Feature-Based Machine Learning Framework for Multi-Source NH3 Dataset Analysis

Abstract

Accurate monitoring of atmospheric ammonia (NH3) is important for air-quality assessment and sustainable agriculture, but available datasets differ in spatial resolution, temporal coverage, units, and physical meaning. This study presents a feature-based machine-learning framework with train–test leakage-controlled preprocessing to evaluate relative NH3 classification consistency across three datasets over China: CAMS GEI, CAMS EAC4, and MEIC. Statistical descriptors were extracted for four spatial–temporal cases: Province–Year, Zone–Year, Province–Season–Year, and Zone–Season–Year. Six classifiers were evaluated using chronological testing and cross-validation. CAMS GEI provided the broadest spatial–temporal coverage, while MEIC showed comparatively stable classifier behavior. The best-performing ensemble models commonly achieved accuracies between 0.97 and 0.99 in the main province-level cases. Mean Absolute Value (MAV) was the leading feature in several province-level analyses, with a maximum reported contribution of 64.26%. These scores describe the separability of threshold-derived reference classes and should not be interpreted as an independent physical prediction of NH3 from external atmospheric drivers.