Seminar Speaker: CliMAS Graduate Student, Alfonso Ladino-Rincon
- Event Type
- Seminar/Symposium
- Sponsor
- Professor Steve Nesbitt
- Location
- 2079 NHB
- Date
- Dec 2, 2025 3:30 pm
- Views
- 23
Machine Learning for Precipitation Retrieval and Radar Big Data: From Microphysics to Scalable Workflows
This work bridges radar-based retrieval of precipitation microphysics with scalable data infrastructure to support next-generation atmospheric science and machine learning applications. The first component investigates the underconstrained nature of dual-frequency precipitation radar (DPR) retrievals, where only two observables, reflectivity at Ku and Ka bands, are available to infer three-parameter gamma drop size distributions (PSDs). Using in situ PSD measurements from NASA’s CAMP2Ex field campaign, we performed synthetic retrieval experiments to quantify the limitations of operational and analytical DPR retrievals. A deep neural network (DNN) trained on matched reflectivity and PSD samples achieved significantly improved performance.
In parallel, a novel data infrastructure called Radar DataTree was developed to address the challenges of working with large-scale, fragmented radar archives. Radar DataTree extends the WMO FM-301 standard from file-level organization to time-resolved, dataset-level structures. Built using open-source tools such as xarray.DataTree, Zarr, and Icechunk, the system transforms radar volumes into FAIR-compliant, cloud-native datasets suitable for scalable and reproducible workflows. This architecture enables efficient parallel processing of large radar collections, with demonstrated gains of 100× in tasks such as Quasi-Vertical Profile (QVP) generation and quantitative precipitation estimation (QPE).
Ongoing research focuses on integrating disdrometer-informed PSD data with machine learning models to retrieve full particle size distributions directly from radar observables. Two strategies are being explored: (1) Conditional Generative Adversarial Networks (CGANs) to generate realistic PSDs conditioned on radar inputs, and (2) self-supervised learning approaches that leverage the structure and metadata capabilities of Radar DataTree to train retrieval models at scale without requiring labeled PSD parameters. Together, these developments offer a pathway toward end-to-end machine learning systems that combine advanced retrieval techniques with robust radar data infrastructure, enabling more accurate, scalable, and AI-ready tools for precipitation science.
ZOOM:
https://illinois.zoom.us/j/82767407475?pwd=7ca6OfzVOPneyOwAULFkntXaYWaBvo.1
Meeting ID: 827 6740 7475
Pass Code: 809774