Combinations of distributional regression algorithms with application in uncertainty estimation of corrected satellite precipitation products

2024-06-29

Machine Learning with Applications

AI Summary

Accurately quantifying uncertainty in satellite remote-sensing precipitation estimates—such as those from GPM IMERG—remains a persistent challenge. To address this, we propose the first uncertainty estimation framework integrating multiple distributional regression models (DNN-based quantile regression, NGBoost, and DeepAR), featuring a novel weighted ensemble strategy to enhance quantile prediction consistency. Methodologically, our approach jointly optimizes quantile loss and employs Bayesian model averaging for end-to-end probabilistic modeling of precipitation error distributions. Evaluated on the GPM IMERG calibration dataset, our framework achieves a 12.7% reduction in Continuous Ranked Probability Score (CRPS), improves 90% confidence interval coverage to 93.5% (+3.5 percentage points), and reduces uncertainty interval width by 18.3%. These results significantly outperform individual models and conventional bias-correction methods, demonstrating superior calibration and sharpness in probabilistic precipitation forecasting.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Relational Probabilistic Models
• Intelligent Robots: State Estimation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Problem

• Satellite Rainfall Estimation
• Error Quantification
• Data Improvement

Innovation

• Combined Learning Method
• Distributed Regression Improvement
• Generalized Additive Models & Random Forests