Conclusions

Summary
This study presented a multi-model comparison for smog level classification using PM2.5 and PM10 AQI values collected from Lahore, Pakistan (2020–2023). Both classical ML and deep learning models were trained and evaluated on the processed dataset.

The CNN, DNN, and LSTM models achieved higher classification accuracy compared to traditional ML methods. Visual analysis using confusion matrices showed that deep models were better at identifying higher pollution levels.

Limitations
1. Data Source Constraints: The OpenWeatherMap API offers limited granularity and may not match precision levels of local environmental monitoring stations.
2. Location Specificity: The dataset represents only Lahore, Pakistan, limiting generalizability to other regions.
3. Feature Limitations: The analysis was based solely on AQI values. Other relevant factors like wind, humidity, or temperature were not included.
4. No Hyperparameter Tuning: Fixed configurations were used for deep learning models without rigorous tuning.
5. Interpretability: Deep learning models offer limited interpretability compared to decision trees or rule-based classifiers.

Future work could involve:
- Incorporating meteorological and spatial features.
- Utilizing real-time multi-sensor fusion.
- Extending the study to other cities and longer periods.