DL-FIRE: a deep learning model for spatiotemporal prediction of the lidar-like aerosol profile over Southeast China

Vertical profiles of the aerosol extinction coefficient (AEC) offer height-resolved information crucial for quantifying aerosol-radiation interactions and aerosol-cloud processes. However, continuous observation of such profiles remains limited due to sparse ground-based lidar coverage and the discontinuous nature of satellite retrievals. To mitigate this spatiotemporal data gap, we present DL-FIRE, a novel, to our knowledge, deep learning framework for spatiotemporal prediction of lidar-like aerosol profiles. The model predicts vertical AEC profiles at hourly lead times from 0 to 48 h using lagged ERA5 meteorological features. Its architecture centers on a specialized transformer encoder that utilizes a learnable aggregation token to distill integrated profile representations and implicitly capture vertical aerosol dynamics. We further introduce a masked loss function augmented with pairwise ordering constraints, enabling effective learning from sparse and incomplete profiles. Evaluated on the full dataset, our approach achieves R=0.5308 and MSE=0.0090, reducing the prediction error by 9.1% and 11.8% compared to the multilayer perceptron (MLP) and convolutional neural network (CNN) baselines, respectively, while demonstrating enhanced stability. Importantly, the model demonstrates strong generalization capability, independently validated against ground-based lidar measurements at Guangzhou (R=0.58, MSE=0.0067) and Shouxian (R=0.53, MSE=0.0189) during 2024-2025, confirming its practical reliability. Collectively, DL-FIRE provides a reliable tool for atmospheric sensing and supports intelligent environmental decision-making, with potential applications in long-term regional aerosol prediction studies.