부산대학교 도서관

Wind turbines can affect radar systems used for missions in defense, aviation safety, and weather forecasting. Both range and Doppler reflections from wind turbines can create clutter’ reduce detection sensitivity, and obscure potential targets. An accurate and up-to-date database of wind turbine locations and status is needed in order to assess their impact on radar functionalities. Overhead and satellite images are important data sources to assess the status of wind turbines. To significantly reduce human workload in classifying the status of wind turbines through overhead imagery, we trained classifiers using convolutional neural networks (CNN) to automate the wind turbine status classification problem. We leveraged the existing AlexNet and GoogLeNet CNN frameworks, paired with overhead imagery from Google maps, to train a classifier to determine whether a turbine structure exists at a location where one has been registered. The automated classifier achieved an average accuracy of 97%. Our classifier can act as a highly efficient prefilter to select wind turbines for status verification by human operators. 1