Segmenting Vegetation from bare-Earth in High-relief and Dense Point Clouds using Machine Learning

These programs are modelled after work originally presented by myself at the AGU Fall Meeting in December 2021 ([abstract][https://agu.confex.com/agu/fm21/meetingapp.cgi/Paper/863844] and recording) and supplementary to the manuscript in review in Remote Sensing. They are designed to segment vegetation from bare-Earth points in a dense point cloud, although they may also be used to segment any two classes that are visually distinguishable from each other by colour alone. The programs are meant to reclassify large and dense point clouds very efficiently, similar to the following (green points represent 'vegetation' and brown points represent 'bare-Earth'):

The out-of-the-box direct transferrability of the pre-trained ML models is further demonstrated using a point cloud for Chimney Bluffs, NY (along Lake Ontario) previously published by the USGS (yellow points represent 'vegetation' and blue points represent 'bare-Earth'):

There are two approaches:

1. The non-machine learning approach utilizes a vegetation index and Otsu's thresholding method. This approach is more computationally demanding and the appropriate vegetation index and threshold value are likely to vary by location and application.
2. The machine learning approach utilizes the Tensorflow API. This approach is more efficient, less subjective, and more robust across geographies and applications. Although it is faster with access to a GPU, a CPU can also be used. (Note: The code is not optimized for use with multiple GPUs.)

The full paper describing this approach is currently in review by Remote Sensing: Wernette, Phillipe A. (in reivew) Vegetation Filtering of Coastal Cliff and Bluff Point Clouds with MLP. Submitted to Remote Sensing.

Contents

1. Vegetation Indices Included
2. Installation
3. Non-Machine Learning Approach
4. Machine Learning Approach
   • Command Line Arguments
   • Option A: Two Separate Programs
     • ML_veg_train.py
     • ML_veg_reclass.py
   • Option B: One Single Program
     • ML_vegfilter.py
5. Feedback
6. Citation
7. References

Vegetation Indices Included

These programs compute the following vegetation indicies, their M-statistics, and Otsu threshold values:

• Excess Red (exr)¹
• Excess Green (exg)²
• Excess Green-Red (exgr)³
• Excess Blue (exb)⁴
• Normal Green-Red Difference Index (ngrdi)⁵
• Modified Green Red Vegetation Index (mgrvi)⁵
• Green Leaf Index (gli)⁶
• Red Green Blue Veggetation Index (rgbvi)⁷
• Kawashima Index (ikaw)⁸
• Green Lead Algorithm (gla)⁶

Each of the vegetation indices only requires some combination of the red, green, and blue color bands. No NIR, SWIR, or other band is required.

The following Python modules are required:

os
ntpath
time
copy
subprocess
laspy
lasrs *(required to write compressed LAZ file with laspy)*
numpy
pandas
tKinter
tensorflow (or tensorflow-gpu) **Only if using the machine learning approach**

Installation:

Clone the repository locally first.

git clone https://github.com/pwernette/point_cloud_vegetation_filtering

Then, create a virtual environment from one of the .yml environment files in the environment sub-directory.

To create an environment for the non-machine learning approach (utilizing Otsu's thresholding method and a vegetation index), create the environment using:

conda env create -f PC_veg_filter_env.yml

To create an environment for the machine learning approach (utilizing Tensorflow), create the environment using:

conda env create -f ML_veg_filter_env.yml

Once you have created the virtual environment, activate the environment by either:

conda activate vegfilter

or

conda activate mlvegfilter

USAGE (PROGRAMS WITHOUT MACHINE LEARNING):

Before running the training program or reclassification program, ensure that you have pre-clipped two separate LAS or LAZ point clouds:

1. Point cloud containing points only of class A (e.g. point cloud containing only vegetation points).
2. Point cloud containing points only of class B (e.g. point cloud containing only bare-Earth or non-vegetation points).

These point clouds can be segmented from larger point clouds using any number of programs. I utilize CloudCompare to manually clip out points for each class. It is important that each of the two segmented point clouds specified above include only points of the same class. Including points actually belonging to another class but included in a different sample point cloud will introduce error in the histogram values and will affect the computed Otsu's threshold value for each vegetation index.

First run the training program:

python veg_train.py

Then, run the reclassification program:

python veg_reclass.py

veg_train.py

Inputs:

The program will automatically request the user to select 2 input files:

1. The point cloud containing vegetation points only (for training).
2. The point cloud containing only bare-Earth points (for training).

Outputs:

An output CSV file will be generated with the following naming scheme:

{veg_filename}_{noveg_filename}.csv

Where {veg_filename} is the file name of the point cloud containing vegetation points only and {noveg_filename} is the name of the point cloud containing bare-Earth points only.

The output CSV will have the following attributes (columns) of information:

{vegetation_index_name} {minimum_possible_index_value} {maximum_possible_index_value} {M-statistic} {Otsu_threshold_value}

veg_reclass.py

Inputs:

The program will automatically request the user to select 2 input files.

1. The point cloud to be reclassified.
2. The CSV file containing the vegetation index value ranges, M-statistics, and Otsu threshold values. --> This file is automatically created when you run veg_train.py

Outputs:

A new LAZ file will be generated with the following naming scheme:

{filename}_reclass_{vegetation_index_name}_veg_noveg.laz

Where {filename} is the original point cloud file name and {vegetation_index_name} is the name of the vegetation index determined or selected to differentiate vegetation from bare-Earth using Otsu's thresholding approach.

The output LAZ file will be saved in the same directory as the input file and will contain all the original points with updated classification values corresponding to either vegetation or bare-Earth.

USAGE (FOR MACHINE LEARNING PROGRAMS):

The machine learning approach can be run (1) as two separate programs, one for ML model training and a second for LAS/LAZ file (re)classification, or (2) as a single program that builds and trains a ML model and then uses that model to reclassify a LAS/LAZ file.

Command Line Arguments

Command line options are available to for both the two program and one program options to cut down on pop-up windows and aid in batch scripting:

Argument	Type(s)	Default value(s)	Description/Function	Program
-v, -veg	string	NA	Point cloud containing vegetation points only	ML_veg_train, ML_vegfilter
-g, -ground	string	NA	Point cloud containing ground points only	ML_veg_train, ML_vegfilter
-r, -reclass	string	NA	Point cloud to be reclassified	ML_veg_reclass, ML_vegfilter
-h5, -model	string	NA	h5 Model file	ML_veg_reclass, ML_vegfilter
-m, -name	string	NA	ML model name	ML_veg_train, ML_vegfilter
-vi, -index	string	rgb	Vegetation index or indices to be calculated	ML_veg_train, ML_vegfilter
-mi, -inputs	list-string	r,g,b	Model inputs (will be used in conjuction with -index flag options)	ML_veg_train, ML_vegfilter
-mn, -nodes	list-integer	8,8,8	Number of nodes per model layer (by default specifies the number of layers)	ML_veg_train, ML_vegfilter
-md, -dropout	float	0.2	Probability of model layer dropout (used to avoid overfitting)	ML_veg_train, ML_vegfilter
-mes, -earlystop	list-integer,float	5,0.001	Early stop criteria ([patience],[change_threshold])	ML_veg_train, ML_vegfilter
-te, -epochs	integer	100	Number of training epochs (maximum number)	ML_veg_train, ML_vegfilter
-tb, -batch	integer	100	Batch size	ML_veg_train, ML_veg_reclass, ML_vegfilter
-tc, -cache	boolean	True	Cache batches (improves training time)	ML_veg_train, ML_veg_reclass, ML_vegfilter
-tp, -prefetch	boolean	True	Prefetch batches (significantly improves training time)	ML_veg_train, ML_veg_reclass, ML_vegfilter
-tsh, -shuffle	boolean	True	Shuffle inputs (use only for training to avoid overfitting)	ML_veg_train, ML_veg_reclass, ML_vegfilter
-tsp, -split	float	0.7	Data split for model training (remainder will be used for model validation)	ML_veg_train, ML_vegfilter
-tci, -imbalance	boolean	True	Adjust data inputs for class imbalance (will use lowest number of inputs)	ML_veg_train, ML_vegfilter
-tdr, -reduction	float	0.0	Data reduction as proportion of 1.0 (useful if working with limited computing resources)	ML_veg_train, ML_vegfilter
-thresh, -threshold	float	0.6	Confidence threshold used for reclassification	ML_veg_reclass, ML_vegfilter
-rad, -radius	float	0.10	Radius used to compute geometry metrics (if specified in inputs)	ML_veg_train, ML_veg_reclass, ML_vegfilter

OPTION A: RUN TWO SEPARATE PROGRAMS

If utilizing the two program approach, first build, train, and save the model (line 1). Then, reclassify a LAS/LAZ file using one or more models (line 2):

python ML_veg_train.py
python ML_veg_reclass.py

ML_veg_train.py

The ML_veg_train.py program will read in the two training point clouds, account for any class imbalance, build a ML model, and train the ML model.

Running ML_veg_train.py without any command line argument will automatically enable a simple graphical interface similar to this:

Inputs:

The following inputs are required for the ML_veg_train.py program. If any of these options are not specified in the command line arguments, a pop-up window will appear for each.

1. The point cloud containing vegetation points only
2. The point cloud containing only bare-Earth points
3. The output model name

Outputs:

All outputs will be saved in a directory with the following scheme:

saved_models_{date}

Where {date} is the date the model was created and is pulled from the computer clock. If this directory does not already exist then it will first be created.

The trained model will be written out as a single h5 file as well as a directory. Both the h5 file and the directory will have the same name, as specified by the user.

A plot of the model will also be saved as a PNG file (see example below), and a summary text file will be written that contains the complete model summary and metadata.

ML_veg_reclass.py

The ML_veg_reclass.py program will automatically read in the model structure, weights, and required inputs (including vegetation indices and geometry metrics) and will reclassify the input point cloud.

Running ML_veg_reclass.py without any command line argument will automatically enable a simple graphical interface similar to this:

Inputs:

The following inputs are required for the ML_veg_reclass.py program. If any of these options are not specified in the command line arguments, a pop-up window will appear for each.

1. The point cloud to be reclassified
2. The h5 model file (can be generated using the ML_veg_train.py program)

Outputs:

The reclassified LAS/LAZ file will be saved in the same directory as the original point cloud.

A new LAZ file will be generated in with the following syntax:

{filename}{model_name}{threshold_value}.laz

Where {filename} is the original point cloud file name, {model_name} is the name of the model used to reclassify the input point cloud, and {threshold_value} is the threshold value used to segment vegetation from bare-Earth.

OPTION B: RUN A SINGLE PROGRAM

ML_vegfilter.py

The ML_vegfilter.py program will use the two training point clouds to generate a machine learning model with the user-specified arguments, and then use this model to reclassify the specified point cloud. The significant advantage of using a single program is eliminating the need to read the model file for reclassification.

Running ML_vegfilter.py without any command line argument will automatically enable a simple graphical interface similar to this:

Inputs:

The following inputs are required for the ML_vegfilter.py program. If any of these options are not specified in the command line arguments, a pop-up window will appear for each.

1. The point cloud containing vegetation points only
2. The point cloud containing only bare-Earth points
3. The output model name
4. The point cloud to be reclassified

Outputs:

The model will be saved as a h5 file and a directory, as well as a PNG of the model structure and a detailed metadata summary text file. The model and all it's associated files (graph as PNG and summary metadata file) will be saved in a saved_models_{date} folder, where {date} is the date the model was created.

The reclassified point cloud will be saved in the results_{date} folder as:

{filename}{model_name}{threshold_value}.laz

Where {filename} is the original point cloud file name, {model_name} is the name of the model used to reclassify the input point cloud, and {threshold_value} is the threshold value used to segment vegetation from bare-Earth.

FEEDBACK

If you have any questions about how to implement the code, suggestions for improvements, or feedback, please leave a comment or report the issue with as much detail as possible.

CITATION

Wernette, Phillipe A. 2024. Segmenting Vegetation from bare-Earth in High-relief and Dense Point Clouds using Machine Learning.

@software{Wernette2024,
  author = {{Wernette, Phillipe A.}},
  title = {Segmenting Vegetation from bare-Earth in High-relief and Dense Point Clouds using Machine Learning},
  url = {https://10.5281/zenodo.10966854},
  version = {1.00},
  date = {2024-04-12},
}

OTHER PUBLICATIONS AND INFORMATION

This work follows my original presentation at the 2021 AGU Fall Meeting in New Orleans, LA. My abstract can be found HERE, and my recorded presentation can be found HERE.

My manuscript in Remote Sensing is also based on this original research and is currently available via Preprints.org:

Wernette, P. Machine Learning Vegetation Filtering of Coastal Cliff and Bluff Point Clouds. Preprints 2024, 2024041387. https://doi.org/10.20944/preprints202404.1387.v1

Point clouds for coastal bluffs near the Elwha River mouth near Port Angeles, WA can be found HERE.

Wernette, Phillipe (2024). Coastal bluff point clouds derived from SfM near Elwha River mouth, Washington from 2016-04-18 to 2020-05-08 [Dataset]. Dryad. https://doi.org/10.5061/dryad.8pk0p2nww

REFERENCES

Footnotes

1. Meyer, G.E.; Neto, J.C. 2008. Verification of color vegetation indices for automated crop imaging applications. Comput. Electron. Agric. 63, 282–293. ↩

2. Woebbecke, D.M.; Meyer, G.E.; Von Bargen, K.; Mortensen, D.A. 1995. Color Indices forWeed Identification Under Various Soil, Residue, and Lighting Conditions. Trans. ASAE, 38, 259–269. ↩

3. Mao, W.; Wang, Y.;Wang, Y. 2003. Real-time detection of between-row weeds using machine vision. In Proceedings of the 2003 ASAE Annual Meeting; American Society of Agricultural and Biological Engineers, Las Vegas, NV, USA, 27–30 July 2003. ↩

4. Neto, J.C. 2004. A combined statistical-soft computing approach for classification and mapping weed species in minimum -tillage systems. Ph.D. Thesis, University of Nebraska – Lincoln, Lincoln, NE, USA, August 2004. ↩

5. Tucker, C.J. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 8, 127–150. ↩ ↩²

6. Louhaichi, M.; Borman, M.M.; Johnson, D.E. 2001. Spatially located platform and aerial photography for documentation of grazing impacts on wheat. Geocarto Int. 16, 65–70. ↩ ↩²

7. Bendig, J.; Yu, K.; Aasen, H.; Bolten, A.; Bennertz, S.; Broscheit, J.; Gnyp, M.L.; Bareth, G. 2015. Combining UAV-based plant height from crop surface models, visible, and near infrared vegetation indices for biomass monitoring in barley. Int. J. Appl. Earth Obs. Geoinf. 39, 79–87. ↩

8. Kawashima, S.; Nakatani, M. 1998. An algorithm for estimating chlorophyll content in leaves using a video camera. Ann. Bot. 81, 49–54. ↩