Automatic Annotation Of 3D Multispectral LiDAR Data For Land Cover Classification

Ongoing advancements in Earth observation technologies have led to an increasing demand for fine-grained 3D maps, particularly in urban areas rich of diverse objects. Unlike traditional monochromatic LiDAR (ML), modern multispectral LiDAR (MSL) systems simultaneously capture high resolution geometric and spectral data, especially beneficial for accurate 3D urban mapping. At the same time, deep learning (DL) models have shown promising results in urban mapping, despite their need for large amount of labeled data. This study presents a new method based on zero-shot and K-means unsupervised learning to automatically label 3D MSL data. The benefits of MSL's spatial-spectral information and autoannotated training data have been explored by using KPConv point-wise DL model. Achieved results indicate that the proposed auto-annotation pipeline, with an overall accuracy (OA) of ca 85% and a mean Intersection over Union (mIoU) of ca 70%, could ease laborious annotation task and facilitate the development of new unsupervised point-based semantic segmentation algorithms for 3D land cover classification.