An accurate two-stage deep machine learning aided air quality estimation based on multiple gases from aerial images

Currently, industry-related environmental factors negatively affect human health. The Air Quality Index (AQI), a measurement of air pollution, is worsening and affecting our daily life. Increasing air pollution, concerns about climate change, evolving technologies, and environmental research estimate air quality as increasingly important. Air pollution management has become crucial, and environmental monitoring conditions and air quality in daily life are necessary. The latest research on AQI does not focus on aerial images or specific pollutants. Additionally, avoid tackling the issue of having a high level of some specific weather elements, such as temperature and humidity. In the proposed research, we tackle the issues mentioned above by splitting our investigation into two phases. Thus, in this paper, an image-based high-accuracy air quality monitoring system is realized in two stages. In the first stage of our examination, popular deep machine learning prediction algorithms, including Artificial Neural Networks (ANN), Convolution Neural Networks (CNN), CNN-Long Short Term Memory (CNN-LSTM), and MobileNetv2 networks, for the assignment into different categories of the AQI, targeting the AQI prediction with the utilization of aerial images. Next, the second stage utilizes ANN, Recurrent Neural Networks (RNNs), LSTM, and RNN-B-LSTM popular deep machine learning techniques to develop a regression prediction system trained with the features and the resulting classification of air quality results from the highest accuracy approach of the first stage (which is the ANN approach with 99.14% accuracy). To tackle the issue of overfitting/underfitting, the oversampling technique (SMOTE) is used. The proposed method considers the contribution of each significant pollutant gas (i.e., PM2.5, PM10, NOx, NH₃, SO₂, O₃) to the overall air quality index (AQI). The weather elements such as temperature and humidity impact air quality; they are also considered in our examination, resulting in a more accurate forecast. In addition, a comparison study is performed to identify the best model architecture for the proposed prediction system. Finally, the results show that the first stage of image classification should be executed with ANN (99.14 % accuracy), which is a highly accurate approach, and the second stage should be executed with the RNN-B-LSTM approach, again because of its high accuracy (98.73 % accuracy).