ST‐AeroClassifier: A Lidar‐Based Aerosol Classification Framework Based on a Comprehensive Optical Property Database

Abstract

Accurate aerosol classification from spaceborne lidar is fundamental for understanding aerosol–radiation interactions and their climatic impacts. Physically consistent aerosol typing remains challenging due to overlapping optical signatures arising from particle non‐sphericity and internal inhomogeneity, as well as the limited labeled observational data. Traditional threshold‐ or clustering‐based approaches lack robustness across diverse aerosol regimes, while machine learning methods improve upon thresholding but are constrained by limited observational data or driven exclusively by synthetic data sets, reducing data utilization efficiency. These limitations hinder the development of classifiers that are physically consistent and robust. To overcome these challenges, we propose ST‐AeroClassifier, a simulation‐driven typing framework bridging comprehensive simulations and observations. Its core is a Transformer‐based kernel pre‐trained on a database generated using advanced super‐spheroidal models. This kernel learns intrinsic relationships between aerosol microphysical properties and lidar observables, enabling more efficient use of limited observational data and achieves balanced discrimination across aerosol types. The framework relies only on the particle depolarization ratio and lidar ratio, standard products of current and upcoming spaceborne lidar missions. Application to Atmospheric Lidar (ATLID) measurements aboard the Earth Cloud Aerosol and Radiation Explorer (EarthCARE) demonstrates that the framework retrieves physically consistent aerosol‐type groupings. Its extensibility is further confirmed by successful adaptation to 532 nm Advanced Carbon Dioxide Detection Lidar (ACDL) observations on Daqi‐1. Embedding aerosol simulations into the classification process, ST‐AeroClassifier provides a transferable, observation‐efficient pathway for spaceborne lidar aerosol typing, underscoring its scalability and broad applicability in future studies.