AI-Driven Spatiotemporal Data Mining: Novel Architectures for Imputation and Forecasting

Accurate spatiotemporal prediction is fundamentally essential for anticipating and managing the dynamic evolutions within global physical, environmental, and social systems. While existing architectures effectively model complex spatiotemporal relationships, their predictive capabilities frequently fail under conditions of severe data sparsity and out-of-distribution (OOD) shifts. The central scientific challenge in this domain is the pervasive unreliability of observational networks during atypical scenarios, such as unforeseen transit disruptions, extreme meteorological anomalies, or massive sensor failures, which create extreme data scarcity, OOD shifts, and highly non-stationary temporal dynamics. Developing advanced, adaptable AI frameworks that can robustly predict spatiotemporal phenomena specifically under these volatile, data-sparse conditions is an urgent scientific necessity.

This Research Topic aims to aggregate cutting-edge AI methodologies that specifically address the systemic challenges of forecasting under extreme data sparsity and volatile environmental shifts. This collection seeks solutions to three fundamental scientific problems. First, we focus on forecasting under extreme data scarcity, exploring novel architectures like physics-informed neural networks that can generate reliable predictions with minimal historical observations. Second, we aim to overcome out-of-distribution (OOD) shifts, seeking generalized models capable of maintaining invariant spatiotemporal representations even when tested on unseen or anomalous data distributions. Third, we emphasize tackling highly non-stationary temporal dynamics, encouraging the development of dynamic architectures and causal inference frameworks that can rapidly adapt to sudden, unforeseen environmental disruptions without relying on fixed historical periodicity. Ultimately, our goal is to advance spatiotemporal knowledge discovery to support resilient civil infrastructure and environmental sustainability capable of withstanding unexpected crises.

We welcome Original Research, Systematic Reviews, and Perspective articles within the “Machine Learning and Artificial Intelligence” section. To ensure a well-defined identity, this collection emphasizes submissions that directly tackle key research questions surrounding spatiotemporal forecasting under extreme data sparsity and OOD shifts. Scientific themes of interest include: novel AI architectures specifically designed for robust forecasting with highly incomplete data; predictive models that dynamically adapt to non-stationary shifts; and physics-informed or causality-aware neural networks for data-scarce environments. We strongly encourage empirical studies applying these solutions to diverse atypical scenarios, including emergency urban computing (e.g., anomalous subway passenger flows), extreme climate event forecasting, and disaster response logistics. Submissions should clearly articulate how their methodological innovations overcome the specific scientific constraints of highly volatile, anomaly-driven spatiotemporal datasets.