Neural Hydrometeor Phase Detection: Deep Learning-Based Prediction of Liquid and Ice Water Content from NEXRAD Radar Observations

To enhance our understanding of hydrometeor phase and cloud microphysical dynamics, a neural network model capable of predicting liquid water content (LWC) and ice water content (IWC) using data from dual-polarization operational NEXRAD S-band radar observations was developed. Accurate quantification of these properties is crucial for differentiating hazardous from non-hazardous aviation flight domains, weather prediction, and climate modeling.

Hydrometeor Classification Algorithms (HCAs) have employed various techniques to interpret radar data. Fuzzy logic classifiers (Serke et al. 2022; Ryzhkov et al. 2005) have been core to many approaches, some have even been supplemented with neural networks to create neuro-fuzzy classifiers (Liu and Chandrasekar 2000). However, these approaches often struggled or did not attempt to distinguish homogeneous supercooled liquid water from ice particle populations in winter storms (Vivekanandan et al. 1999; Elmore 2011). Recent efforts have explored new directions to address these challenges. Schuur et al. (2012) incorporated numerical weather prediction model data to improve classification in winter scenarios. Grazioli et al. (2015) proposed an unsupervised clustering technique, comparing results to video disdrometer imagery. The Radar Icing Algorithm (RadIA) by Serke et al. (2022) specifically targeted in-flight icing detection using dual-polarization radar fields and model temperature profiles. This work builds upon these approaches, introducing a supervised learning method enabled by a novel spatiotemporal weather data fusion workflow. By leveraging in-situ measurements from a recent research field campaign to create a robust training dataset, we aim to address the persistent challenges in winter precipitation classification and to improve upon existing methods.

This study leverages data from the Seeded and Natural Orographic Wintertime Clouds: The Idaho Experiment (SNOWIE) campaign, combining operational S-band radar measurements with high-resolution in situ data collected from a Nevzorov probe mounted on a research aircraft. A Graphical Processing Unit-accelerated spatiotemporal fusion workflow was employed to interpolate between the sparse, volumetric radar data and the dense, trajectory-based aircraft measurements.

The data fusion process involved several key steps to align radar data with aircraft measurements. First, NEXRAD radar volume data were matched to flight path timestamps, considering a plus or minus thirty minute window around each flight point. For each radar file, additional fields were calculated such as specific differential phase and spatial statistics. These spatial statistics included standard deviations and texture fields, computed using a 3x3 window along each radar tilt angle, providing information about local variability in radar measurements. Radar and flight path coordinates were converted to a common Earth-centered reference system to enable accurate spatial interpolation. We then used inverse distance weighted interpolation to estimate radar variables at each flight path point. This process was repeated for all relevant radar files, resulting in multiple sets of radar values for each flight point at different times. Finally, we used temporal interpolation to align these radar estimates with the exact flight timestamps. This two-step approach effectively combined the broad spatial coverage of radar data with the precise, in-situ measurements from the aircraft, creating a comprehensive dataset for analysis.

The resulting fused dataset comprises 18 radar-derived variables and corresponding in situ liquid water content (LWC) and ice water content (IWC) measurements from approximately 50,000 flight data points covering a 3 month period during the 2017 flight campaign. To ensure representative sampling, we implemented a 50%/50% train-test split stratified by flight day, obtaining equal proportions of data from each day for both training and evaluation sets.

Our prediction model was based on a feed-forward neural network consisting of 24 hidden layers, each with 1024 neurons. Rectified Linear Unit (ReLU) activation functions were utilized to introduce non-linearity and help prevent the vanishing gradient problem, alongside batch normalization after each hidden layer to improve training stability. To address the challenge of predicting two different outputs (LWC and IWC) with potentially different scales, the authors employed a normalized mean squared error (MSE) loss function. This approach ensured that the model attributed equal importance to accurately predicting both LWC and IWC, regardless of their relative magnitudes The model used the 18 radar-derived variables as input and simultaneously predicted LWC and IWC. This dual-output design allowed the model to learn shared representations that benefit both predictions, potentially capturing complex relationships between liquid and ice water content in mixed-phase conditions.

The neural network model achieved a mean squared error (MSE) of 0.003, averaged across liquid water content (LWC) and ice water content (IWC) predictions (units of g/m³). For context, the typical range of LWC and IWC values in our dataset is 0.0-0.3 g/m³. To benchmark our approach, we compared output to XGBoost regression, a powerful ensemble learning technique based on decision trees. The XGBoost model achieved an MSE of 0.007 g/m³, indicating that the neural network outperformed this traditional machine learning approach on this dataset. The key insight from this study was the potential for further improvement through increased model complexity and utilization of larger datasets of similar quality. Drawing parallels from advancements in other domains such as natural language processing, the authors anticipate that scaling up both the model architecture and the volume of training data could lead to significant performance gains. This scalability is a particular strength of neural network approaches.

Future work will focus on expanding the training dataset to include a wider range of geographic locations and atmospheric conditions, potentially incorporating data from multiple field campaigns. Additionally, the authors plan to explore the integration of other sensing platforms to further enhance the model's predictive capabilities. Overall, this research highlights the potential for machine learning to enhance radar-based cloud microphysics measurements. The ability to derive high-resolution LWC and IWC estimates from radar data could lead to more accurate weather predictions, improved cloud seeding operations, enhanced aviation safety through better icing forecasts, and deeper insights into cloud-climate interactions.