LiDAR Remote Sensing Meets Weak Supervision: Concepts, Methods, and Perspectives

TL;DR

This paper integrates LiDAR data interpretation and LiDAR-based inversion within a unified weakly supervised learning (WSL) framework, motivated by high labeling costs and pervasive domain shifts. It surveys how incomplete, inexact, and inaccurate supervision, along with domain adaptation and generalization, can be harnessed through techniques like pseudo-labeling, consistency regularization, self-training, and label refinement to enable robust 3D perception and large-scale parameter retrieval. The authors detail LiDAR-specific challenges—irregular geometry, sparsity, and cross-domain heterogeneity—and discuss how sparse LiDAR signals can guide joint learning with multi-source remote sensing data for canopy height, building height, biomass, and depth mapping. They also outline future directions, including leveraging foundation models, open-world adaptation, and richer multimodal datasets to advance scalable, annotation-efficient LiDAR remote sensing at regional to global scales.

Abstract

Light detection and ranging (LiDAR) remote sensing encompasses two major directions: data interpretation and parameter inversion. However, both directions rely heavily on costly and labor-intensive labeled data and field measurements, which constrains their scalability and spatiotemporal adaptability. Weakly Supervised Learning (WSL) provides a unified framework to address these limitations. This paper departs from the traditional view that treats interpretation and inversion as separate tasks and offers a systematic review of recent advances in LiDAR remote sensing from a unified WSL perspective. We cover typical WSL settings including incomplete supervision(e.g., sparse point labels), inexact supervision (e.g., scene-level tags), inaccurate supervision (e.g., noisy labels), and cross-domain supervision (e.g., domain adaptation/generalization) and corresponding techniques such as pseudo-labeling, consistency regularization, self-training, and label refinement, which collectively enable robust learning from limited and weak annotations.We further analyze LiDAR-specific challenges (e.g., irregular geometry, data sparsity, domain heterogeneity) that require tailored weak supervision, and examine how sparse LiDAR observations can guide joint learning with other remote-sensing data for continuous surface-parameter retrieval. Finally, we highlight future directions where WSL acts as a bridge between LiDAR and foundation models to leverage large-scale multimodal datasets and reduce labeling costs, while also enabling broader WSL-driven advances in generalization, open-world adaptation, and scalable LiDAR remote sensing.

Figures (17)

• Figure 1: An overview of LiDAR remote sensing meets weak supervision. LiDAR remote sensing data are primarily obtained from satellite, airborne, vehicle-mounted, and ground-based platforms, providing rich surface information. The interpretation tasks for LiDAR data include denoising, filtering, registration, semantic segmentation, instance segmentation, and 3D reconstruction. Inversion tasks involve estimating forest height, building height, biomass, leaf area index, sea ice thickness, and snow depth. These tasks face challenges such as high annotation costs, domain shift, difficulties in field validation, and sparse supervisory signals. To address these challenges, weakly supervised learning techniques such as incomplete, inaccurate, inexact supervision, domain generalization, and domain adaptation can be employed. Additionally, methods like consistency regularization, pseudo-labeling, self-training, active learning, and feature alignment can effectively improve model generalization. These images are obtained from Wang2024IntroductionTLduan2023denoisingKOLLE2021100001HAN2024500f15071257BESSO2024113843rs1616298391506221063626610679232qiao2024frameworkGUENTHER2024114293.
• Figure 2: A taxonomy of methods and applications for LiDAR remote sensing meets weak supervision. Weakly supervised approaches for LiDAR remote sensing include incomplete supervision, inexact supervision, inaccurate supervision, domain adaptation and domain generalization. These approaches jointly support point cloud interpretation and LiDAR-based inversion tasks.
• Figure 3: Incomplete supervised learning framework for point clouds.
• Figure 4: Bounding box annotations (left) and point-wise semantic labels (right) from the H3D dataset KOLLE2021100001.
• Figure 5: Illustration of scribble-labeled LiDAR point cloud scenes (left) and corresponding overlaid frames (right) from the DALES dataset 9150622.