Multimodal Particulate Matter Prediction: Enabling Scalable and High-Precision Air Quality Monitoring Using Mobile...

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Persistent air pollution challenges necessitate innovative monitoring solutions.
• Traditional monitoring systems can be costly and centralized.
• Mobile devices offer potential for distributed environmental sensing.

Purpose of the Study:

• To develop and evaluate a novel system for predicting Particulate Matter (PM) concentrations using mobile camera devices.
• To leverage deep learning and multimodal frameworks for air quality estimation from imagery.
• To assess the feasibility of edge deployment for real-time environmental monitoring.

Main Methods:

• Utilized transformer-based deep learning architectures, including Contrastive Language-Image Pre-Training (CLIP).
• Developed a baseline using time-series models (NLinear) for 1D PM signal prediction.
• Implemented a CLIP-based system for 2D image analysis of smartphone-captured environmental scenes.

Main Results:

• Linear models (NLinear) outperformed complex transformers for short-term 1D PM forecasting.
• The CLIP-based system achieved Top-1 accuracy of 0.24 and Top-5 accuracy of 0.52 on diverse images.
• Demonstrated viable edge deployment with processing times of 0.29s (GPU) and 2.68s (SBC) per image.

Conclusions:

• Transformer-based multimodal models show promise for mobile sensing applications in environmental monitoring.
• The research provides a practical framework for distributed air quality monitoring using accessible hardware.
• Future work may involve seasonal data expansion and architectural refinements for improved accuracy.