Image-based Classification of Air Pollution Using Different Pretrained Cnn Models And A Small Dataset

Abstract

Because contaminants and particles are volatile, dynamic, and highly variable in both time and place, predicting air quality is a challenging endeavor. At the same time, the ability for modeling, predicting, and monitoring the quality of air is becoming increasingly pertinent. In this study, we demonstrate that utilizing several pretrained convolutional neural network models, such as ResNet18, ResNet50, ResNet101, Mobilenetv2 and Shufflenet is feasible to anticipate fine particulate matter (PM2.5) concentrations with minimal computation time. The results show that, it is possible to estimate the PM2.5 level through pretrained models using a small single scene dataset. The models are tested with 2500 images and trained with (50% for training, 25% for validation, and 25% for testing). Among all the models, ResNet101 has the highest accuracy prediction (Acc = 86.27% at LR = 0.0007) with an average learning time about (92 minutes) followed by ResNet50 that achieved a prediction accuracy equal to (Acc = 84.19% at LR=0.00007) with about half the needed learning time that is about (40 minutes). These followed by Shufflenet (Acc = 83.97% with about 44 minutes), and Mobilenetv2 (Acc = 82.70% with about 40 minutes). It is also noticeable that ResNet18 has a reasonable accuracy (Acc = 83.28%) with the least needed learning time about (16 minutes).