Decadal Analysis of Agricultural Ammonia Distribution in China Using CAMS Data and Machine Learning Classifiers

Ammonia (NH<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf>) is a major atmospheric pollutant largely emitted from agricultural activities, contributing to air quality degradation, ecosystem imbalance, and secondary particulate matter formation. This study analyzes NH<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf> emissions across China from 2014 to 2025 using satellite data from the Copernicus Atmosphere Monitoring Service and spatial processing in QGIS. Both statistical and machine learning (ML) approaches were employed to evaluate spatiotemporal patterns and predict future emissions. Statistical analysis revealed distinct regional and seasonal trends: Beijing and Central China consistently exhibited the highest fluxes (up to ~6.04×10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−11</sup> kg m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−2</sup> s<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−1</sup>), whereas Tibet and South China recorded the lowest (~3.60×10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−13</sup> kg m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−2</sup> s<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−1</sup>), with spring and summer showing peak values. Machine learning models, including LightGBM, XGBoost, and SVM, were assessed under various feature scaling strategies. XGBoost provided the most geographically consistent predictions, capturing high emissions in eastern and northern provinces, medium in central and southern regions, and low in western areas. LightGBM achieved excellent performance under favorable conditions but exhibited variability across scenarios, while SVM offered stable but moderate accuracy. Among preprocessing methods, Z-score normalization proved optimal, and LightGBM achieved superior multi-class classification metrics, including high accuracy, F1-scores, and AUC. These results demonstrate that well-designed ML workflows can substantially enhance NH<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf> monitoring and prediction, providing a robust scientific basis for environmental management and policy planning in China.