Enhanced Adaptive Learning Model for Accurate Dual-Polarization Radar Rainfall Mapping

This study investigates the generalization capability of a deep learning model for dual-polarization radar-based quantitative precipitation estimation (QPE). The model is built upon a ShuffleNet architecture and trained using an extensive dataset, including radar reflectivity, differential reflectivity, and specific differential phase, from the KMLB radar in Florida covering the years from 2017 to 2021. Rain gauge data was also incorporated to enhance accuracy in estimating surface rainfall rates. Although the model demonstrates great performance in Florida, its adaptability was assessed by testing it in two contrasting regions: Oklahoma and California. Oklahoma, with its complex weather systems influenced by topography and the nearby Gulf of Mexico and Rocky Mountains, presents unique challenges for QPE models. Similarly, California's diverse precipitation patterns, ranging from coastal rain events to mountainous snowfall, offer another rigorous testing environment. The model's performance in these regions was validated against high-precision rain gauge measurements, providing key insights into its strengths and limitations. The results indicate that while the model performs well in its training region, accuracy challenges arise when applied to areas with differing climatic conditions, highlighting the need for adaptive learning techniques or region-specific model adjustments to improve QPE performance across diverse environments.