Manual to run the GBSeparation algorithm for leaf-wood separation

This is a manual on how to run the GBSeparation algorithm developed by Tian et al. (2022) (doi: 10.1109/TGRS.2022.3218603).

The GBSeparation folder was downloaded in july 2023 from https://zenodo.org/record/6837613. Some slight modifications have been done.

Steps

Get the code

1. Download this Github repo. OR download the original code (rar file on https://zenodo.org/record/6837613) and unzip, and download the GBS_demo.py file from this repo.

Set a conda environment

There are some environment requirements to be able to run the code

• Python 3.7 and above (The code is compatible with Python 3.7);
• Other 3rd party packages (numpy, networkx,sklearn,open3d ...).

Therefore you should make a new conda environment:

1. Open Anaconda Prompt

2. Run

   conda create --name GBSeparation python=3.7

3. Activate this environment

   conda activate GBSeparation

4. Install the required packages (numpy, open3d, networkx, scikit-learn, laspy)

   conda install -c anaconda numpy
   conda install -c open3d-admin open3d
   conda install -c anaconda networkx
   conda install -c anaconda scikit-learn 
   conda install -c conda-forge laspy
   conda install -c conda-forge matplotlib
   pip install pillow==9.0.0
   

Input data requirements

There are also some data requirements to be able to run the code

• The input data should be a single tree 3D point cloud and no large gaps in existence.
• The input tree data should be in .pcd, .txt, .las or .ply format

Run the code

1. Modify the GBS_demo.py file by chaning the input and output paths to your paths:

   # the input files path
   pathin = "F:/EUCFACE/trees/" #change to your input folder 
   # the output files path
   outpath = "F:/EUCFACE/leaf-wood/wood/" #change to your output folder
   

2. Navigate to the folder where the GBSeparation folder is at:

   cd ….

3. Run the GBS_demo script:

   python GBS_demo.py

4. Check the separation results which should appear in your output folder.

Tips

• Downsampling and filtering (for reflectance) before running GBSeparation can improve the results.

• GBS often fails in extracting the bottom of the trunk completely. It's an option to add a line to the code where you set points below 2 m (or another value) to wood.

• Different settings give different results. A good idea is to use different parameters/reflectance settings and merge the results. In this example we:

  • Only retained the last returns
  • Used six reflectance thresholds, resulting in six filtered point clouds (reflectance > -20 dB, -15 dB, -11dB, -9dB, -7.5 dB and -6dB)
  • Downsampled each filtered point cloud until it contained fewer than 11 million points (computational limit of our infrastructure)
  • Ran GBS separation on each filtered, downsampled point cloud, with two different input settings: split_interval=[0.1, 0.2, 0.3, 0.5, 1], max_angle=0.25π; and split_interval=[0.5, 1, 1.5, 2, 3], max_angle=0.15 π. The first are the default settings, while the authors recommended the second to use on larger trees. We obtained, for each point cloud, the wood-separated point cloud.
  • merged all the point clouds, and did an extra 'downsampling' to remove double points

  In this figure, each colour is a different GBS run:

Test file

In the in folder on this repo you can find a test file to run the code. The out folder also contains the leaf-wood separated results for this file so you have an idea of the results.