MS-DGCNN++: Multi-Scale Dynamic Graph Convolution with Scale-Dependent Normalization for Robust LiDAR Tree Species Classification
Not on the map. Why?
Abstract (arXiv)
Graph-based deep learning on LiDAR point clouds encodes geometry through edge features, yet standard implementations use the same encoding at every scale. In tree species classification, where point density varies by orders of magnitude between trunk and canopy, this is particularly limiting. We prove it is suboptimal: normalized directional features have mean squared error decaying as O(1/s2) with inter-point distance~s, while raw displacement error is constant, implying each encoding suits a different signal-to-noise ratio (SNR) regime. We propose MS-DGCNN++, a multi-scale dynamic graph convolutional network with scale-dependent edge encoding: raw vectors at the local scale (low SNR) and hybrid raw-plus-normalized vectors at the intermediate scale (high SNR). Five ablations validate this design: encoding ablation confirms +4--6% overall accuracy (OA) gain; density dropout shows the flattest degradation under canopy thinning; a noise sweep locates the theoretical crossover near SNR2 ≈ 1.22; max-pooling provenance reveals far neighbors win 85% of competitions under raw encoding, a bias eliminated by normalization; and isotropy analysis shows normalization nearly doubles effective rank. On STPCTLS (seven species, terrestrial laser scanning), MS-DGCNN++ achieves the highest OA (92.91%) among 56 models, surpassing self-supervised methods with 7--24× more parameters using only 1.81M parameters. On HeliALS (nine species, airborne laser scanning, geometry-only), it achieves 73.66% OA with the best balanced accuracy (50.28%), matching FGI-PointTransformer which uses 4× more points. Robustness analysis across five perturbation types reveals complementary variant strengths for deployment in heterogeneous forest environments. Code: https://github.com/said-ohamouddou/MS-DGCNN2.
Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Builds on 42 of its 58 references that are in the map. First posted on arXiv 16 Jul 2025.
Builds on
42 of 58 references.
References · 42
- 2016PointNet: Deep Learning on Point Sets for 3D Classification and Segmentationcited by 5,248
- 2017PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Spacecited by 4,401
- 2015ShapeNet: An Information-Rich 3D Model Repositorycited by 2,897
- 2018Dynamic Graph CNN for Learning on Point Cloudscited by 2,223
- 2020Point Transformercited by 1,470
- 2018PointCNN: Convolution On X-Transformed Pointscited by 764
- 2020PCT: Point cloud transformercited by 732
- 2018PointConv: Deep Convolutional Networks on 3D Point Cloudscited by 657
- 2021Point-BERT: Pre-training 3D Point Cloud Transformers with Masked Point Modelingcited by 653
- 2023Point Transformer V3: Simpler, Faster, Strongercited by 589
- 2019RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Cloudscited by 585
- 2019DeepGCNs: Can GCNs Go as Deep as CNNs?cited by 527
30 more
- 2020PointContrast: Unsupervised Pre-training for 3D Point Cloud Understandingcited by 504
- 2022Masked Autoencoders for Point Cloud Self-supervised Learningcited by 459
- 2022PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategiescited by 436
- 2022Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Frameworkcited by 379
- 2019Relation-Shape Convolutional Neural Network for Point Cloud Analysiscited by 346
- 2022Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-trainingcited by 264
- 2018SO-Net: Self-Organizing Network for Point Cloud Analysiscited by 210
- 2022CrossPoint: Self-Supervised Cross-Modal Contrastive Learning for 3D Point Cloud Understandingcited by 198
- 2018PPFNet: Global Context Aware Local Features for Robust 3D Point Matchingcited by 181
- 2021Walk in the Cloud: Learning Curves for Point Clouds Shape Analysiscited by 172
- 2024PointMamba: A Simple State Space Model for Point Cloud Analysiscited by 167
- 2023Contrast with Reconstruct: Contrastive 3D Representation Learning Guided by Generative Pretrainingcited by 164
- 2022Autoencoders as Cross-Modal Teachers: Can Pretrained 2D Image Transformers Help 3D Representation Learning?cited by 122
- 2023PointGPT: Auto-regressively Generative Pre-training from Point Cloudscited by 115
- 2022Surface Representation for Point Cloudscited by 84
- 2022P2P: Tuning Pre-trained Image Models for Point Cloud Analysis with Point-to-Pixel Promptingcited by 74
- 2020Learning Geometry-Disentangled Representation for Complementary Understanding of 3D Object Point Cloudcited by 71
- 2023Instance-aware Dynamic Prompt Tuning for Pre-trained Point Cloud Modelscited by 68
- 2024Dynamic Adapter Meets Prompt Tuning: Parameter-Efficient Transfer Learning for Point Cloud Analysiscited by 53
- 2021PVT: Point-Voxel Transformer for Point Cloud Learningcited by 43
- 2024Parameter-Efficient Fine-Tuning in Spectral Domain for Point Cloud Learningcited by 43
- 2024PCP-MAE: Learning to Predict Centers for Point Masked Autoencoderscited by 24
- 2024Parameter-efficient Prompt Learning for 3D Point Cloud Understandingcited by 16
- 2024Positional Prompt Tuning for Efficient 3D Representation Learningcited by 15
- 2024PointNet with KAN versus PointNet with MLP for 3D Classification and Segmentation of Point Setscited by 15
- 2023Decoupled Local Aggregation for Point Cloud Learningcited by 12
- 2024Benchmarking tree species classification from proximally-sensed laser scanning data: introducing the FOR-species20K datasetcited by 12
- 2022Points to Patches: Enabling the Use of Self-Attention for 3D Shape Recognitioncited by 9
- 2022PointSCNet: Point Cloud Structure and Correlation Learning Based on Space Filling Curve-Guided Samplingcited by 4
- 2025Introducing the Short-Time Fourier Kolmogorov Arnold Network: A Dynamic Graph CNN Approach for Tree Species Classification in 3D Point Cloudscited by 4
Built on by
Citing papers per year: 2026: 1.
1 paper.