This workflow performs tree species classification on cross section images of individual tree point clouds using a YOLOv5 classification model.
Download YOLOv5 and setup a working environment as explained here. Or use yolo calssification scripts from this repository (see "./yolo_classification").
Download Julia from here.
Clone this repository and use the folder ./treeprojection, which contains a Project.toml, a Manifest.toml and the julia script (treeprojection.jl) as Julia environment.
The julia package LasIO, which is used here to load the point clouds, only supports las-format 1.1 - 1.3. If your files are in las-format==1.4 you can use the rscript "transform_las-format_1.2_.r" to transform your las files into las-format 1.2.
Create a tree spiecies oriented folder structure as shown below and store your point clouds into these folders.
├── pointclouds
│ ├── tree_species_1
│ │ ├── tree.las
│ │ ├── tree.las
│ │ ├── ...
│ ├── tree_species_2
│ │ ├── tree.las
│ │ ├── tree.las
│ │ ├── ...
│ ├── ...
│ │ ├── tree.las
│ │ ├── tree.las
│ │ ├── ...
│ ├── tree_species_99
│ │ ├── tree.las
│ │ ├── tree.las
│ │ ├── ...Create a similar tree spiecies oriented folder structure with empty folders for the output images. The treeprojection.jl script expects the shown folder structure.
The Julia script "treeprojection.jl" creates four cross section images (600px x 800px) from different angles (angles: 0, 45, 90, 135) of each passed las-file. A example of a cross section image is presented below.
Go to the folder of your julia installation and run the command below. Note that the script is opted for parallel computing. You can set the number of processors with -p. For example, if you have only one processor set -p 1.
./julia -p 7 -O ./treeprojection/treeprojection.jl ./pathto/pointclouds ./pathto/output
For Training of the YOLOv5 model split the dataset of cross section images into a training and validation dataset. Save the split dataset in a folder structure like this, because YOLOv5 expects it:
├── images
│ ├── train
│ │ ├── tree_species_1
│ │ │ ├── treeid_0.png
│ │ │ ├── treeid_45.png
│ │ │ ├── treeid_90.png
│ │ │ ├── treeid_135.png
│ │ │ ├── ...
│ │ ├── ...
│ ├── val
│ │ ├── tree_species_1
│ │ │ ├── treeid_0.png
│ │ │ ├── treeid_45.png
│ │ │ ├── treeid_90.png
│ │ │ ├── treeid_135.png
│ │ │ ├── ...
│ │ ├── ...
After setting up a YOLOv5 environment activate this environment and navigate to the YOLOv5 classification folder for training YOLOv5/classify/. Run the following code from the terminal
python train.py --model yolov5l-cls.pt --data /pathto/images --epochs 200 --optimizer AdamW
After training apply the trained YOLOv5 classification model to a test data set. Activate the YOLOv5 environment and navigate to the YOLOv5 classification folder YOLOv5/classify/. Use the command below for prediction.
python predict.py -- weight ./runs/train-cls/Your_EXP/weights/best.pt --sources path/to/test_data_images --save-txt
As a result YOLOv5 saves the predicted classes of each cross section image in a result folder as text files. Use these text files for the next step. Since you have four cross section images of a single tree point cloud, you get four classification results from your trained yolo model for each single tree point cloud. In order to get the final tree species prediction for the individual tree point cloud apply the r script "" on the text files created by the application of the trained YOLOv5 classification model on the test dataset (see below).
Open the R-Srcipt "final_predictions.r" and use the path to the text files containing the yolo model predictions from the step above as input path. Additionally, sepcify the output folder and the output file name in the R-Script. Afterwards just run the R-Script. As a result it will save a csv file containing the final class prediction results for each individual point cloud.
Apply the R-Script "" on the text files created by the application of the trained YOLOv5 classification model on the test dataset. The result is a csv file containing the tree id's and the final casts for tree species.
In order to train the YOLO classification model and to validate the performance during training we randomly split the original tre3d training-dataset into a training- (90%) and validation (10%) dataset. We performed this 90% / 10% split for each tree species. Using the above presented workflow our trained YOLOv5 classification model achieved an overall accuracy of 84 % over all 33 tree species classes on the validation dataset.The best performing YOLO classification model was trained after 26 epochs. The class specific accuracies achieved by the application of the trained YOLO model on the validataion dataset and the model summary are depicted here:
The hyperparamters we used for training are uploaded as a yaml file ("opt.yaml").