were discussed. By incorporating the concept of user-guided geo-intervention design and generation, the divide between spatially focused research and practice greatly decreases by linking spatial methods under the practicality of predicting real-world outcomes with geo-interventions.