A Probability Density-Based Visual Analytics Approach to Forecast Bias Calibration.

Renpei Huang,Quan Li,Li Chen,Xiaoru Yuan

DOI: https://doi.org/10.1109/tvcg.2020.3025072

IF: 5.2

2022-01-01

IEEE Transactions on Visualization and Computer Graphics

Abstract:Biases inevitably occur in numerical weather prediction (NWP) due to an idealized numerical assumption for modeling chaotic atmospheric systems. Therefore, the rapid and accurate identification and calibration of biases is crucial for NWP in weather forecasting. Conventional approaches, such as various analog post-processing forecast methods, have been designed to aid in bias calibration. However, these approaches fail to consider the spatiotemporal correlations of forecast bias, which can considerably affect calibration efficacy. In this article, we propose a novel bias pattern extraction approach based on forecasting-observation probability density by merging historical forecasting and observation datasets. Given a spatiotemporal scope, our approach extracts and fuses bias patterns and automatically divides regions with similar bias patterns. Termed BicaVis, our spatiotemporal bias pattern visual analytics system is proposed to assist experts in drafting calibration curves on the basis of these bias patterns. To verify the effectiveness of our approach, we conduct two case studies with real-world reanalysis datasets. The feedback collected from domain experts confirms the efficacy of our approach.