VegPoint 🌱💧

Postharvest Moisture Prediction using LiDAR Point Clouds: Integrating 3D Structure and Spectral Intensity

(coming soon)

This repository contains the code and dataset processing pipeline for our paper:

From 3D Points to Drying Insight: Hierarchical Deep Learning on LiDAR-Based Geometry and Spectra
Majharulislam Babor, Arman Arefi, Barbara Sturm, Valentin Vierhub-Lorenz, Marina M.-C. Höhne, Manuela Zude-Sasse
Submitted, 2026

---

🌟 Overview

Moisture content is a critical quality parameter in postharvest storage. Traditional spectral sensors capture only 2D information and miss the 3D geometric heterogeneity of produce.
VegPoint is the first annotated 3D LiDAR dataset for broccoli and mushroom, providing:

• 3D point clouds with intensity at 1320 nm, 1450 nm, and a dual‑wavelength moisture index (MI).
• Per‑segment moisture labels for core and periphery regions (high/medium/low classes).
• A complete pipeline for preprocessing, augmentation, and deep learning with PointNet++.

Our work demonstrates that combining geometry and intensity improves moisture classification, especially under spatial transfer (core→periphery) and when one modality is occluded.

---

📂 Dataset

Each sample is a point cloud of 5,000 points (after resizing) with the following attributes:

Field	Description
x, y, z	3D coordinates (normalized per segment)
intensity	Reflectance at the given LiDAR wavelength (1320 nm, 1450 nm or moisture index MI)
region	Core or periphery (derived from XY distance)
moisture	Class: 0 (low, <35%), 1 (medium, 35‑88%), 2 (high, >88%)

The dataset is available at Zenodo:

🔗 10.5281/zenodo.19001042

After downloading, place the contents in data/cloud/ following this structure:

data/
├── cloud/
│   ├── mushroom/
│   │   ├── sample1/
│   │   │   ├── a.csv # x,y,z and intensity at wavelength 1320 nm
│   │   │   ├── b.csv # x,y,z and intensity at wavelength 1450 nm
│   │   │   └── c.csv # x,y,z and intensity at moisture index (MI)
│   │   ├── sample2/
│   │   └── ...
│   └── broccoli/
│       ├── sample1/
│       └── ...
└── mc.csv # ground-truth moisture per sample & region

The file mc.csv must contain at least the columns: sample_id, moisture_core, moisture_periphery.

---

⚙️ 🌱 💧 Installation and Moisture Prediction Experiments

1. Clone the repository

git clone https://github.com/majharB/vegPoint.git
cd vegPoint

2. Set up a virtual environment (conda or venv)

conda create -n vegpoint python=3.11
conda activate vegpoint
# or
python -m venv vegpoint

3. Install dependencies

pip install -r requirements.txt

If you use a GPU, ensure PyTorch with CUDA is installed separately (see pytorch.org).

🚀 Usage All experiments are organized as scripts in the scripts/ directory. The core modules are in src/ (package name pcdg).

Data preparation The dataset is automatically loaded and preprocessed by src/pcdata.py when you run any training script. No manual preprocessing is needed.

4. Training a model

4a. standard random split

Train a PointNet++ model on a broccoli/ mushrooms and wavelength intensity with a simple train/val/test split:

source vegpoint/bin/activate
python scripts/train.py \
    --veg mushroom \
    --channel a \
    --split_method random \
    --aug \
    --epochs 30 \
    --cuda 0

Options:

--veg: mushroom, broccoli, or combine

--channel: a, b, or c (intensity option)

--split_method: random, core_periphery (train on core, test on periphery), periphery_core (reverse)

--aug: enable data augmentation (rotation, scaling, jittering)

--semi_supervised: (for spatial transfer) include a few labeled samples from the test domain in training (no data from target region was used in paper)

--model: geo (only geometry) or geo_int (geometry + intensity)

--epochs, --cuda, etc.

4b. Nested cross‑validation

To perform 10‑fold nested CV with hyperparameter tuning (as used in the paper):

python scripts/cross_val.py \
    --veg mushroom \
    --channel a \
    --aug \
    --model geo_int \
    --cuda 0 \
    --n_splits 10 \
    --epochs 30

The script saves per‑fold results (test indices, best config, losses, predictions) in results/folds/mushroom/. The inner loop trains each configuration for 10 epochs; the best configuration is then evaluated on the test set.

4c. Spatial transfer experiments

Train on one region and evaluate on the other (core→periphery or periphery→core) as reported in the paper:

python scripts/train_spatial_transfer.py \
    --veg broccoli \
    --channel b \
    --split_method core_periphery \
    --aug \
    --semi_supervised \
    --model geo_int \
    --epochs 30 \
    --cuda 0

📝 Citation

BibTeX

@article{babor_2026_vegpoint,
  title={From 3D Points to Drying Insight: Hierarchical Deep Learning on LiDAR-Based Geometry and Spectra},
  author={Babor, M. and Arefi, A. and Sturm, B. and Vierhub-Lorenz, V. and Höhne, M. M.-C. and Zude-Sasse, M.},
  journal={Manuscript submitted for publication.},
  year={2026}
}

APA

Babor, M., Arefi, A., Sturm, B., Vierhub-Lorenz, V., Höhne, M. M.-C., & Zude-Sasse, M. (2026). From 3D points to drying insight: Hierarchical deep learning on LiDAR-based geometry and spectra. Manuscript submitted for publication.