Dynamic feature-adaptive representation learning for spatiotemporal air pollution forecasting

Area of Science:

• Environmental Science
• Data Science
• Public Health

Background:

• Atmospheric pollution is a significant threat to public health and urban sustainability.
• Existing air quality prediction models often fail to capture complex spatiotemporal variations due to static feature selection.
• Densely populated metropolitan areas present unique challenges for accurate pollution forecasting.

Purpose of the Study:

• To introduce a novel dynamic feature selection approach for spatiotemporal air quality forecasting.
• To improve the accuracy and interpretability of air pollution prediction models.
• To address the limitations of traditional static feature selection methods in environmental modeling.

Main Methods:

• Development and application of Feature-Adaptive Selection and Transformation (FAST), an attention-based method for dynamic feature selection.
• Evaluation of FAST using real-world air quality data from over 90 monitoring stations across Delhi, Mumbai, and Bengaluru.
• Benchmarking FAST against conventional feature selection techniques (filter, wrapper, embedded).

Main Results:

• FAST demonstrated consistent improvements in forecasting accuracy across different time horizons (6-hour and 48-hour).
• The system achieved up to 18% lower Mean Absolute Error (MAE) and up to 12% improvement in R² values compared to traditional methods.
• FAST provided interpretable insights into seasonal and spatial pollution patterns.

Conclusions:

• FAST offers a dynamic and adaptive approach for data-driven air quality forecasting in urban environments.
• The method effectively captures spatiotemporal variations, leading to enhanced predictive performance and interpretability.
• This approach can aid in developing more effective strategies for managing urban air quality.