Tree Species Classification from Very High-Resolution UAV Images Using Deep Learning

Kurzfassung

Very high-resolution red-green-blue (RGB) imagery captured by uncrewed aerial vehicles (UAVs) enables a new perspective on forest mapping, focusing 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 classification per tree due to low resolution and the inability to capture individual tree-level details. Driven by deep learning (DL), tasks like tree species (TS) classification using spatial data, such as the texture of a tree, are possible. This use of very high-resolution imagery for forest mapping has yet to be fully explored. This work uses the convolutional neural network (CNN) U-Net and a transformer model named Fully Convolutional Transformer (FCT) for semantic segmentation with the benchmark dataset "Bamforest" with a two cm resolution and 27,160 annotated trees within 105 hectares. In the summer of 2022, it was collected in Germany with a UAV. Ten classes were chosen to distinguish seven common German TS, coniferous, and deciduous TS, one combined class for all other trees, one for dead trees, and one for background. The objective is to develop a semantic segmentation model that accurately classifies each pixel into distinct classes. An individual tree crown (ITC) loss function is added to the training to improve results by using the ground truth of each tree crown to refine predictions through a voting process. In postprocessing, the semantic segmentation is combined with instance segmentation, which was previously trained separately. This process involves assigning the semantic segmentation results to each correct tree shape and deciding on one species using a voting mechanism, which helps refine the tree shape and species decision. The final results show the importance of data availability for each class with the F1-score and the intersection over union (IoU) and the advantage of using the second ITC loss for better classification with the disadvantage of additional computational time. The results show 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 results support forest management, biodiversity assessment, and environmental monitoring by leveraging UAV and advanced DL techniques. The extensive dataset makes the generalization of models for other similar forest areas possible and creates the ability to generate single-tree information in dense forests.