Spatially Gated Mixture of Experts for Missing Data Imputation in Pavement Management Systems

Accurate imputation of missing pavement-condition data is critical for proactive infrastructure management, yet it is complicated by spatial non-stationarity—deterioration patterns and data quality vary markedly across regions. This study proposes a Spatially Gated Mixture-of-Experts (SG-MoE) imputation model that explicitly encodes spatial heterogeneity by (i) clustering road segments using geographic coordinates and (ii) supervising a gating network to route each sample to region-specialized expert regressors. Using a large-scale national pavement management database, we benchmark SG-MoE against a strong baseline under controlled missingness mechanisms (MCAR: missing completely at random; MAR: missing at random; MNAR: missing not at random) and missing rates (10–50%). Across scenarios, SG-MoE consistently matches or improves upon the baseline; the largest gains occur under MCAR and the challenging MNAR setting, where spatial specialization reduces systematic underestimation of high crack-rate sections. The results provide practical guidance on when spatially aware ensembling is most beneficial for infrastructure imputation at scale. We additionally report comparative results under three missingness mechanisms. Across five random seeds, SG-MoE is comparable to the single LightGBM baseline under MCAR/MAR and achieves its largest gains under MNAR (e.g., sMAPE improves by 0.82 points at 10% MNAR missingness).