DAMAGED TREES DETECTION USING THE EXPANSION OF DEEP LEARNING MODEL FROM UAV RGB IMAGES TO MULTISPECTRAL IMAGES

초록

Deep learning-based object detection models require great amounts of training data. Furthermore, deep learning-based models trained from a particular image are not suitable for other test images with different characteristics. In this study, we aim to implement the deep learning-based approach for detecting damaged trees by expanding the training data obtained from the RGB images to the multispectral images in which two images have different characteristics with regard to sensors, spectral bands, spatial resolution, region, and tree disease. Initially, the target multispectral images were translated to the domain of source RGB images by using the CycleGAN. The Faster R-CNN object detector originally developed from the source RGB images was then applied to the translated target multispectral images. Our experimental results showed that the expansion of deep learning-based model works effectively to the target images with higher accuracy.