A Machine Learning-Based Framework for Bias Correction of Doppler Weather Radar Observations

Radar constant miscalibration is one of the major sources of uncertainty in the radar-derived rainfall products. The traditional bias correction techniques often struggle to account for the nonstationary and nonlinear nature of bias. This study proposes a novel machine learning-based framework using the extreme gradient boosting (XGBoost) algorithm to model reflectivity bias as a function of ground radar (GR) and spaceborne radar (SR) reflectivity differences, along with radar geometrical parameters (range, azimuth, and elevation). Furthermore, a strategy for near real-time bias correction is proposed based on an ensemble approach that combines an offline-pretrained model with an adaptive online learning component, incrementally updating the output as new data becomes available. This allows the model to adapt to evolving bias patterns over time. The results indicate that the proposed technique consistently outperforms the traditional iterative method. The results point toward its potential in reducing bias in near real-time for improved quantitative precipitation estimation (QPE) and other applications.