Error Structure Analysis for Multiple Dual-polarization Radar using Ensemble Learning

Abstract

Accurate quantitative precipitation estimation (QPE) was essential for the prediction and prevention of natural disasters. Recently, radar has been attracting attention as a technique for performing QPE with high spatiotemporal resolution. In particular, the QPE was improved by introducing a dual-polarization technique that observed several hydrometeorological variables at various scales compared to the single polarization radar that utilized only the existing reflectivity-precipitation relationship. This study aimed to analyze the error structure of dual-polarization radar by predicting gauged rainfall using ensemble models. The location of the radar was Gwanaksan, Garisan, and Gwangdeoksan which belonged to the Bukhan river basin. The Pearson correlation coefficients between reflectance-precipitation and gauged rainfall were examined initially. After that, each gauged rainfall was predicted using ensemble learning, which included random forest (RF), gradient boost, and XGboost. Mean absolute error, root mean squared error, and R squared were evaluated as the predictive performance. Hyper-parameters were optimized by 5-fold cross-validation, and the reliability of the research results was obtained by 10 iterations. The result showed that RF had better predictive performance than other models. Gauged stations that operated at high altitudes were not good enough for the QPE because of the mountain effect. Due to the mountain range effect, the importance of the Gwanaksan radar at an altitude of 3 km was very high.