Comparing Mask R-CNN and Mask2Former architectures for individual tree crown delineation

Weather anomalies caused by the anthropogenic climate crisis challenge environmental workers such as foresters with an increasing number of responsibilities. In addition to providing attractive conditions for new generations of workers, deep-learning tools can ease processes for better efficiency. Tree instance segmentation has potential for many functionalities that support arborists and foresters by detecting and classifying singular trees. In the past, Mask R-CNN was primarily applied due to its outstanding performance on the 2016 COCO dataset challenge. As an alternative, we suggest Mask2Former, which outperforms Mask R-CNN on the COCO dataset. Additionally, we test whether additional digital canopy height model data can improve training. While the latter is shown to have no, if not a negative impact on the results, Mask2Former indeed outperforms Mask R-CNN in tree instance segmentation by up to 3.8%. Our code is publicly available.2