Machine Learning Interpretability Methods to Delineate the Aerosol Formation in the Arabian Sea Near Kerala Coast | ||
| Pollution | ||
| دوره 12، شماره 2، تابستان 2026، صفحه 601-613 اصل مقاله (1.15 M) | ||
| نوع مقاله: Original Research Paper | ||
| شناسه دیجیتال (DOI): 10.22059/poll.2026.409240.3256 | ||
| نویسندگان | ||
| Sherin Babu* 1؛ Marina Aloysius2؛ Sana S Navas1؛ Binu Thomas3 | ||
| 1Department of Computer Science, Assumption College Autonomous, Changanassery, Kottayam, Kerala, India | ||
| 2Department of Physics, Assumption College Autonomous, Changanassery, Kottayam, Kerala, India | ||
| 3Department of Computer Applications, Marian College, Kuttikanam, Idukki, Kerala, India | ||
| چکیده | ||
| Atmospheric aerosols have a significant function in atmospheric systems and hence play a crucial role in climatic changes. Machine learning (ML) models are highly preferred for aerosol estimation because of their exceptional predictive capability. However, it is challenging to justify and understand the predictions made by these ML models. The purpose of this research is to show how model-agnostic interpretation methods - permutation feature importance (PFI) and SHapley Additive exPlanations (SHAP) can be used to enhance and clarify machine learning model prediction of aerosols in the Arabian Sea region near the Kerala coast. Initially, the performance of 3 ML models, Polynomial regression, Bayesian ridge regression and Support Vector Regression (SVR) models are analyzed for estimating the aerosol optical depth (AOD). The study employed Pearson correlation to investigate the relationships between AOD and the various input features and to find the best features for building the ML models. Mean Squared Error (MSE) and Coefficient of Determination (R2) are the performance metrics used to assess these models' performance. Results indicated that SVR model (with R2 = 0.7933 and MSE = 0.0063) provided better predictive performance. Then the predictions of the most accurate model are explained by PFI and SHAP. The ML interpretability analysis showed that the main factors strongly associated with aerosol formation are aerosol radiative forcing at the top of the atmosphere (ARF_TOA), radiative forcing at the surface of the atmosphere (ARF_SURF), sea salt and temperature profile at 250hPa. | ||
| کلیدواژهها | ||
| AOD؛ Interpretable ML؛ PFI؛ SHAP؛ SVR | ||
| مراجع | ||
|
| ||
|
آمار تعداد مشاهده مقاله: 207 تعداد دریافت فایل اصل مقاله: 180 |
||