Authors: Marshall Rosenhoover – University of Alabama-Huntsville
Title: Fusing Radar and Personal Weather Station Data for Spatial Rain Rate Field Generation Using Machine Learning
Presentation Type: Poster
Abstract: This study introduces a machine learning framework for generating spatially continuous rain rate fields, which represent rainfall rates across a given area. Generating these fields requires significant amounts of ground truth data, which must be densely populated. To address this challenge, we propose a method that combines radar precipitation rate estimates with Personal Weather Station (PWS) rain gauge measurements to derive rainfall accumulation functions during the active period of the PWS. While PWS rain gauges are known to contain errors and biases, they are difficult to quantify in practice. Thus to validate these accumulation functions, we compare the normalized accumulation derived from the radar-PWS fusion with that from radar alone, ensuring correlation within an acceptable margin of error while preserving location-specific information. Once the validated accumulation functions are established as reliable ground truth, they are used to train Convolutional Neural Networks and Graph Neural Networks to generate spatially continuous rain rate fields. Our study is conducted over Hawaii. The dataset consists of eight clusters of PWSs, and to test the generalization ability of the model, we leave one cluster out as the testing dataset to evaluate how well the model performs on unseen areas. Additionally, we compare the performance of our method against the Radar Qualitative Precipitation Estimate with gauge bias correction over both trained and test areas.