Machine learning-based prediction of drone noise propagation in three-dimensional urban environments

Predicting noise emissions from drones and urban air mobility is challenging due to the wide variety of vehicle configurations, rich sources of sound, and operation in complex urban terrain environments. In this work, a supervised U-shaped neural network, specifically a convolutional neural network with multiple output layers, was designed to efficiently and accurately predict sound propagation in a three-dimensional urban environment. Five distinct layers divide an urban area into different altitudes, incorporated into the neural network to supervise the learning process. The results indicate that a well-trained neural network accurately predicts sound propagation in diverse urban environments, achieving a mean error of pressure below 5% and significantly reducing computation time from hours to minutes compared to traditional methods like Gaussian beam tracing. The results demonstrate potential for acoustic assessment in urban acoustic studies.