Heterogeneity-aware deep learning for spatiotemporal data: a methodological review

Spatiotemporal heterogeneity is an important property of geographic phenomena and a challenge for moving deep learning from generic data fitting toward geographic knowledge-constrained modeling. Although common implementations of convolution, graph propagation, and attention capture spatiotemporal dependencies, their parameter-sharing mechanisms and static inductive biases often assume global stationarity, hindering adaptation when variable distributions and relationships vary across space and time. Existing heterogeneity-aware designs remain fragmented across tasks, data forms, and architectures, lacking a synthesis of explicit heterogeneity modeling mechanisms. This review presents a mechanism-level framework centered on variable and relational heterogeneity. By integrating stratified and local organizational structures, the framework organizes existing methods around three stages: feature representation, model parameterization, and objective optimization. It identifies six core strategies: stratified and local feature representation, stratified and local parameterization, and feature self-supervised and parameter regularization constraints. We delineated the boundary between explicit heterogeneity-aware modeling and generic spatiotemporal context encoding, and analyzed the theoretical basis, mechanisms, applicability, and limitations of each strategy. We further discussed challenges in coupled heterogeneity modeling, spatiotemporal transferability, interpretability, computational complexity, uncertainty quantification, and spatiotemporal foundation models. This review provides a methodological reference for understanding how deep learning models represent, adapt to, and constrain spatiotemporal heterogeneity in GeoAI.