An image-based PM2.5 prediction model using a supervised learning algorithm with deep features fusion

Air pollution has been increasing due to different factors, such as urbanization, industrial growth, vehicle emissions, climatic changes and so on. Particulate Matter 2.5 (PM2.5) is extremely fine and is more harmful than other pollutants, as it can easily penetrate the respiratory system of living beings. Due to its size, it is extremely difficult to detect PM2.5 from images. This study presents a PM2.5 prediction model based on fused deep features and assisted by a Machine Learning (ML) algorithm. The study is based on key phases such as pre-processing, deep feature extraction, feature fusion, feature selection and PM2.5 prediction. CLAHE and Z-score normalization techniques are utilized to enhance contrast and normalize the images. Deep features are extracted from EfficientNet, ResNet50 and MobileNetV3, followed by feature fusion. Due to the maximum dimensionality, Recursive Feature Elimination with Cross-Validation (RFECV) is employed to select key features. Random Forest (RF) classifier is employed for classifying between six Air Quality (AQ) classes, through which the PM2.5 range is predicted. The performance of the proposed study is analyzed in terms of standard performance measures such as accuracy, precision, recall and F-measure in different aspects. This study proves better performance by showing generalization and robustness.