New method improves detection of rare, intense rainfall using satellite data

A research team has developed a novel machine-learning technique to address a persistent problem in satellite-based rainfall monitoring: artificial intelligence systems tend to underestimate rare, high-intensity storms because training data is heavily skewed toward light rain and no-rain conditions.

The Problem with Imbalanced Data#

When training AI models for rainfall retrieval from satellite infrared brightness temperatures, most observations in the dataset represent either zero rainfall or light precipitation. This imbalance causes models to become biased toward predicting common, lower-intensity conditions and systematically underestimate the frequency and magnitude of extreme rainfall events—precisely the events most critical for flood forecasting and severe weather monitoring.

How Hurdle-RMIL Works#

The proposed Hurdle-Retrieval Model Imbalanced Learning (Hurdle-RMIL) approach uses a divide-and-conquer strategy. Rather than treating all rainfall values as a single imbalanced distribution, the method separates the problem into two distinct challenges:

• A hurdle model addresses the zero inflation (the abundance of no-rain observations)
• RMIL (Retrieval Model Imbalanced Learning) handles the long-tailed distribution of positive rain amounts using a Bayes-based mathematical transformation

This transformation exploits the physical invariance of the rainfall-to-satellite forward process under fixed observation conditions, allowing the model to learn a balanced-distribution behavior from naturally imbalanced data without requiring synthetic rebalancing.

Performance Improvements#

Testing across multiple regions in China demonstrated that Hurdle-RMIL delivers meaningful gains for extreme events while maintaining competitive accuracy for routine predictions:

• At 30 mm per hour, the method achieved an equitable threat score (ETS) of 0.051 versus 0.015 for the best conventional baseline
• Mean error improved to −25.41 mm per hour compared to −28.98 mm per hour for baseline approaches
• Root mean square error remained competitive with top baselines across the 0.1–10 mm per hour range
• At most high-intensity thresholds, Hurdle-RMIL achieved the highest ETS scores

Case studies also revealed improved spatial and intensity representation of rainfall patterns, confirming the method's ability to capture both extreme and moderate rainfall events more faithfully.