Daily PM2.5 Estimation for Pakistan Using Sparse Ground Monitoring
Pakistan experiences severe PM2.5 episodes, yet routine ground monitoring remains sparse and concentrated in a few major cities. Large parts of the country have no reliable daily measurements. This study presents a daily PM2.5 estimation model for Pakistan that combines nationally available satellite, atmospheric and meteorological features with observations from monitored stations.
The Problem
Recent PM2.5 history improves estimates near active monitoring sites, but it is unavailable across most of the national domain. A standard temporal model trained on that history loses substantial accuracy when it is applied where no monitor exists.
Approach
We train a single tree-based model and randomly withhold the recent-monitor-history features for a subset of the training data. The model therefore learns to operate both with and without local PM2.5 history:
- Near monitors: it uses recent ground observations to sharpen the estimate.
- Elsewhere: it relies on satellite, atmospheric and meteorological features alone.
Results
Evaluated on the held-out 2025 test year:
| Setting | Mean absolute error |
|---|---|
| Local PM2.5 history available | 14.3 µg/m³ |
| Local PM2.5 history absent | 23.2 µg/m³ |
The benefit of local history is strongest during severe episodes. At the 150 µg/m³ threshold, the F1 score for identifying episode days rises from 0.729 without local history to 0.838 with it.
Why It Matters
The approach preserves the value of recent ground observations without the accuracy loss that a standard temporal model suffers where monitoring is absent. The result is a daily national PM2.5 estimate that makes full use of Pakistan’s limited monitoring network.
Daily national PM2.5 estimates for Pakistan are published on Hawanama.