Key points are not available for this paper at this time.
Introduction Aerosol behavior plays a crucial role in atmospheric research, environmental monitoring, and climate modeling. Understanding aerosol composition, sources, and temporal dynamics is essential for improving air quality assessment and climate predictions. Methods This study proposes a comprehensive machine learning-based framework for aerosol classification and long-term forecasting in the Dibrugarh region of India. A multiannual dataset comprising key aerosol parameters—Aerosol Optical Depth (AOD), Angstrom Exponent (AE), Fine Mode Fraction (FMF), and Single Scattering Albedo(SSA)—was utilized. Various machine learning and deep learning models, including Random Forest, XGBoost, CatBoost, LSTM, and Transformer architectures, were applied. The framework integrates composition-based, source-based, and seasonal classification, along with correlation analysis using PCA, t-SNE, and Chi-square tests. Data imbalance was addressed using SMOTE, ADASYN, and sampling techniques, and long-term forecasting was performed using a BiLSTM-Attention model. Results The proposed models achieved high performance across multiple tasks. The BiLSTM-Attention model demonstrated approximately 90% accuracy in aerosol type forecasting over a 10-year horizon. The results reveal a strong dependency between aerosol composition and source types, along with significant seasonal variations in aerosol behavior. Discussion The findings highlight the effectiveness of the proposed hybrid framework in capturing complex aerosol patterns and temporal dynamics. The study provides a robust and interpretable approach for aerosol monitoring, with potential applications in environmental policy planning and climate research.
Kukkar et al. (Wed,) studied this question.