Geometric, Photometric, and Semantic Multi-Layer 3D Mapping Towards a Forest Digital Twin

Accurate, rich 3D forest representations are essential for perception, analysis, scene understanding and decision-making in woodland and off-road rugged terrain scenarios. This paper presents an approach for multi-layer 3D mapping for forest environments, integrating geometric, photometric, and semantic representations towards unified spatial models.The geometric layer is constructed using a state-of-the-art factor graph–based SLAM formulation with global georeferencing, enabling consistent large-scale mapping. The photometric layer is obtained through precise multi-sensor registration between LiDAR and RGB imagery, allowing the projection of color information onto the 3D structure. For the semantic layer, a transformer-based deep learning model is employed to perform dense semantic segmentation, enriching the map with comprehensive classification of entities, terrain, and structures.Experimental results in a real-world wildland-urban interface environment highlight the capabilities and limitations of each layer, as well as their combined suitability for off-road rural scene representation. The proposed mapping methods constitute a key step towards interpretable digital twins for forest management.