Detection of Oak Wilt Disease Using Convolutional Neural Network From Uav Natural Color Imagery

In this study, we applied CNN for detecting oak wilt disease from UAV natural color imagery. CNN was trained using training dataset and then applied into multiple test datasets (image A, A’, and B) which have various scene characteristics. In overall, CNN produced high recall for detecting oak wilt disease, while showed relatively low precision. False alarms mainly resulted from shadows between tree crowns. CNN yielded higher detection accuracy for image A which has a very similar scene distribution with training dataset. In contrast, CNN yielded lower detection accuracy for image A’ which acquired at different illumination conditions and for image B which has a different forest structure and species. We modified the patch size of training dataset and CNN architecture related to hyper-parameters such as size and the number of filters and evaluated their performance. The detection accuracy depended on the patch size and CNN architecture.