Real time image-based air quality forecasts using a 3D-CNN approach with an attention mechanism

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Accurate air quality forecasting is crucial for public health and environmental monitoring.
• Existing methods often struggle with complex environmental variables and temporal dynamics.
• Image-based data offers a rich, underutilized source for air quality assessment.

Purpose of the Study:

• To develop and validate an image-based deep learning model for improved air quality recognition and multi-horizon forecasting.
• To integrate a three-dimensional convolutional neural network (3D-CNN) and gated recurrent unit (GRU) with an attention mechanism.
• To enhance the extraction of spatio-temporal features and mitigate fluctuations in particulate matter data.

Main Methods:

• A hybrid deep learning architecture combining 3D-CNN for feature extraction and GRU with an attention mechanism for temporal analysis.
• Utilized the Shanghai scenery dataset with corresponding air quality monitoring data for model training and validation.
• Incorporated an attention mechanism to dynamically adjust feature influence and reduce noise in particulate matter predictions.

Main Results:

• The proposed model demonstrated superior forecasting accuracy compared to existing state-of-the-art methods.
• Achieved reliable multi-horizon predictions through efficient feature extraction and effective noise reduction.
• Validated the model's feasibility and reliability using real-world site images and air quality data.

Conclusions:

• The developed image-based deep learning approach offers a promising solution for accurate air quality forecasting.
• The hybrid 3D-CNN-GRU model with attention effectively captures complex environmental dynamics.
• The method provides valuable early warning guidelines for air pollutants, aiding environmental management.