YOLO Models for Instance Segmentation of Individual Tree Crowns from Aerial Imagery in Wellington

<p><strong>The instance segmentation task of individual tree crowns is an important real-world application that facilitates forest management, carbon storage estimation, and biodiversity modelling. Recently, Convolutional Neural Networks (CNNs) have achieved great success in computer vision. Several efforts have applied CNNs to perform instance segmentation of tree canopies. Unlike typical segmentation scenarios, aerial imagery of tree crowns often features densely distributed small and medium crowns, overlapping crowns, varied species, and challenging backgrounds. This variability poses significant challenges to traditional instance segmentation methods. You Only Look Once (YOLO) has recently gained popularity as a rapid and powerful approach for object detection and instance segmentation. Its one-stage design is computationally efficient and particularly appealing for large-scale imagery analysis. Nevertheless, standard YOLO models may struggle with very small or overlapping tree crowns, scale variation, and subtle inter-class differences, underscoring the need for specialized enhancements when applied to canopy segmentation tasks.</strong></p><p>The overall goal of this thesis is to address these challenges by leveraging YOLO-based instance segmentation methods and tailoring it to the unique requirements of aerial tree crown identification and species classification. Specifically, this research focuses on designing robust detection frameworks optimized for small and medium objects, devising novel multi-scale feature extraction and fusion techniques, and crafting effective yet efficient network architectures for aerial canopy data.</p><p>First, this thesis proposes a new YOLO method for identifying individual tree crowns based on YOLOv7. It introduces a specialized detection mechanism for small and medium tree crowns, employs dense connectivity in the backbone to reuse feature maps across layers, and incorporates an efficient attention module to capture long-range dependencies. Experiments on the Wellington, New Zealand (NZ), aerial canopy dataset demonstrate that this method achieves higher detection and segmentation accuracies than other commonly used baselines.</p><p>Second, this thesis proposes a new efficient YOLOv8 method, optimized for precise instance segmentation and species classification of tree crowns. This method includes new schemes for selecting candidate positive samples for each instance and a refined network design tailored for small and medium tree crowns. Adjustments in hyperparameters, particularly within the Task-Aligned Assigner, are also discussed to better suit canopy segmentation tasks. Comprehensive experiments conducted on the canopy dataset demonstrate that the new approach not only outperforms a number of advanced methods in terms of the Box AP and Mask AP metrics but also achieves a substantial decrease in parameters and model complexity.</p><p>Finally, this thesis proposes a feature fusion technique based on the YOLOv8 architecture to address diverse canopy sizes. The new method incorporates a feature fusion mechanism that includes both cross-scale and same-scale fusion methods, enhancing the model's ability to integrate information across different layers and scales. Large convolution operations are employed to effectively extract key features, helping the model capture richer and deeper global information in the image. Experimental results on the canopy dataset demonstrate that the new method further advances performance, marking a promising solution for accurate and efficient instance segmentation of individual tree crowns in aerial imagery.</p>