CAT-GAN-UKF: uncertainty-aware spatiotemporal reconstruction for resilient urban air-quality monitoring under partial observability

Urban air-quality monitoring networks are inherently incomplete: measurements disappear intermittently, sustained outages occur, and the reliability of the remaining evidence changes with pollutant, location, temporal regime, and network state. We present CAT-GAN-UKF, an uncertainty-aware framework for multivariate reconstruction under partial observability that combines nonlinear spatiotemporal state propagation, adversarial learning of context-dependent uncertainty, and masked recursive Bayesian updating. Pollutant-specific conditional Wasserstein generative adversarial networks with gradient penalty (WGAN-GP) are trained on signed empirical process- and observation-residual distributions derived exclusively from the chronological training period. Monte Carlo variance of generated residuals defines time-varying diagonal \(Q_t\) and \(R_t\) , which are injected into a nonlinear network UKF. The primary model uses causal temporal, lagged, rolling, missingness, and spatial context; temperature and humidity are evaluated only in a sensitivity analysis because they did not improve reconstruction in this benchmark. The evaluation uses 153,260 hourly station-time records from an OpenAQ-derived 2024 New York City metropolitan-area dataset. After predefined quality control, 27 of 33 monitoring locations were retained. All learned parameters are fitted on training data only; a separate validation interval is used for pollutant-specific multiplicative interval calibration, and the held-out test period is never used for fitting or calibration. Robustness is assessed under four random-masking levels, four contiguous outage lengths, and an empirical-gap replay protocol whose outage-length distribution is learned from naturally occurring training-period gaps. Each stress test is repeated over 30 masking seeds. CAT-GAN-UKF is compared with matched linear KF, fixed-covariance UKF and heuristic adaptive UKF baselines, as well as Extra Trees, random forest, histogram gradient boosting, multilayer perceptron, ridge, lasso, persistence, and temporal means. CAT-GAN-UKF significantly improves RMSE over adaptive UKF in seven of nine scenarios, linear KF in all nine, Extra Trees in all nine, and persistence in eight of nine after Holm correction. Relative to a strong fixed-covariance UKF, point-estimation accuracy is generally comparable rather than uniformly superior; however, CAT-GAN-UKF reduces average 95% interval width by approximately 38% across the nine scenarios while retaining mean empirical coverage near 95%. Spatial coupling is the dominant structural contribution: removing it increases RMSE by approximately 5.9% under 10% random masking and by 42.7% under a 24-hour outage. Conversely, adding meteorological covariates worsens RMSE throughout the sensitivity analysis. These findings establish an uncertainty-efficient research framework for spatial reconstruction under degraded sensing; real-time latency, memory use, and resource-constrained deployment feasibility were not evaluated.