Regime-Dependent Transferability of Hourly NO₂ Forecasts: A Multi-Source Machine-Learning Study of Two Climatically Contrasting Cities in Kazakhstan

This repository contains the processed dataset, the trained LightGBM model, SHAP analysis outputs, evaluation results, and Python code used in the study "Regime-Dependent Transferability of Hourly NO₂ Forecasts: A Multi-Source Machine-Learning Study of Two Climatically Contrasting Cities in Kazakhstan". Study region: Almaty (43.24° N, 76.92° E) and Astana (51.17° N, 71.45° E), Kazakhstan. Study period: 2021-01-01 to 2025-12-31 (hourly resolution). Data sources integrated in this study: Ground-station NO₂ observations from 11 regulatory stations of RSE Kazhydromet (7 stations in Almaty, 4 in Astana). Raw Kazhydromet data are subject to the provider's access terms and are not redistributed in this repository. ERA5 meteorological reanalysis (Copernicus Climate Data Store). CAMS EAC4 atmospheric composition reanalysis (Copernicus Atmosphere Data Store). Sentinel-5P TROPOMI tropospheric NO₂ column, retrieved via Google Earth Engine. Contents of this repository: Processed feature tables (station-hour records with meteorological, chemical, satellite, and derived features). SHAP values, feature matrices, and metadata used for interpretability analysis. Model evaluation results (metrics by horizon and city, test-set predictions). Station catalog with coordinates and metadata. Python scripts for feature engineering, model training, SHAP analysis, and the decision-support prototype. Model performance summary: R² = 0.70 and +25.6% relative improvement over the persistence baseline at the +6 h horizon on the 2025 test set (109,082 hourly records). Intended use: reproducibility of the results reported in the associated publication; benchmarking of air-quality forecasting models for Central Asia; educational and research use in atmospheric science and applied machine learning. How to cite: Please cite both the associated publication and this Zenodo record. License: Creative Commons Attribution 4.0 International (CC BY 4.0) for the derived datasets and code produced by the authors. Third-party data (ERA5, CAMS, Sentinel-5P) remain subject to their original licenses and terms of use.