Prediction of Atmospheric Refractivity From Clutter Power Images Using a Convolutional Neural Network and a Trilinear Atmospheric Model

In this paper, we propose a method for predicting atmospheric conditions from clutter power images using a convolutional neural network (CNN). The proposed refractivity from clutter (RFC) method employs a CNN model that utilizes clutter power images as input. These input data, generated by the Advanced Refractive Prediction System (AREPS) EM simulator, using four variables of a theoretical trilinear atmospheric model. The proposed RFC method outputs values for these four variables, which can be used to predict the current atmospheric conditions in a short time. This CNN model is trained on 50,000 data samples, with a training-to-validation data ratio of 4:1. Furthermore, as cross-validation, the trained CNN model is utilized to predict atmospheric conditions using actual data measured in Baengnyeongdo, and the findings show that increasing the number of data samples improves performance. These results demonstrate the capability of the proposed RFC method to predict various atmospheric conditions in real time.