Real-Time Long Horizon Air Quality Forecasting via Group-Relative Policy Optimization

AI summary

Overview

Research area: Computer vision and spatiotemporal machine learning for environmental forecasting, specifically real-time particulate matter (PM) air quality prediction. Technical level: Advanced. Scope: The paper introduces an East Asia air quality dataset and a Group-Relative Policy Optimization framework for 48-120 hour real-time PM forecasts that reduce false alarms while maintaining accuracy.

What This Paper Is About

Existing global foundation models such as Aurora are general but often miss region-specific atmospheric dynamics and rely on non-real-time inputs, limiting their usefulness for local air quality warnings. The goal is to build a real-time, long-horizon PM forecasting system for East Asia that aligns with public health priorities, where false alarms harm trust and missed severe pollution events endanger populations.

Key Contributions

1. Constructs and releases the real-world observations and high-resolution CMAQ-OBS dataset for East Asia, supporting real-time 48-120 hour forecasts and reporting a 59.5% reduction in regional error.
2. Identifies a mismatch between standard point-wise training objectives and asymmetric operational costs, showing that supervised fine-tuning (SFT) models over-predict and produce high false alarm rates.
3. Introduces Group-Relative Policy Optimization (GRPO) with class-wise rewards and curriculum rollout to align forecasts with operational priorities.
4. Demonstrates a 47.3% reduction in False Alarm Rate over the SFT-only baseline while achieving a competitive F1-score, with code and dataset publicly available.

Main Findings

• Dataset advances: The CMAQ-OBS East Asia dataset reduces regional error by 59.5% and enables real-time 48-120 hour forecasting.
• Baseline failure mode: SFT models trained with standard point-wise losses over-predict because they do not account for asymmetric operational costs.
• Operational alignment works: GRPO with class-wise rewards and curriculum rollout lowers the False Alarm Rate by 47.3% compared with SFT-only.
• Accuracy preserved: The approach achieves a competitive F1-score, so fewer false alarms do not come at the expense of overall forecast skill.
• Practical readiness: The framework improves reliability for long lead time air quality alerts in a region with complex terrain and strong atmospheric dynamics.

Methodology in Plain English

The researchers first assembled a regional dataset for East Asia by combining real observations with high-resolution CMAQ atmospheric chemistry model output. They used this dataset to train a model that predicts PM concentration fields several days ahead. Standard training treats every error equally, but air quality warnings have unequal costs: a false alarm can erode public trust, while a missed severe pollution event can harm health. To fix this, they adapted GRPO, a policy optimization method that compares groups of candidate predictions and updates the model toward better relative choices. They added class-wise rewards so the model learns to avoid false alarms while still catching severe events, and used a curriculum rollout to gradually train on harder forecasting scenarios. They then compared this against a supervised fine-tuning baseline.

Why This Matters

Impact on research: It moves environmental forecasting beyond point-wise accuracy toward operationally aligned objectives, introduces policy optimization to air quality prediction, and provides a new East Asia benchmark for long-horizon PM forecasting.

Real-world applications:

• Public health agencies can issue 2-5 day PM warnings with fewer false alarms, preserving trust while protecting vulnerable populations.
• Local governments and environmental regulators can prioritize interventions during severe pollution episodes.
• Schools, hospitals, outdoor workers, and smart-city systems can use forecasts for exposure-reduction advisories.
• Regional air quality management across East Asia can benefit from better real-time, terrain-aware predictions.

Industry relevance: Environmental monitoring and forecasting companies, weather and air quality service providers, geospatial AI developers, and public health technology platforms can use the dataset and GRPO framework to build more reliable operational warning systems.

Future Directions

• Extend the dataset and GRPO framework to other pollutants and regions with complex terrain or strong atmospheric dynamics.
• Add uncertainty quantification and calibration so alerts can express confidence levels for decision-makers.
• Integrate more real-time data streams, such as satellites and ground sensor networks, to improve latency and coverage.
• Test the reward design and curriculum rollout in operational agency settings and other geophysical forecasting tasks.

Target Audience

This paper is most useful for advanced machine learning researchers working on spatiotemporal forecasting, reinforcement learning from operational feedback, and geospatial AI; environmental scientists and meteorologists; air quality and public health officials; and engineers building operational forecasting or warning systems.

Authors’ abstract

Accurate long horizon forecasting of particulate matter (PM) concentration fields is essential for operational public health decisions. However, achieving reliable forecasts remains challenging in regions with complex terrain and strong atmospheric dynamics such as East Asia. While foundation models such as Aurora offer global generality, they often miss region-specific dynamics and rely on non-real-time inputs, limiting their practical utility for localized warning systems. To address this gap, we construct and release the real-world observations and high-resolution CMAQ-OBS dataset for East Asia, reducing regional error by 59.5% and enabling real-time 48-120 hour forecasts critical for public health alerts. However, standard point-wise objectives cannot reflect asymmetric operational costs, where false alarms deteriorate public trust while missed severe events endanger populations. This cost mismatch causes SFT models to over-predict and yield high False Alarm Rates. We introduce Group-Relative Policy Optimization (GRPO) with class-wise rewards and curriculum rollout to align predictions with operational priorities. Experimental results demonstrate that our framework significantly improves the reliability of the forecast. Compared to the SFT-only baseline, our model reduces the False Alarm Rate by 47.3% while achieving a competitive F1-score, proving its effectiveness for practical, real-world air quality forecasting systems on long lead time scenarios. Code and dataset are publicly available at https://github.com/kaist-cvml/FAKER-Air.