Crowd aggregation event-coupled traffic flow prediction based on geospatial sensing data

Crowd aggregation events induce significant perturbations in urban transportation systems, triggering unconventional disruptions in traffic flow patterns. Geospatial sensing technology provides critical spatiotemporal data for characterizing this complex uncertainty, which stems from the interactions between aggregation events, people's travel decisions, and the dynamic response mechanism of transportation service supply. However, effectively extracting and integrating heterogeneous features from different geospatial sensing resources remains a key challenge in capturing the underlying evolution mechanism of traffic flows, rendering existing models inadequate. To address this, we propose a heterogeneous interaction framework that disaggregates traffic heterogeneity into three components: spatial multiscales, temporal non-stationarity and uncertainty interference. A dynamic interaction tensor model is established, in which self-evolving spatiotemporal convolutional kernels autonomously capture nonlinear coupling relations among heterogeneous elements. The experimental results show that the method is effective in capturing the features of conditional uncertainty under crowd aggregation events to complement traffic prediction. This paradigm expands the theoretical boundaries of traffic prediction and provides novel cognitive tools for dynamic optimization of smart city systems. The code can be accessed at github.com/thenicelx/EC-DFN . • Novel Heterogeneity Disaggregation Framework. We innovatively decompose traffic uncertainty into three interactive dimensions: spatial multiscales, temporal non-stationarity, and uncertainty interference, enabling a holistic characterization of complex urban traffic dynamics. • Dynamic Interaction Tensor Model. Our approach introduces self-evolving spatiotemporal convolutional kernels that autonomously capture nonlinear couplings between heterogeneous elements (e.g., multi-scale infrastructure status, dynamic population distribution), overcoming limitations of static models in handling real-world complexity. • Geospatial Sensing Integration. The framework synergizes multi-source geospatial data to resolve the long-standing challenge of extracting and integrating heterogeneous features for capturing the evolution mechanism of traffic flow disruptions. • Expands theoretical boundaries of traffic prediction by formalizing crowd-event interactions, • Provides actionable cognitive tools for dynamic optimization in smart city systems.