Object-Guided Tree Species Classification Using Deep Learning

Kurzfassung

Very high-resolution RGB imagery captured by unmanned aerial vehicles (UAVs) enables a new perspective on forest mapping, allowing more focus on spatial information. For now, it is common to use satellite imagery using spectral information. However, this limits the research to forest areas rather than specific classification at the tree level due to low resolution and the inability to capture individual tree-level details. Driven by deep learning (DL), tasks like tree species classification using spatial data, such as the texture of a tree, are possible. Especially the use of very high-resolution imagery for forest mapping has yet to be fully explored. We propose a region-based deep learning architecture for tree species classification based on the refined individual tree delineation results from our previous work. We test our approach on the "Bamforest" benchmark dataset with a 2 cm resolution and 27,160 annotated trees within 105 hectares. The data was obtained in a natural German forest and collected in the summer with a UAV. We evaluated ten classes to distinguish seven common German tree species, coniferous and deciduous tree species, one combined class for minority tree species, one for dead trees, and one for the background. The objective is to develop a semantic segmentation model that accurately classifies each pixel into distinct classes. In addition to studying the impact of the pixel amount of each class on the classification output, we propose an accuracy loss function to the training to improve results by using the spatial configuration of the ground truth to refine predictions through a voting process. In postprocessing, we combine our semantic segmentation with instance segmentation. This process involves assigning our results to each correct tree shape with a single tree species through a voting mechanism, which helps refine the tree shape and decide on one species. Our results show the importance of data availability for each class with the F1-score and the Intersection over Union (IoU) per class as well as the advantage of using the second accuracy loss function for minority classes with the disadvantage of additional computational time. Our experiments also prove that the voting process in postprocessing has a positive effect on the output by comparing first, a test dataset similar to the training data and second, a different test area. The further application of this framework has the potential to support forest management, biodiversity assessment, and environmental monitoring by leveraging UAV and advanced DL techniques. The extensive and diverse dataset makes the generalization of models for other forest areas possible and creates the ability to generate single-tree information in dense forests.