Precipitation Modeling | Machine Learning & Atmospheric Processes Group

Understanding global precipitation remains one of the most important and most difficult problems in Earth system science. Rain and snow drive the water cycle and influence nearly every aspect of life from agriculture to energy to natural hazards. But the processes that govern them, convection, particle growth, and melting, occur at scales that are incredibly small compared to the resolution of satellite instruments and climate models.

We are developing machine learning methods that bridge this gap. Using both supervised and unsupervised learning we extract patterns from radar and radiometer data to improve how we retrieve precipitation and represent it in models. From snowflake particle size distributions to global precipitation maps, we are working to reveal what the data are already telling us but has remained hidden in the noise.

Convective storms and complex cloud microphysics make global precipitation difficult to measure and simulate accurately

Some of our recent projects include: developing deep learning models to convert between microwave emissivity channels for better radiative transfer modeling of global precipitation, building physically interpretable networks that classify precipitation phase from satellite observations, and clustering precipitation profiles using nonlinear dimensionality reduction to identify evolutionary pathways in snow and rain formation.

We use data from satellite missions like GPM and products like IMERG to train and validate our models on global scales

We are continually looking to examine new and exciting precipitation products (e.g., satellite-based, reanalysis, model simulations), so please reach out if you have data you’d like to share!

Related Publications

2025

Sci. Adv.

Decoding global precipitation processes and particle evolution using unsupervised learning

Fraser King, Claire Pettersen, Brenda Dolan, and 2 more authors

Science Advances, 2025

Abs DOI

High-quality hydrometeor microphysical observations are essential for accurate precipitation estimates and for evaluating weather and climate models. However, analyzing these properties is challenging due to their high variability, complex interactions, and large data volumes. In this study, we examine more than 1.5 million minute-scale rain and snow particle attributes and collocated meteorological variables from seven global measurement sites over 9 years. Applying Uniform Manifold Approximation and Projection (UMAP) for nonlinear dimensionality reduction, we reduce the dataset’s dimensionality by 75%, identifying nine distinct precipitation groups and associated particle evolution pathways. UMAP effectively captures the global structure of precipitation phases such as rain, snow, and mixed-phase types, revealing clear patterns that linear methods struggle to resolve. The resulting UMAP manifold offers a unique perspective on precipitation phase and intensity, advancing our understanding of particle evolutionary processes and offering valuable insights for improving weather and climate models and remote sensing precipitation estimates. Analyzing rain and snow particles reveals hidden patterns, improving precipitation classification for better model predictions.

2024

JAS

Primary Modes of Northern Hemisphere Snowfall Particle Size Distributions

Fraser King, Claire Pettersen, Brenda Dolan, and 2 more authors

Journal of the Atmospheric Sciences, Dec 2024

Publisher: American Meteorological Society Section: Journal of the Atmospheric Sciences

Abs DOI

A comprehensive understanding of snowfall microphysics is crucial for enhancing the accuracy of remote sensing snowfall retrievals. However, variations in regional and seasonal snow particle size distributions (PSDs) contribute substantial uncertainty. Here, we examine snowfall PSDs from across the Northern Hemisphere, applying principal component analysis (PCA) to disdrometer observations with the aim of identifying dominant modes of variability across varying regional climates. The PCA revealed three empirical orthogonal functions (EOFs) that account for a combined 95% of the variability across the dataset, which are attributed to latent linear embeddings of snowfall intensity (EOF1), snowfall character (EOF2), and snowfall regime (EOF3). Examining point clusters with the highest combined EOF values reveals six distinct modes of variability [i.e., principal component (PC) groups] with unique PSD traits. These groups are then correlated with environmental factors using data from collocated vertically pointing radar, surface meteorology, and reanalysis to assist in assigning physical attributes. The first and second PC groups, linked to EOF1’s intensity embedding, are described by their PSD intercepts, snowfall rates, and reflectivity and Doppler velocity values, representing low- and high-intensity snowfall modes, respectively. The third and fourth PC groups, associated with EOF2’s character embedding, are defined by temperature, fall speed, and density, indicative of cold, fluffy snowfall and warm, dense snowfall, respectively. The fifth and sixth PC groups, related to EOF3’s regime embedding, are distinguished by their PSD slope, snowfall rate, and reflectivity profiles, signifying shallow, weak convective systems with small particles and deep, stratiform snowfall events with large aggregates, respectively. Significance Statement This research enhances our understanding of varying snow particle size distributions in the Northern Hemisphere, offering valuable new insights for improving future remote sensing–based snowfall retrieval algorithms. Using a statistical technique called principal component analysis, we found that 95% of the variability in observed snowfall could be explained by three primary features: how intense the snowfall is, how dense the particles are, and the depth of the storm. We identified six unique snowfall groups, each with its own set of traits, such as the snow being light and fluffy, or heavy and packed. By linking these traits to external environmental observations, we can better understand the driving physical mechanisms within each group.

2022

AMT

DeepPrecip: a deep neural network for precipitation retrievals

Fraser King, George Duffy, Lisa Milani, and 3 more authors

Atmospheric Measurement Techniques, Oct 2022

Publisher: Copernicus GmbH

Abs DOI

Remotely-sensed precipitation retrievals are critical for advancing our understanding of global energy and hydrologic cycles in remote regions. Radar reflectivity profiles of the lower atmosphere are commonly linked to precipitation through empirical power laws, but these relationships are tightly coupled to particle microphysical assumptions that do not generalize well to different regional climates. Here, we develop a robust, highly generalized precipitation retrieval algorithm from a deep convolutional neural network (DeepPrecip) to estimate 20 min average surface precipitation accumulation using near-surface radar data inputs. DeepPrecip displays a high retrieval skill and can accurately model total precipitation accumulation, with a mean square error (MSE) 160 % lower, on average, than current methods. DeepPrecip also outperforms a less complex machine learning retrieval algorithm, demonstrating the value of deep learning when applied to precipitation retrievals. Predictor importance analyses suggest that a combination of both near-surface (below 1 km) and higher-altitude (1.5–2 km) radar measurements are the primary features contributing to retrieval accuracy. Further, DeepPrecip closely captures total precipitation accumulation magnitudes and variability across nine distinct locations without requiring any explicit descriptions of particle microphysics or geospatial covariates. This research reveals the important role for deep learning in extracting relevant information about precipitation from atmospheric radar retrievals.

2022

JAMC

A Centimeter-Wavelength Snowfall Retrieval Algorithm Using Machine Learning

Fraser King, George Duffy, and Christopher G. Fletcher

Journal of Applied Meteorology and Climatology, Aug 2022

Publisher: American Meteorological Society Section: Journal of Applied Meteorology and Climatology

Abs DOI

Remote sensing snowfall retrievals are powerful tools for advancing our understanding of global snow accumulation patterns. However, current satellite-based snowfall retrievals rely on assumptions about snowfall particle shape, size, and distribution that contribute to uncertainty and biases in their estimates. Vertical radar reflectivity profiles provided by the vertically pointing X-band radar (VertiX) instrument in Egbert, Ontario, Canada, are compared with in situ surface snow accumulation measurements from January to March 2012 as a part of the Global Precipitation Measurement (GPM) Cold Season Precipitation Experiment (GCPEx). In this work, we train a random forest (RF) machine learning model on VertiX radar profiles and ERA5 atmospheric temperature estimates to derive a surface snow accumulation regression model. Using event-based training–testing sets, the RF model demonstrates high predictive skill in estimating surface snow accumulation at 5-min intervals with a low mean-square error of approximately 1.8 × 10−3 mm2 when compared with collocated in situ measurements. The machine learning model outperformed other common radar-based snowfall retrievals (Ze–S relationships) that were unable to accurately capture the magnitudes of peaks and troughs in observed snow accumulation. The RF model also displayed consistent skill when applied to unseen data at a separate experimental site in South Korea. An estimate of predictor importance from the RF model reveals that combinations of multiple reflectivity measurement bins in the boundary layer below 2 km were the most significant features in predicting snow accumulation. Nonlinear machine learning–based retrievals like those explored in this work can offer new, important insights into global snow accumulation patterns and overcome traditional challenges resulting from sparse in situ observational networks.

2021

Atmosphere

Seasonal Estimates and Uncertainties of Snow Accumulation from CloudSat Precipitation Retrievals

George Duffy, Fraser King, Ralf Bennartz, and 1 more author

Atmosphere, Mar 2021

Number: 3 Publisher: Multidisciplinary Digital Publishing Institute

Abs DOI

CloudSat is often the only measurement of snowfall rate available at high latitudes, making it a valuable tool for understanding snow climatology. The capability of CloudSat to provide information on seasonal and subseasonal time scales, however, has yet to be explored. In this study, we use subsampled reanalysis estimates to predict the uncertainties of CloudSat snow water equivalent (SWE) accumulation measurements at various space and time resolutions. An idealized/simulated subsampling model predicts that CloudSat may provide seasonal SWE estimates with median percent errors below 50% at spatial scales as small as 2° × 2°. By converting these predictions to percent differences, we can evaluate CloudSat snowfall accumulations against a blend of gridded SWE measurements during frozen time periods. Our predictions are in good agreement with results. The 25th, 50th, and 75th percentiles of the percent differences between the two measurements all match predicted values within eight percentage points. We interpret these results to suggest that CloudSat snowfall estimates are in sufficient agreement with other, thoroughly vetted, gridded SWE products. This implies that CloudSat may provide useful estimates of snow accumulation over remote regions within seasonal time scales.

2020

ESS

Using CloudSat-CPR Retrievals to Estimate Snow Accumulation in the Canadian Arctic

Fraser King and Christopher G. Fletcher

Earth and Space Science, 2020

Abs DOI

Snow is a critical contributor to our global water and energy budget, with profound impacts for water resource availability and flooding in cold regions. The vast size and remote nature of the Arctic present serious logistical and financial challenges to measuring snow over extended time periods. Satellite observations provided by the Cloud Profiling Radar instrument—installed on the National Aeronautics and Space Administration satellite CloudSat—allow the retrieval of snowfall rates in high-latitude regions, which have been used to estimate surface snow accumulation. In this study, a validation of CloudSat-derived terrestrial snow estimates is presented at four Environment and Climate Change Canada weather stations situated in the Arctic for the common period 2007–2015. Comparisons of monthly climatological snow accumulation show mean biases of less than 1.5-mm snow water equivalent annually. Monthly time series exhibit correlations above 0.5 and root-mean-square error below 10-mm snow water equivalent at the two highest-latitude stations (Eureka and Resolute Bay) with correlations falling below 0.5 south of 70°N. CloudSat was also found to underestimate annual mean snow accumulation at the majority of sites, suggesting a potential negative bias in CloudSat’s accumulation estimates, or underestimation related to sampling. These results imply that CloudSat can provide reliable estimates of snow accumulation across similar high-latitude regions above 70°N.