MS-DGCNN++

MS-DGCNN++: Multi-Scale Dynamic Graph Convolution with Scale-Dependent Normalization for Robust LiDAR Tree Species Classification

Resource	Link
Preprint	arXiv:2507.12602 · PDF
Unified 3D library (large benchmarks & preprocessing)	github.com/said-ohamouddou/LIDARLearn

---

What this repository is for

MS-DGCNN++ is a hierarchical multi-scale dynamic graph CNN with scale-dependent edge encoding: raw displacement features at the local scale (small neighborhoods, low SNR for directions) and hybrid raw-plus-normalized directional features at an intermediate scale (larger neighborhoods, high SNR). A theoretical noise-sensitivity argument motivates why normalized directions have mean squared error decaying as O(1/s²) with neighbor spacing s, while raw displacements do not—supporting asymmetric encoding across scales.

This repo holds the structured ablation suite, robustness evaluations, training utilities, and tables/figures pipelines used in Section “Ablation studies and robustness” of sn-article.tex.

The large-scale state-of-the-art comparisons on STPCTLS (seven species, terrestrial laser scanning) and HeliALS (nine species, airborne laser scanning, geometry-only), including 56 models on STPCTLS as reported in the paper, were run using our open-source library LIDARLearn, which integrates MS-DGCNN++ with shared preprocessing, configs, and training scripts. For dataset preprocessing, multispectral / ALS workflows, and replication of those benchmark tables, start from LIDARLearn (preprocessing/, datasets/, scripts/, docs/).

Paper experiments ↔ scripts

§ Experiment	Topic	Entry point
Per-scale encoding	Raw vs hybrid vs asymmetric normalization	core_experiments/experiment1_ablation.py
Density dropout	Canopy thinning / retention	core_experiments/experiment2_density_dropout.py
Noise sweep	SNR crossover and degradation	core_experiments/experiment3_noise_sweep.py
Max-pooling provenance	Which neighbors win max pooling	core_experiments/experiment4_maxpool_provenance.py
Isotropy / effective rank	Feature-space geometry	core_experiments/experiment5_isotropy.py
Component / fusion / k-scale	Broader ablations	core_experiments/run_ablations.py, run_kscale_ablation.py

Batch runners: run_core_experiments.sh, run_robustness.sh (noise, outliers, dropout, n-points, few-shot—after training clean checkpoints).

Repository layout

Path	Purpose
models/	msdgcnn2.py (MS-DGCNN++), msdgcnn.py, optional pointm2ae/ baseline
core_experiments/	Experiments 1–5 and aggregated ablation drivers
robustness_eval/	Checkpoint training + perturbation protocols
utils/	CV splits, dataloaders, training loop, plots, LaTeX helpers
data/STPCTLC/	Local layout for STPCTLS HDF5 and CV split JSON
results/	Example outputs (JSON, TeX, figures)

The directory name STPCTLC follows this codebase; the dataset is referred to as STPCTLS in sn-article.tex.

Requirements

• Python 3.10+ recommended
• NVIDIA GPU recommended (training falls back to CPU if CUDA is unavailable)

pip install -r requirements.txt

Install PyTorch for your stack from pytorch.org (CUDA wheel vs CPU).

Point-M2AE baseline

The models/pointm2ae/ copy here is adapted for these experiments. For the official Point-M2AE release (paper code, weights, full dependencies), see github.com/zrrskywalker/point-m2ae.

MS-DGCNN++ core paths use pure PyTorch k-NN graph features. Enabling the Point-M2AE baseline in this repo may still require CUDA extensions (pointnet2_ops, knn_cuda, etc.).

Data (STPCTLS)

Place the HDF5 bundle expected by the loaders as:

data/STPCTLC/point_cloud_data.h5

On first run, stratified CV splits are created or reused, e.g. data/STPCTLC/cv_splits_k5_seed42.json.

Path tip: core_experiments/*.py joins relative --data_path with the core_experiments/ directory. From the repo root use an absolute path:

python core_experiments/experiment1_ablation.py --data_path "$(pwd)/data/STPCTLC"

Running experiments

From the repository root:

Core experiments

chmod +x run_core_experiments.sh
./run_core_experiments.sh           # experiments 1–5
./run_core_experiments.sh 1       # single experiment
./run_core_experiments.sh ablations
./run_core_experiments.sh kscale

Robustness

chmod +x run_robustness.sh
./run_robustness.sh train
./run_robustness.sh dropout
./run_robustness.sh noise
./run_robustness.sh outlier
./run_robustness.sh npoints
./run_robustness.sh few_shot
./run_robustness.sh all

Use --help on individual Python scripts for epochs, batch size, folds, and output dirs.

Outputs

Artifacts land under results/ or --output_dir: configs, metrics JSON, LaTeX tables, and figures aligned with the paper’s supplementary material workflow.

Citation

Please cite the arXiv preprint. When you use the unified benchmarking stack, cite LIDARLearn as indicated there.

@misc{ohamouddou2025msdgcnnpp,
  title         = {{MS-DGCNN++}: Multi-Scale Dynamic Graph Convolution with Scale-Dependent Normalization for Robust {LiDAR} Tree Species Classification},
  author        = {Ohamouddou, Said and El Afia, Hanaa and Boulaich, Mohamed Hamza and El Afia, Abdellatif and Chiheb, Raddouane},
  year          = {2025},
  eprint        = {2507.12602},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2507.12602}
}