다양한 CNN 모델을 이용한 드론 영상 기반 태양광 패널 청결 상태 분류: 구미 지역 사례 연구

초록

This study utilizes drone imagery to classify the cleanliness of photovoltaic (PV) panels in Gumi, South Korea supplemented by publicly available Kaggle datasets. Despite the presence of an industrial complex, contamination from air pollution was minimal, and consisted primarily of bird droppings due to seasonal factors (during the period from September to November). Four CNN-based models (VGG, ResNet, MobileNet, and EfficientNet), pre-trained on ImageNet, were evaluated. All models except VGG achieved significant classification performance, with EfficientNet performing the best. However, the supplementary Kaggle data played a limited role in classification accuracy, highlighting the need for region-specific datasets. This study confirms the importance of regional factors in PV panel contamination monitoring. Future research should explore seasonal contamination patterns and methods to address class imbalance for improved classification performance.

키워드