| 21 Sep 2026
Machine Learning-Enhanced Aerosol Type Classification: Global Multi-Parameter Aerosol Modeling and an Automated Classifier Driven by Active–Passive Remote Sensing Parameters
Abstract. Accurate classification of aerosol types is crucial for regional aerosol pollution source tracking and optical remote sensing studies. However, conventional classification methods are mostly relying on the single-device observation that provides limited optical parameters, and the use of fixed empirical thresholds that limits the classification accuracy. This study develops a machine learning-enhanced aerosol types classification method that integrates active and passive optical remote sensing data. First, the global multi-parameter aerosol model is built from high-quality aerosol product datasets of 19 representative AERONET sites worldwide, 7 distinct aerosol type classes were established using the K-means clustering algorithm. Then,
based on the established aerosol database with multiple characteristic parameters, we propose a random forest-based aerosol types classifier, which uses only active-passive remote sensing directly retrieved parameters as inputs. This classifier is evaluated to achieve an overall accuracy of approximately 80% and successfully classifies the dominant aerosol types at 27 typical AERONET sites. This study offers a practical and reliable method for combining active and passive remote sensing data to automated aerosol types classification.
The authors describe a method for classifying different aerosol types based on AERONET inversion products using k-means clustering. In terms of scope and results, this work is strikingly similar to Shin, S.-K., Tesche, M., Noh, Y., and Müller, D.: Aerosol-type classification based on AERONET version 3 inversion products, Atmos. Meas. Tech., 12, 3789–3803, https://doi.org/10.5194/amt-12-3789-2019, 2019.
In this well-received paper (cited 74 times since 2019), we use lidar parameters as provided in the AERONET version 3 dataset for aerosol classification. Based on selected sites that are expected to represent different aerosol types in their pure form, we define 7 aerosol types that can be identified based on threshold values of the particle linear depolarization ratio and the single-scattering albedo, both at 1020 nm, from AERONET inversion products. Our aerosol types follow a physically meaningful definition based on shape and absorptivity, and include mineral dust, dust-pollution mixtures that are either dominated by dust or not, and non-dust aerosols with different levels of absorption.
Having said this, I am surprised that our earlier work is not considered here, particularly as other papers based on our classification are. I recommend that the authors put their findings in relation to those of Shin et al. (2019). Both studies end up with 7 aerosol types. As stated above, physical reasoning went into our choice and the classification can be done analytically based on two parameters. The machine-learning retrieval on the other hand includes a multitude of parameters and bases its aerosol-type definition on statistically robust differences in the clusters. From looking at the results, it seems to me that there is quite some consistency between the aerosol types in the different clusters. Trying to reconcile the two approaches could thus strengthen the interpretation of the results obtained with the machine-learning method. In that context, the authors should also explicitly state how their work (a back box that uses 13 input parameters) is an advancement beyond our earlier method (threshold-based analytical solution with two input parameters). If it is primarily related to the application of machine learning methods, I recommend moving this work to AMT as that would be outside the scope of ACP.