The propagation of air contaminant PM2.5 that threatens public health is hard to predict because it is affected by long short-term measurements involving many atmospheric variables. Air quality prediction systems provide initial information to increase public awareness and are expected to reduce the long-term health impact on public health. However, these heterogeneous sensory systems are not feasible. They are essentially incompatible and computationally expensive due to their massive deployments of sensory nodes. In this study, a collaborative prediction model is proposed to extract spatiotemporal features from real-world scientific datasets which are collected from government monitoring sites and from community-driven microsites in Taiwan. This study inherits the basic idea of horizontal aggregated learning to generate a more robust prediction model by enhancing features of the dataset. In this study, a prediction model i.e., called sparse-fault-tolerant deep learning (SFT-DL) model is designed using combinations of convolutional neural network (CNN) layers and long-short-term memory (LSTM) layers to forecast the PM2.5 propagations. In a nutshell, the proposed model achieves accurate predictive results than the baseline CNN and LSTM by considering the relationship among long short-time measurements. In addition, the collaborative learning framework boosts the robustness of the prediction model, which is assessed using point-based evaluation.
