Designing the Forest of Tomorrow: Generating Virtual Trees with Adversarial Autoencoders

Abstract. This paper explores the generation of “realistic” 3D representations of individual trees to enhance visualizations of forest simulation tool outcomes. By leveraging remote sensing data, we aim to capture individual tree features and characteristics accurately, linking them to dynamic simulations of forest structures and composition. Employing a deep learning approach, we train models on existing 3D scanned data to produce diverse and realistic visual representations of specific tree species. Our method addresses the limitations of existing synthetic tree generation techniques, which often overlook species-specific characteristics. Our approach emphasizes the generation of diverse tree forms, accounting for differences in trunk shape, canopy size, and branching structures. The resulting 3D data offers potential applications for realistic future forest visualizations and improved data augmentation in tree classification models, ultimately contributing to the creation of virtual forests that represent rich species diversity.