News

With climate change as threat and staff shortage in forestry, tree detection and vitality classification from UAV-based data is becoming a crucial task in order to preserve and maintain our forests. However it is very challenging, particularly in deciduous and mixed forests. While most studies have relied on multispectral data (RGB + NIR + RE), recent advancements in deep learning models have shown promising results using only RGB data, which would make these kinds of analyses cheaper and more accessible. This study aims to compare the efficacy of these data sources as the foundation for tree crown detection and vitality classification. I trained models on 3-band RGB and 5-band multispectral data collected from forests around Bamberg, Germany. Two instance segmentation algorithms, Mask R-CNN and Mask2Former, were employed for Individual Tree Crown Delineation (ITCD) along with ResNet-50 for vitality classification. At an IoU (Intersection over Union) threshold of 0.5, the best segmentation algorithm (Mask R-CNN on RGB data) achieved a precision of 0.46, recall of 0.31 and thus an F1-score of 0.37 on the test set. The ResNet classifiers based on both types of data demonstrated an overall accuracy exceeding 0.8, with the RGB-trained model achieving the highest accuracy of 0.85. For vitality classification, the RGB model achieved an F1-score of 0.91 for the vital class, 0.79 for the dead class, and 0.48 for the degrading class. The multispectral model showed slightly lower performance with F1-scores of 0.89, 0.77, and 0.51 for the vital, dead, and degrading classes, respectively. Notably, multispectral data increased the recall for degrading trees by 13 percent points, which resulted in a slightly higher F1-score. Incorporating the genus label as additional information for the model significantly enhanced the precision of the dead class by 8% on RGB data and by 13% on multispectral data. This study demonstrates that RGB data is a viable alternative for tree health monitoring and that genus labels contain valuable information for vitality classification. Hence, a multi-label learning approach is proposed to leverage it. The aim of this thesis is to contribute to the overarching goal of an end-to-end application for the segmentation and classification of trees from airborne data, both RGB or multispectral.