Enhancing tree species classification: insights from YOLOv8 and explainable AI applied to TLS point cloud projections

Abstract Deep learning-based tree species classification models applied to terrestrial laser scanning data achieve high accuracies, yet their classification decisions remain poorly understood. Addressing this gap, we present a novel explanation framework that systematically links salient regions of Finer-CAM saliency maps—which highlight image features contributing to the classification of a target species while suppressing features shared with similar species—to structural tree features in 2D side-view images of terrestrial laser scanning (TLS) point clouds. We applied this framework using YOLOv8-based tree species classification models trained on a subset of the FOR-species20K benchmark data set. Faithfulness analysis confirmed that Finer-CAM reliably identifies the most discriminative image regions for model predictions, validating its suitability for enhancing model interpretability. Analysis of 630 saliency maps indicate that the models primarily rely on image regions associated with tree crowns across most species, particularly for Silver Birch, European Beech, English Oak, and Norway Spruce, while stem regions contribute more strongly to the differentiation of European Ash, Scots Pine, and Douglas-fir. Image perturbation experiments further demonstrated that the visibility of tree structure—such as branching patterns—in the 2D side-view images enhances classification performance, indicating YOLOv8’s capabilities to leverage detailed structural information encoded in high-density TLS point clouds. Using a cross-validation approach our trained models achieved a mean overall accuracy of 96% (SD = 0.24%) on a test data set unseen during training. Beyond assessing the contribution of image regions associated with structural tree features to tree species differentiation, our methodological framework can facilitate a better understanding of model limitations, the detection of data set biases and shortcut learning, and contributes in building confidence in deep learning-based tree species classifiers.