Accurately characterizing the spatio-temporal evolution of urban air pollution and its meteorological response remains challenging because public multi-source observations are often asynchronous, incomplete, and heterogeneous. To address this issue, this study proposes UWMR-Net, an uncertainty-aware and wind-guided spatio-temporal framework that integrates public remote-sensing products, meteorological reanalysis, static land-surface attributes, and activity-related proxies for joint pollutant forecasting and meteorological-response diagnosis. The model combines an uncertainty-aware fusion module for quality-, lag-, and missingness-sensitive multi-modal integration, a wind-guided dynamic propagation backbone for directional transport modeling, and a counterfactual response branch for separating atmospheric response from background emission-related signals. Experiments on two public study domains, Los Angeles and the Po Valley, show that UWMR-Net achieves the best performance with five-seed mean ± SD reporting across PM2.5 and O3 1-day-ahead and 7-day-ahead forecasting tasks. The UWMR-Net entries are calculated from cleaned row-level test predictions over five independent random seeds. Rolling-origin backtests, macro-area diagnostics, extreme-event analyses, and split-conformal prediction intervals further characterize temporal stability, spatial heterogeneity, and calibrated forecast uncertainty. Counterfactual case studies additionally reveal spatially coherent meteorological enhancement and mitigation patterns during representative particulate and ozone episodes. These results demonstrate that uncertainty-aware fusion and wind-guided response modeling provide an effective and interpretable framework for urban air-pollution forecasting and meteorological-response analysis from public multi-source data.