A hybrid deep learning and handcrafted feature extraction with outliers detection for image-based air quality index (AQI) classification

Air quality remains a critical global concern due to its direct impact on human health and the environment. Traditional air quality monitoring methods has limitations in terms of coverage, cost, and accuracy, especially in extreme environmental conditions. This research introduces an AQI classification model using the same dataset as the Eff-AQI model, which leverages image-based data to assess air quality levels in urban environments. The model combines deep learning-based feature extraction with handcrafted features and image outlier detection techniques to address challenges such as the inability to capture complex pollution-specific patterns effectively and the lack of robust mechanisms for detecting outliers or anomalies in data. These issues often lead to inaccurate classifications, particularly in extreme environmental conditions, like high pollution levels, or due to technical issues in image capture, including poor image quality. The model uses VGG16 to extract high-dimensional features from AQI images, which represent visual data captured from various urban locations with different pollution levels. These images offer a comprehensive view of the environmental conditions affecting air quality. Handcrafted features, such as color (RGB) and texture (LBP), further enhance classification performance by capturing finer details, complementing the high-dimensional features extracted by the deep learning model. For training, a Fully Connected Neural Network (FCNN) is employed to classify the AQI levels based on the extracted features. The key contribution in this research is the incorporation of three outlier detection techniques using DBSCAN, LOF, and Isolation Forest, which enable the model to identify and accurately classify extreme pollution events, improving its robustness in challenging scenarios. The model was validated on two datasets, achieving an average precision, recall, and F1-score of 0.97, and an overall accuracy of 96% on the original dataset. For the unseen dataset, the model achieved an average precision of 0.77, recall of 0.75, and an F1-score of 0.75, with an overall accuracy of 77% demonstrating its generalizability across different data distributions. Confusion matrices were also evaluated, providing additional insight into model performance across AQI categories. These results show that the model can effectively classify AQI levels across diverse environmental conditions and offer a more scalable, reliable solution for real-time air quality monitoring. The scope of this study focused on leveraging image-based data for AQI classification, specifically addressing the challenges of capturing pollution-specific patterns and detecting outliers in urban environmental images. By overcoming the limitations of current image-based techniques, this model shows promise for wider application in environmental monitoring and public health initiatives. Future work will focus on expanding the dataset, improving outlier detection methods, and using temporal and weather information to enhance model accuracy and real-time performance.