A Unified Deep Learning Framework for Biomass Burning Plume Detection and Domain-Adaptive PM 1 Estimation

Biomass burning is a major source of atmospheric pollution. However, rapid and quantitative assessment of particulate matter in smoke plumes remains challenging, owing to the physical uncertainties, limited coverage, and labor-intensive quality control of conventional monitoring approaches. Existing image-based deep learning methods typically address either smoke detection or air quality assessment separately. To address this gap, we develop a Unified Smoke Detection and Aerosol Estimation Framework (SDAF), a three-stage deep learning approach evaluated using a smoke-rich airborne dataset. The framework integrates smoke localization with PM 1 estimation by combining a YOLOv11-based detector with an optimized convolutional neural network. The model achieves high accuracy under in-plume conditions (R 2 of 0.985). However, its performance degrades under out-of-plume conditions due to substantial differences in visual features between the two domains. Consequently, direct across-domain transfer performs poorly, whereas region of interest (ROI)-level fine-tuning substantially improves performance for out-of-plume images (R 2 of 0.621). Despite these promising results, fundamental limitations remain. Image-based PM 1 estimation is intrinsically ill-posed due to the non-unique mapping between visual observations and particle mass. Overall, the framework enables an integrated workflow from smoke localization to quantitative PM 1 estimation using image data alone, offering a scalable solution for biomass burning monitoring and air quality assessment while highlighting the fundamentally indirect nature of image-based PM 1 inference relative to spatially resolved retrievals.