Deep learning sharpens GPM radar precipitation classification

A new deep-learning framework improves satellite precipitation-type classification by combining engineered radar statistics with convolutional features from dual-frequency profiles. Trained on Multi-Radar Multi-Sensor (MRMS) labels, it separates stratiform rain, snow, convective rain, and hail more accurately than the operational Global Precipitation Measurement (GPM) Dual-Frequency Precipitation Radar (DPR) Version 07 product. It promises reliable precipitation records for forecasting, climate monitoring, and water-risk assessment.

Satellite radar is essential for global precipitation monitoring, but operational precipitation-type maps still depend largely on threshold rules applied to vertical reflectivity profiles and auxiliary variables. These rules are transparent and fast, yet they can be sensitive to calibration drift, environmental variability, and mixed-phase or rapidly evolving storms. They may also miss nonlinear relationships between dual-frequency signals and precipitation structure. Such limitations matter because precipitation-type labels feed weather models, climate studies, and hydrological applications. Errors in separating convective rain, stratiform rain, snow, and hail can propagate into later products. Based on these challenges, research on machine-learning precipitation classification is needed.

Researchers from Southern Illinois University Edwardsville, the University of Oklahoma/National Severe Storms Laboratory, and Southern Illinois University Carbondale published (DOI: 10.34133/remotesensing.1078) a peer-reviewed study in the Journal of Remote Sensing on 6 October 2026. The work addresses a practical gap: Global Precipitation Measurement (GPM)’s Dual-Frequency Precipitation Radar (DPR) provides global observations, but threshold-based classification can mislabel convective cells, snow, and hail in complex or marginal cases. The team tested whether a deep neural network (DNN) could produce more consistent surface precipitation-type information from the same radar profiles.

The hybrid DNN delivered clear gains over a composite GPM/DPR Version 07 benchmark. Across stratiform, snow, convective, and hail, macro precision rose from 0.70 to 0.87 and macro recall from 0.67 to 0.88. Macro F1 improved from 0.66 to 0.87; class F1 scores reached 0.91, 0.98, 0.84, and 0.77. The model reduced the benchmark’s bias toward stratiform labels and eliminated prominent snow–convective confusion. In illustrated severe-weather cases, it placed hail cores more precisely within convective regions, while some smaller hail areas were better captured by existing algorithms. These gains support more reliable precipitation-type inputs for downstream studies.

The study used 2022 GPM overpasses intersecting the Multi-Radar Multi-Sensor (MRMS) domain. MRMS Surface Precipitation Type data, time-matched within one minute, provided labels; a footprint was retained only when all MRMS grid points inside its roughly 5.2 km diameter shared one class. The final set contained 10,000 randomly selected stratiform and snow samples plus all 5,385 convective and 1,203 hail samples, split 70/15/15. Inputs combined 13 engineered features—storm-top height, 0°C height, reflectivity maxima, and dual-frequency ratio measured (DFRm) values—with convolutional features from smoothed Ku, Ka, and DFRm slope profiles. Parallel one-dimensional filters of 5, 10, and 15 points captured multiscale vertical structure. Hyperband tuned width, dropout, batch size, and optimizer. Receiver operating characteristic (ROC) area under curve (AUC) values were 0.98 for stratiform, nearly 1.00 for snow, 0.97 for convective, and nearly 0.98 for hail; an ablation study favored the hybrid model.

The authors note that the gains matter where precipitation labels feed climate and hydrological studies. They caution that MRMS labels are an approximate reference and that tropical precipitation, transitional hail edges, and multi-year validation remain future targets. They view the hybrid design as a scalable foundation for more reliable satellite precipitation classification.

The team quality-controlled GPM Level 2 Ku- and Ka-band profiles, smoothed reflectivity over five range bins, computed DFRm, and derived slope profiles. Engineered features captured storm depth, freezing level, reflectivity maxima, and dual-frequency behavior. Convolutional branches extracted localized gradients; a dense branch processed statistical features. The fused representation passed through dense layers to a four-class softmax before classification, as described. Categorical cross-entropy trained the network, while Hyperband selected architecture and training settings. MRMS Surface Precipitation Type products supplied supervised labels.

Future work could add spatial context from neighboring radar bins, probability outputs for ambiguous cases, and multi-year cross-validation. Including tropical precipitation and improving hail-edge labeling would broaden applicability. If validated beyond the United States, such models could support global precipitation records, extreme-rain monitoring, snowfall detection, and climate-model evaluation. The approach may help agencies produce more consistent precipitation-type maps from satellite radar, though ground validation and quality-index filtering will be essential before deployment. Operational use would also require clear uncertainty communication.

###

References

DOI

10.34133/remotesensing.1078

Original Source URL

https://doi.org/10.34133/remotesensing.1078

About Journal of Remote Sensing

The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.