Deep Learning for Aerial Tree Detection. A Comparative Analysis of YOLOv7 and DETR

Abstract

This thesis explores the application of modern object detection techniques for automated tree identification in aerial imagery. The primary objective is to develop a robust and accurate method for detecting and counting trees in Austria, using the city of Vienna as a case study. Since no suitable publicly available datasets for tree detection in Austrian aerial imagery exist, custom datasets were created through an iterative process, consisting of over 5,500 manually and semi-manually annotated aerial images with more than 40,000 tree labels. All datasets have been made publicly available. Two state-of-the-art object detection architectures were compared: the convolutional YOLOv7 model and the transformer-based DETR model. Beyond analyzing the model architectures, the focus was on data preparation, the training process, and comprehensive evaluation using standard metrics such as Intersection over Union (IoU), Precision, Recall, and Mean Average Precision (mAP). The results show that the DETR model, despite its higher complexity and longer training times, delivers significantly better results in terms of accuracy and robustness. YOLOv7, on the other hand, offers marginally faster inference times and simpler handling. This work demonstrates that transformer-based models are highly suitable for precise detection of individual trees in aerial imagery, although complete detection of all trees is practically unattainable due to occlusions and overlapping tree canopies. At the same time, it highlights how critical high-quality, domain-specific training data is for the success of machine learning projects in remote sensing. Finally, potential approaches for future work are discussed, such as the use of advanced training methods and newer model generations.