An XGBoost-GAN framework for mapping PM2.5 and NO2 and exploring land-use scenarios in the Sichuan Basin

Understanding the spatial distribution of air pollutants is essential for supporting air-quality management and land-use planning. This study developed an exploratory XGBoost–GAN framework to reconstruct annual PM2.5 and NO2 surfaces and to examine model-indicated pollutant-pattern differences under land-use scenarios in the Sichuan Basin, China. Annual observations from 113 monitoring stations during 2014–2024 were combined with multi-source auxiliary rasters, including aerosol optical depth, meteorological variables, land-use/land-cover data, nighttime light, elevation, and population density. Random Forest, Long Short-Term Memory, and XGBoost models were trained using 2014–2023 data and independently evaluated using 2024 station observations. XGBoost achieved the best overall performance, with R² values of 0.71 and 0.81 for PM2.5 and NO2, respectively, and was therefore used to generate 1-km baseline pollutant surfaces for 2024. Two Generative Adversarial Network (GAN) models were then trained to translate land-use images into XGBoost-estimated pollutant maps, enabling scenario experiments involving plant, built-up land, and water-body configurations. Under the condition of exploratory mode response rather than causal analysis, the scenario analysis results indicate that multiple land use layouts are related to the local decrease in pollutant concentrations estimated by the model. This study provides a data-driven approach for joint pollutant mapping and preliminary screening of land-use planning scenarios, while highlighting the need for validation with observed land-use transitions in future work.