From Space to Flow: Modeling Human Mobility with Satellite Imagery and Deep Learning

This thesis presents a framework for modeling and predicting spatial demand by combining human mobility data, computer vision, and advanced spatial interaction models. Mobility traces are employed as a proxy for real-world activity, with demand estimated by quantifying flows within accurately delineated spatial units. A segmentation pipeline is applied to satellite imagery to extract refined spatial footprints of target areas. These footprints are integrated with enriched origin–destination flows that incorporate contextual features. Several modeling strategies are investigated, ranging from classical approaches such as Gravity and Radiation models to neural network–based regressors. A key contribution is the adaptation of the Deep Gravity Model, which allocates flows across competing destinations using a distributional learning objective, capturing interaction dynamics more effectively than independent regression techniques.The results demonstrate that this framework improves the accuracy of demand estimation and offers a versatile tool for applications in spatial planning, mobility analysis, and decision support.