Ideally, a forest monitoring system should be cost-effective, time-efficient, and accurate. However, existing tools fall short, raising scalability concerns. Traditional forest inventory and terrestrial laser scanning are slow and costly. Airborne laser scanning covers larger areas but remains expensive. Spaceborne missions are scalable but lack accuracy. The rapid development of UAVs and 3D vision algorithms offers a promising alternative. Yet, scalability depends on both learning and data — creating comprehensive, annotated forest datasets is challenging. Synthetic data from virtual environments holds promise, as seen in other domains, but is still largely unexplored in forest ecology.
The goal of my research is to develop a toolkit that creates an augmented digital twin of a real forest to enable scalable forest monitoring algorithms using synthetic data. The research is divided into two stages: 1) Real-to-Sim: Algorithmically constructs a virtual forest as a digital twin from sensed representations of a real forest. 2) Sim-to-Real: Generate synthetic data from augmented digital forests to train 3D perception algorithms for forest property inference.
