Classifying thermodynamic cloud phase using machine learning models

Abstract. Vertically resolved thermodynamic cloud-phase classifications are essential for studies of atmospheric cloud and precipitation processes. The Department of Energy (DOE) Atmospheric Radiation Measurement (ARM) Thermodynamic Cloud Phase (THERMOCLDPHASE) value-added product (VAP) uses a multi-sensor approach to classify the thermodynamic cloud phase by combining lidar backscatter and depolarization, radar reflectivity, Doppler velocity, spectral width, microwave-radiometer-derived liquid water path, and radiosonde temperature measurements. The measured pixels are classified as ice, snow, mixed phase, liquid (cloud water), drizzle, rain, and liq_driz (liquid+drizzle). We use this product as the ground truth to train three machine learning (ML) models to predict the thermodynamic cloud phase from multi-sensor remote sensing measurements taken at the ARM North Slope of Alaska (NSA) observatory: a random forest (RF), a multi-layer perceptron (MLP), and a convolutional neural network (CNN) with a U-Net architecture. Evaluations against the outputs of the THERMOCLDPHASE VAP with 1 year of data show that the CNN outperforms the other two models, achieving the highest test accuracy, F1 score, and mean intersection over union (IOU). Analysis of ML confidence scores shows that ice, rain, and snow have higher confidence scores, followed by liquid, while mixed, drizzle, and liq_driz have lower scores. Feature importance analysis reveals that the mean Doppler velocity and vertically resolved temperature are the most influential data streams for ML thermodynamic cloud-phase predictions. Lidar measurements exhibit lower feature importance due to rapid signal attenuation caused by the frequent presence of persistent low-level clouds at the NSA site. The ML models' generalization capacity is further evaluated by applying them at another Arctic ARM site in Norway using data taken during the ARM Cold-Air Outbreaks in the Marine Boundary Layer Experiment (COMBLE) field campaign. The models demonstrated similar performance to that observed at the NSA site. Finally, we evaluate the ML models' response to simulated instrument outages and signal degradation and show that a CNN U-Net model trained with input channel dropouts performs better when input fields are missing.