Deep Learning-Based Hydrometeor Classification from E-Profile Ceilometers Using Cloudnet Reference Data

Ground-based lidar networks have expanded rapidly in recent years, providing continuous, high-resolution profiles of aerosols, precipitation and clouds for both operational meteorology and climate research. Among them, the EUMETNET E-Profile network now operates more than 400 single-wavelength ceilometers, enabling unprecedented spatial and temporal coverage of backscatter measurements. However, unlike synergistic radar-lidar systems such as Cloudnet, ceilometers alone do not provide operational target classification of hydrometeors or aerosol/clear-sky discrimination.

In this study, we explore the capability of artificial intelligence methods to infer Cloudnet-level target classifications directly from ceilometer backscatter profiles. The approach treats standardized 24-h time-height backscatter as image-like inputs and applies convolutional encoder-decoder architectures for semantic segmentation of atmospheric structures. Training and validation were performed using data from multiple Cloudnet reference stations at different latitudes under diverse meteorological conditions, enabling the model to learn station-agnostic spatio-temporal patterns associated with hydrometeors and aerosol layers.

Initial results demonstrate that key Cloudnet hydrometeor categories and clear-sky/aerosol regions can be recovered from ceilometer-only input, even in the absence of synergistic radar information. These findings indicate that single-wavelength backscatter can be used as input in computer-vision models, in order to extract physically meaningful patterns from the temporal evolution of the signal.

This work establishes the basis for a future near-real-time classification framework scalable to the E-Profile network. The methodology also opens new opportunities for cross-validation with spaceborne lidar and radar products, particularly from the EarthCARE mission, and for generating long-term occurrence statistics that may inform studies on cloud processes, aerosol-cloud interactions and model performance.

Acknowledgements:

This research is part of the Spanish national project PID2023-151817OA-I00, titled DeepAtmo, funded by MICIU/AEI/10.13039/501100011033 and Horizon Europe program under the Marie Sklodowska-Curie Staff Exchange Actions with the project GRASP-SYNERGY (grant agreement No. 101131631). This work is also part of the 2024 Leonardo Grant for Researchers and Cultural Creators from the BBVA Foundation. Ana del Águila is part of Juan de la Cierva programme through grant JDC2022-048231-I funded by MICIU/AEI/10.13039/501100011033 and by European Union “NextGenerationEU”/PRTR.