Application of YOLOv5 and ReLU for Weather Forecasting in Aeronautical Meteorological Services

Abstract

This paper compares the use of deep learning with YOLOv5 and ReLU to create a prediction system for Aeronautical Meteorological Services. The system forecasts five types of weather conditions: BKN, CAVOK, FEW, OVER, and SCT. The evaluative information for the images is based on observing the characteristics of grouped clouds and comparing them with quantities in the sky using the "okta" unit of measurement, which divides each image into eight parts. Thus, the deep learning aspect of this paper involves teaching the model to recognize various image characteristics both during the day and at night to use the forecast results to inform sky conditions. The process described in this article involves three steps: The first step is to train the data with images that have varying characteristics from the five data types. The next step is to test the accuracy of the weights generated from the training and training steps. The final step is to use the weights to create a decision-making system for users. From the experiment, a private dataset in this article used more than 10,000 images in the testing. The experimental results found that the average accuracy results for the YOLOv5 and ReLU algorithms could be measured at 80.88% and 76.82%, respectively.