@

초록

OpticSatNet is a lightweight convolutional neural network developed for onboard satellite image selection. It operates in orbit to transmit only useful images to the ground station, thus reducing communication costs for satellite constellations. In this study, we simulated satellite constellation operations with and without artificial intelligence-based image selection using Ansys STK(Systems Tool Kit). Highway images were classified within a full dataset and then downlinked only when selected by the model. We compared the communication costs of conventional operations with those of artificial intelligence-assisted selective downloads. The proposed method reduced communication costs by up to 90.74% in a hypothetical scenario. Classification performance of OpticSatNet was evaluated against seven baseline models, including LeNet-5, CNN + DA + DO + Adam, ShallowNet, ZFNet + RF, S-CNN-RGB, miniVGGNet and AlexNet on independent test datasets. OpticSatNet achieved accuracy of Matthews correlation coefficient of 0.97 and F1 score of 0.96. Compared with the miniVGGNet used in the KITSUNE Sat mission, OpticSatNet required 57.6% fewer Multiply-accumulate operations, demonstrating efficiency and suitability for space missions.

키워드