Low-cost video-based air quality estimation system using structured deep learning with selective state space...

Area of Science:

• Environmental Science and Engineering
• Computer Science (Artificial Intelligence, Machine Learning)

Background:

• Accurate air quality prediction is vital for public health and environmental sustainability.
• Existing models often rely on static images, neglecting the dynamic, temporal nature of air pollution.
• Video-based air quality estimation research is limited, especially for multi-pollutant prediction.

Purpose of the Study:

• To propose Air Quality Prediction-Mamba (AQP-Mamba), a video-based deep learning model for estimating air quality.
• To accurately predict multiple pollutants (PM2.5, PM10) and the Air Quality Index (AQI) using video data.
• To address the limitations of static image analysis by incorporating spatiotemporal features from videos.

Main Methods:

• Developed AQP-Mamba, integrating a structured Selective State Space Model (SSM) with a hybrid predictor.
• Employed spatiotemporal SSM with selective scan and bidirectional processing for dynamic feature extraction.
• Utilized the LMSAQV dataset, comprising 13,176 outdoor videos from Lahore, Pakistan, for training and validation.

Main Results:

• AQP-Mamba achieved high regression performance: R² of 0.91 (PM2.5), 0.90 (PM10), and 0.92 (AQI).
• Excellent classification metrics were obtained: 94.57% accuracy, 93.86% precision, 94.20% recall, and 93.44% F1-score for AQI.
• The model significantly outperformed state-of-the-art video analysis models (VideoSwin-T, VideoMAE, I3D, VTHCL, TimeSformer).

Conclusions:

• AQP-Mamba offers an efficient, scalable, and cost-effective solution for real-time, multi-pollutant air quality estimation.
• The video-based approach captures dynamic air pollution variations, overcoming limitations of static image analysis.
• This method has the potential to supplement data from expensive instruments globally, improving air quality monitoring.