Abstract
Air quality index (AQI) monitoring is critical for public health, yet traditional methods struggle to balance spatial resolution with operational efficiency. We present ARMS, a UAV-based sensing system designed to generate fine-grained, 3D AQI maps in real-time. ARMS utilizes a novel hybrid model, the Gaussian Plume Model-based Neural Network (GPM-NN), to characterize particle dispersion. Complemented by an adaptive monitoring algorithm that prioritizes high-variance areas, the system achieves significant energy savings—up to 90%—while maintaining high estimation accuracy.
Introduction
Traditional AQI monitoring relies on sparse static stations that provide high accuracy but low spatial resolution, often missing "micro-climates" in dense urban environments. While mobile sensing via ground vehicles has been explored, it typically produces 2D maps and lacks the ability to capture vertical pollution profiles. UAVs offer the mobility needed for 3D profiling, but their limited battery life poses a significant challenge for exhaustive area coverage. ARMS addresses this by combining physical dispersion modeling with intelligent, selective sampling.
Methodology
GPM-NN: A Hybrid Dispersion Model
To accurately predict AQI across a volume with sparse data, we developed the GPM-NN model. This architecture integrates the physical principles of the Gaussian Plume Model (GPM) into a neural network framework. The GPM provides a structured baseline for how pollutants disperse from sources (like roads), while the neural network layers learn the complex, non-linear perturbations caused by urban geometry and local weather conditions.
Adaptive Monitoring Algorithm
To overcome the battery constraints of UAVs, ARMS employs an adaptive monitoring strategy. Instead of a standard grid-based "Complete Monitoring" sweep, the system identifies critical "Target Cubes" using a Partial Derivative Threshold (PDT). By focusing on areas where AQI levels are changing most rapidly, the UAV can capture the most informative data points with minimal flight time.
The resulting path is optimized using a greedy algorithm to solve a Traveling Salesman Problem (TSP), ensuring the most energy-efficient trajectory between selected points.
Experimental Results
The system was validated using a custom dataset collected over 100 days at Peking University, Beijing. Experiments covered both 2D open spaces and 3D enclosed courtyards.
- Accuracy: GPM-NN consistently outperformed baseline models like Linear Interpolation and Multi-variable Linear Regression.
- Efficiency: In 2D scenarios, the adaptive algorithm reduced battery consumption by 90% while maintaining ~80% estimation accuracy.
- Vertical Insights: The system successfully identified vertical pollution "pooling" in courtyards, a phenomenon invisible to ground-level sensors.
Conclusion
ARMS demonstrates that combining physical models with machine learning and adaptive sampling can enable efficient, high-resolution environmental monitoring. Future work will explore multi-UAV coordination to expand coverage and refine models for more complex pollution sources.
Reference: Yang, Y., Zheng, Z., Bian, K., Song, L., & Han, Z. (2017). Realtime Profiling of Fine-Grained Air Quality Index Distribution using UAV Sensing. arXiv:1711.02821v1 [eess.SP].
Summary updated