Deep Learning Based Detection of Insect Damage in Forests Using UAV Imagery

Abstract

Early timely diagnosis of the location and extent of damage caused by pests is ecologically and economically crucial for early detection of damage to forests. Automatic classification of damaged and healthy trees using high-resolution images acquired with Unmanned Aerial Vehicle (UAV) sensors can be an alternative solution to determine the local extent and impact areas of damage before entomological control. Therefore, in this study, we detect the damage of Thaumetopoea wilkinsoni Tams 1924 (Pine processionary moth) species on larch trees using UAV images. We propose a system for automatic damage detection by creating binary segmentation models with Convolutional Neural Networks (CNN) from deep learning architecture. We experimentally used different CNN models in different combinations as backbones and encoders. We used UAV images with healthy and damaged trees as datasets. To determine the best performing UAV image and segmentation model, we split the dataset into UAV images and orthoimages obtained from the images. We analyzed the DeepLabV3 + + and Unet + + architectures as segmentation models, SE-NET, Efficientnet-B6 and Efficientnet-B7 architectures as encoders as binary segmentation models for the two different types of datasets. In the binary CNN models used to detect the damage area, Unet + + architecture with SE-NET encoder has the best overall accuracy rate (0.90) on the UAV image set, while DeepLabV3 + + architecture with SE-NET encoder has the best overall accuracy rate (0.81) on the UAV orthoimages. The experimental results showed higher accuracy for raw UAV imagery compared to ortho-images in our experiments; however, this difference may be influenced by dataset size and balance in addition to image processing steps such as orthorectification.