<p>Using observations from a single monitoring site in Hefei, eastern China, we developed an enhanced random-forest-gated generalized additive modeling framework (RF–GAM) for visibility nowcasting and particulate-matter (PM_2.5, PM₁₀) forecasting. The visibility and local meteorological observations covered 23 May 2024 to 9 April 2025, and the autumn visibility-prediction experiments used the subset from 1 September 2024 to 30 November 2024, whereas the PM task used a longer hourly dataset assembled from Department Of Ecology and Environment of Anhui Province and meteorological records from NOAA’s Climate Data Online (NCEI CDO). The PM modeling dataset contained 37,299 hourly samples from 2 January 2020 to 30 April 2024 after temporal alignment and quality control. In the first stage, a random forest classifier identified atmospheric regimes to mitigate class imbalance; in the second stage, regime-specific GAMs modeled conditional nonlinear responses to key meteorological predictors. For visibility, the stage-1 classifier achieved an AUC of 0.992 for fog identification, and the nowcast attained a tolerance-based accuracy of 84.3%. For PM forecasting, incorporating meteorological lag improved PM₁₀ prediction skill, increasing R² from 0.566 to 0.639. Additional ablation and error-propagation analyses showed that the RF–GAM framework outperformed ungated baselines while retaining conditional interpretability within RF-defined regimes. These results should be interpreted as autumn-season visibility performance and PM forecasting performance for a single-site case study in Hefei. External validation using multi-season and multi-site observations is still required before broader regional or operational application. </p> Graphical Abstract <p></p>