Authors: Muhammad Adil, Patrick J. Clemins – University of Vermont
Title: Reliability Aware Multi-Source Fusion for QPE Bias Correction in Radar-Blocked and Gauge-Sparse Regions
Presentation Type: Poster Presentation
Abstract: Accurate quantitative precipitation estimation is critical for hydrologic forecasting, flood risk assessment, and climate analysis. Operational radar-only products such as MRMS RadarOnly QPE are spatially complete but physically unreliable in mountainous, beam-blocked, and gauge-sparse regions, where systematic underestimation is most severe.
Gauge-corrected analyses like NCEP Stage IV partially address these biases but may themselves remain biased in the hardest blocked and ungauged areas, making simple radar-to-Stage IV mapping risky as a supervision strategy. We propose a reliability-aware multi-source fusion framework that integrates MRMS RadarOnly QPE with radar accumulation quality indices (RAQI), gauge influence indicators (GII), HRRR precipitation guidance, and terrain features to perform precipitation reconstruction and bias correction under uncertainty.
Rather than assuming uniform target reliability, the model constructs per-pixel target confidence weights and employs a supervision hierarchy – strong at gauge-supported locations, medium under Stage IV coverage, and weak in blocked or ungauged regions – paired with a weighted, uncertainty-aware loss. The model outputs corrected hourly QPE and per-pixel confidence estimates. Results will evaluate bias correction performance in blocked terrain, and calibration of predictive uncertainty.