No abstract

Accurate prediction of air temperature is of great significance to outdoor activities and daily life. However, it is important and more challenging to predict air temperature in complex terrain areas because of prevailing mountain and valley winds and variable wind directions. The main innovation of this paper is to propose a regional temperature prediction method based on deep spatiotemporal networks, designing a spatiotemporal information processing module to align temperature data with regional grid points and further transforming temperature time series data into image sequences. Long Short-Term Memory network is constructed on the images to extract the depth features of the data to train the model. The experiments demonstrate that the deep learning prediction model containing the spatiotemporal information processing module and the deep learning prediction module is fully feasible in short-term regional temperature prediction. The comparison experiments show that the model proposed in this paper has better prediction results for classical models, such as convolutional neural networks and LSTM networks. The experimental conclusion shows that the method proposed in this paper can predict the distribution and change trend of temperature in the next 3 h and the next 6 h on a regional scale. The experimental result RMSE reached 0.63, showing high stability and accuracy. The model provides a new method for local regional temperature prediction, which can support the planning of production and life in advance and tend to save energy and reduce consumption.

The real-time information on surrounding air quality index (AQI) is important for the public to protect themselves from air pollution. Traditional methods have some shortages regarding the estimation time and running efficiency. Consequently, the AQI results cannot meet the needs of personal protection and environmental management. With the popularity of smart terminals, it is easier to collect particular environmental images for AQI estimation tasks. Therefore, a real-time and image-based deep learning model named YOLO-AQI is proposed. Based on the object detection algorithms, the model has better performance regarding the AQI estimation speed. By optimizing the parameter transfer and network structure, the model takes an average of 0.0582 s to perform feature analysis and achieves 75.15% accuracy on AQI estimation tasks. Comparing YOLO-AQI with several image recognition models (VGG, AlexNet, GoogLeNet, MobileNet, and ResNet), it shows that YOLO-AQI outperforms other models by 14.8% on accuracy and by 71.1% on running speed. This method can provide real-time AQI level information for remote areas such as rural.