Title: Label-Efficient 3D Forest Mapping: Self-Supervised and Transfer Learning for Instance Segmentation, Semantic Segmentation, and Species Classification

Abstract

Accurate, detailed structural and species-level data for individual trees are becoming essential for advancing precision forestry, conserving biodiversity, and establishing reference datasets for biomass and carbon mapping. Currently, point clouds generated by airborne and terrestrial laser scanning represent the most viable data source for extracting this information rapidly and at scale. While recent breakthroughs in deep learning have enhanced the ability to segment and classify individual trees and identify their semantic components, these models generally demand extensive annotated training data, thereby hindering further progress. Generating dense, high-quality annotations for 3D point clouds—particularly within complex forest environments—is both labor-intensive and difficult to scale.

To address this bottleneck, this study investigates strategies that minimize reliance on large labeled datasets by leveraging self-supervised and transfer learning techniques. Our primary goal is to enhance performance in three key areas: instance segmentation, semantic segmentation, and tree classification, utilizing realistic and operational training sets. When compared to models trained from scratch, our approach yields improvements across all evaluated tasks. Specifically, for instance segmentation, the combination of self-supervised learning and domain adaptation resulted in a 16.98% increase in AP50. In semantic segmentation, the application of self-supervised learning alone boosted the mean Intersection over Union (mIoU) by 1.79%. Furthermore, for tree classification, hierarchical transfer learning achieved a 6.07% improvement in mean Jaccard index.

To facilitate adoption and simplify implementation, we have consolidated these tasks into a unified framework. This system streamlines the workflow, enabling users to transition seamlessly from raw point cloud data to tree delineation, structural analysis, and species classification. Additionally, the use of pretrained models contributes to environmental sustainability by reducing energy consumption and carbon emissions by approximately 21%. As an open-source resource, this work aims to accelerate the operational extraction of individual tree information from laser scanning data, ultimately supporting advancements in forestry, biodiversity conservation, and carbon mapping.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC