Digital traces, Earth observation imagery and socioeconomic records provide complementary perspectives on how people use space, how places function and how regions change. Analysing these heterogeneous sources requires methods that capture spatial context, temporal dynamics and relationships across scales. This project brings together five complementary research strands at the Department of Human Geography, Lund University, to develop and apply geospatial AI methods for understanding mobility, urban environments and regional transitions. Representation learning provides a common methodological foundation, complemented by generative modelling to investigate activity-travel behaviour and alternative mobility scenarios. At the individual level, we investigate how time–space constraints, daily routines and individual capabilities shape access to opportunities. Generative models of activity-travel behaviour incorporate these constraints to compare feasible and realised activity spaces and examine differences in experienced accessibility, segregation and spatial inequality. Scenario simulations explore how changes in transport provision and the distribution of opportunities could affect different population groups. At the place level, two research strands examine the relationship between urban environments and human activity. The first combines representations learned from satellite imagery with human mobility networks to characterise urban places through both their physical form and their functional connections. The second uses vision-language models to analyse crowd-collected photographs of local marketplaces, examining the visual and textual information through which local consumption environments are presented and communicated. At the regional level, two further strands investigate economic restructuring and green transitions. Self-supervised sequence and graph models examine connections among people, firms, industries and regions over time, with attention to how knowledge and industrial structures relate to regional adaptation. Language models analyse multilingual news archives to study how these transitions are narrated and contested across geographical and linguistic contexts. The research strands share methods for learning from complex spatial and temporal data, adapting pre-trained models and evaluating transferability beyond the settings used for model development. Validation will use held-out observations and comparisons across locations, languages and time periods. Shared software components and coordinated computational workflows will support reproducible experimentation while allowing each strand to retain its own research design. The requested resources will support model training and adaptation, systematic comparisons and large-scale inference that exceed the capacity of local computing resources. Expected outputs include research publications, reproducible analytical workflows, derived indicators and, where data and model licences permit, openly released code and models. Together, these outputs will contribute evidence and methods for urban and transport planning, regional development and sustainability research.