Python package for automatic tree crown delineation in aerial RGB and multispectral imagery based on Mask R-CNN. Pre-trained models can be picked in the model_garden.
A tutorial on how to prepare data, train models and make predictions is available here. For questions, collaboration proposals and requests for data email James Ball. Some example data is available to download here.
Detectree2是一个基于Mask R-CNN的自动树冠检测与分割的Python包。您可以在model_garden中选择预训练模型。这里提供了如何准备数据、训练模型和进行预测的教程。如果有任何问题，合作提案或者需要样例数据，可以邮件联系James Ball。一些示例数据可以在这里下载。
 |
Code developed by James Ball, Seb Hickman, Thomas Koay, Oscar Jiang, Luran Wang, Panagiotis Ioannou, James Hinton and Matthew Archer in the Forest Ecology and Conservation Group at the University of Cambridge. The Forest Ecology and Conservation Group is led by Professor David Coomes and is part of the University of Cambridge Conservation Research Institute. |
|---|
Please cite this article if you use detectree2 in your work:
Ball, J.G.C., Hickman, S.H.M., Jackson, T.D., Koay, X.J., Hirst, J., Jay, W., Archer, M., Aubry-Kientz, M., Vincent, G. and Coomes, D.A. (2023), Accurate delineation of individual tree crowns in tropical forests from aerial RGB imagery using Mask R-CNN. Remote Sens Ecol Conserv. 9(5):641-655. https://doi.org/10.1002/rse2.332
Independent validation has been performed on a temperate deciduous forest in Japan.
Detectree2 (F1 score: 0.57) outperformed DeepForest (F1 score: 0.52)
Detectree2 could estimate tree crown areas accurately, highlighting its potential and robustness for tree detection and delineation
Gan, Y., Wang, Q., and Iio, A. (2023). Tree Crown Detection and Delineation in a Temperate Deciduous Forest from UAV RGB Imagery Using Deep Learning Approaches: Effects of Spatial Resolution and Species Characteristics. Remote Sensing. 15(3):778. https://doi.org/10.3390/rs15030778
- Python 3.8+
- gdal geospatial libraries
- PyTorch ≥ 1.8 and torchvision versions that match
- For training models GPU access (with CUDA) is recommended
e.g.
pip3 install torch torchvision torchaudio
pip install git+https://github.com/PatBall1/detectree2.git
Currently works on Google Colab (Pro version recommended). May struggle on clusters if geospatial libraries are not configured. See Installation Instructions if you are having trouble.
Under development
Detectree2, based on the Detectron2 Mask R-CNN architecture, locates trees in aerial images. It has been designed to delineate trees in challenging dense tropical forests for a range of ecological applications.
This tutorial takes you through the key steps. Example Colab notebooks are also available but are not updated frequently so functions and parameters may need to be adjusted to get things working properly.
The standard workflow includes:
- Tile the orthomosaics and crown data (for training, validation and testing)
- Train (and tune) a model on the training tiles
- Evaluate the model performance by predicting on the test tiles and comparing to manual crowns for the tiles
- Using the trained model to predict the crowns over the entire region of interest
Training crowns are used to teach the network to delineate tree crowns.
Here is an example image of the predictions made by Detectree2.
In development
In development
- Functions for multiple labels vs single "tree" label
├── LICENSE
├── Makefile
├── README.md
├── detectree2
│ ├── data_loading
│ ├── models
│ ├── preprocessing
│ ├── R
│ └── tests
├── docs
│ └── source
├── model_garden
├── notebooks
│ ├── colab
│ ├── colabJB
│ ├── colabJH
│ ├── colabKoay
│ ├── colabPan
│ ├── colabSeb
│ ├── exploratory
│ ├── mask_rcnn
│ │ ├── testing
│ │ └── training
│ ├── reports
│ └── turing
├── report
│ ├── figures
│ └── sections
└── requirements
To automatically format your code, make sure you have black installed (pip install black) and call
black .
from within the project directory.
Copyright (c) 2022, James G. C. Ball
