Self-Supervised Learning with Human Feedback for tree segmentation based on LIDAR data

Forests in alpine and mountainous regions are facing growing challenges from climate change, making reliable monitoring increasingly important. Although LiDAR technology captures rich 3D data that can support detailed tree-level analysis, training deep learning models for segmentation still typically requires extensive manual labeling. An effort that is both time-consuming and resourceintensive. This work explores a different approach: instead of training a segmentation model from scratch, a pre-trained segmenter is paired with a lightweight classifier trained on the dataset of interest. In this work, a different approach is assessed: instead of training a segmentation model entirely from scratch, a pre-trained segmenter wa paired with a small classifier, trained on the dataset of interest, and whose role is to correct the segementer prediction. The idea with this setup is to automatically build pseudo-ground truth based on the corrected predictions of the segmenter in order to fine-tune it. The interest being that the only hand-labeling needed with this pipeline is on classification samples, which is orders of magnitude faster than segmentation annotation. The pipeline managed to reach an increase of 18% of detected individual trees while reducing by half the number of false positives produced by the Segmenter. Those results are very promising. Not only, did this setup increase the number of correctly identified trees, but it also significantly reduced false positives. Although there is still room for improvement, the method successfully met its objective and hopefully lays a promising foundation for future research in this direction. Contents