A Mobile Image-Driven PM2.5 Estimation Framework Using Deep Learning Techniques

Particulate pollution (PM2.5) is an important concern in Asian countries owing to its health hazards. When planning outdoor activities, understanding the PM2.5 concentration measurement is essential. Because of the lower number of government-run Air Quality Monitoring Stations, other options for obtaining location-specific PM2.5 concentration values are sought. This paper proposes using photo image processing to estimate the PM2.5 concentration. This research aims to improve the efficacy and reduce the computational complexity of the PM2.5 concentration estimation process. The proposed Efficient PM2.5 estimation framework uses EfficientNet-B1 and BiLSTM to estimate PM2.5 concentrations. Met-EfficientNet-B1-BiLSTM was designed and implemented to incorporate the meteorological features - temperature, wind speed, and humidity to further improve the estimation accuracy. The EfficientNet-B1 neural network is applied in the image feature vector extraction process. EfficientNet-B1, with a resolution of $240\times 240$ pixels, was determined to be the optimal variant of EfficientNet for a small dataset of images needed for the estimation of the PM2.5 concentration value. The BiLSTM was used for the regression of these image features with PM2.5 concentration values to obtain the estimated PM2.5 concentration. A dataset comprising HDR and non-HDR images was explicitly created for this study to compare the types of images that improve the accuracy of PM2.5 concentration estimation and the feature extraction process. The proposed Efficient PM2.5 estimation framework reduces computational complexity and outperforms ResNet-18-LSTM by improving the efficacy by 5.75% in MAE and 11.43% in SMAPE metrics. The proposed Efficient PM2.5 estimation framework demonstrates that the mobile image can be efficiently used for PM2.5 concentration estimation.