Towards automated monitoring of tropical forest ecosystems through the largest trees.

The monitoring of individuals and forest plots in Central Africa is a complex task. Establishing experimental monitoring sites and conducting inventories requires a significant amount of time, effort, and resources. One solution to reduce the effort is to summarize a forest plot by focusing on its largest trees. These trees play a crucial role in the structure, dynamics, and carbon cycle of forests. The development of remote sensing methods and deep learning enables the automatic detection and segmentation of tree crowns. We also propose an innovative method using Detectree2 for tree detection and Segment Anything Model from Meta for crown contour segmentation. The objective of this research is to compare the results obtained from commonly used detection and segmentation methods, as well as this new algorithm, in the case study of the largest trees in the Luki landscape (DRC). Validation is carried out by comparing the results with a dataset of manually segmented 500 individuals, based on on-site observations, and compared to very high-resolution ortho-images. The results aim to demonstrate an improvement in the quality of tree crown segmentation based on RGB sensors compared to LiDAR, while also considering variations in acquisition conditions. This provides new perspectives for forest monitoring.