DeepAQI: A Vision-Based EfficientNet Framework for Air Quality Index Prediction from Environmental Metadata

Abstract This research presents a deep learning–based framework for predicting the Air Quality Index (AQI) using outdoor webcam images from U.S. National Parks. Traditional AQI measurement relies on ground‑based air sensors, which require continuous calibration and are often geographically sparse. Recent advances in computer vision have motivated the exploration of image-based AQI estimation, enabling scalable, low‑cost, and real‑time monitoring in remote regions. Leveraging the publicly available NPS_AQI_DB, which includes more than 120,000 webcam images paired with corresponding pollutant data (O₃, SO₂, RH, temperature), we investigate both image‑only and multi‑modal (image + tabular) models for AQI regression. Our primary architecture is EfficientNet‑B0, chosen for its strong performance–efficiency trade‑off. The model is trained on augmented 224×224 images using AdamW optimization, mixed‑precision training, and robust preprocessing to handle missing and corrupted data. To enhance interpretability, Grad‑CAM visualizations highlight regions influencing AQI prediction, often corresponding to sky visibility, haze thickness, and lighting conditions. Additionally, we evaluate a two‑tower fusion model combining CNN features with meteorological variables, demonstrating improved stability across pollution categories. Experimental results show that the best-performing model achieves MAE ≈ 11.2 and RMSE ≈ 15.3 on the test set, reflecting competitive performance given the inherent noise of visual air estimation. Comprehensive evaluation includes residual analysis, calibration curves, AQI‑bin MAE breakdown, per‑park performance, and temporal error trends. These findings confirm that vision-based AQI prediction is feasible and can supplement traditional monitoring networks, especially in visually accessible but sensor-limited environments. Future work will explore temporal modeling, domain adaptation, and deployment on low-power edge devices.