This toolbox uses machine learning methods to analyse and process time series data. A multi-channel time series is expected as input and a modified version of the original time series will be provided as output.
TreeNet (treenet.info) is an international monitoring and research network in which automated tree stem radius fluctuations measured with point dendrometers are analyzed in terms of forest ecosystem responses to climate change. A continuous stream of microclimate and tree physiology data provides realtime information on tree water relations and tree growth.
TreeNet aims to link research results from carbon flux sites with dendrometer data to entire landscapes. Further it provides online-tools to its partners to automatically interpret stem radius fluctuations in terms of tree water deficit, wood growth and related indicators for forest ecosystem carbon sink and drought stress.
The project was initiated by Roman Zweifel, WSL and Werner Eugster, ETHZ in 2009 and is supported by an increasing number of partners.
The main components of the software are the following:
models/: Contains the architecture of each model.raw_data_elaboration/: Scripts for downloading, processing, and converting raw data.utils/: Utility scripts for loading models and other helper functions.scripts/: Bash scripts for downloading raw data, creating TFRecords, training and evaluating the model, and gap-filling data.config.py: Configuration file for hyperparameters and other settings.training.py: Main script for training the neural network.evaluation.py: Script for evaluating the trained model.
In more detail, the relevan directory strucutre looks like the following:
TreeNetAI/
|-- models/
|-- templates/
|-- trained/
|-- raw_data_elaboration/
|-- tools/
|-- config_server.yml (not included)
|-- data_organisation.py
|-- data_statistics.py
|-- server_tools.py
|-- data_processing_library.py
|-- feature_descriptions.py
|-- get_raw_data.py
|-- tfrecords_make.py
|-- tfrecords_load.py
|-- transform_raw_data.py
|-- scripts/
|-- 0_get_data.sh
|-- 1_convert_raw_data.sh
|-- 2_make_tfrecords.sh
|-- 3_train_model.sh
|-- 4_evaluate_model.sh
|-- 5_fill_gaps.sh
|-- utils/
|-- evaluation_tools.py
|-- get_model.py
|-- logger.py
|-- utils.py
|-- config.py
|-- evaluation.py
|-- training.py
For the moment it is only possible to clone the repository to a PC and use the toolbox as a python library. Make sure you have Python installed on your computer and the Python packages listed in the file requirements.txt. You can install all the packages with the following command pip install -r requirements.txt. It is best to first create a virtual environment dedicated to this code and then install the packages inside it. For details on how to create and use a virtual environment see here.
In principle, only the shell script files in the scripts folder should be modified for normal usage. Explanations and details related to the parameters used are explained within these scripts. There are two parts to the software in the repository. The first part is the creation (including training and verification) of the model and the second part is the application of the model for gap-filling.
Data preparation for model input: 1_make_TFrecords.sh
The first step of the tool work-flow is to convert the format of the time series data so that it is compatible with the input requirements of the machine learning model and store it in the tensorflow TFrecords format. There is a script that (load_and_convert_raw_data.py) that downloads the raw data and corresponding metadata directly from a PostgreSQL database. The raw data is processed so that it can be used as input for the machine learning models. In particular, there is a script (tfrecord_make.py) that combines the downloaded data and metadata into the tfrecords format, so that the neural network training input and corresponding labels are always together in the tfrecords data structure.
Model training: 2_train_model.sh
The second part of the project is the training phase. The main script for this phase is the training.py file. It is combined with the utils.py script to load the tfrecords (with the help of the tfrecords_load.py script), to load the desired model (with the help of the get_model.py script) and to train the neural network. The architecture of each model is stored in a separate python script inside the "models" directory. The hyperparameters and other parameters for the training of the model are stored in the config.py file. There is a bash script that is used to call the tfrecord_make.py script that initiates the creation of the tfrecords. There is a second bash script that calls the training.py function together with the variables from the config.py file.
Model evaluation: 3_evaluate_model.sh
There is a third bash script that calls the evaluate.py file to evaluate the model.
Gap-filling: 4_fill_gaps.sh
Contact Mirko Lukovic for help.
So far, the toolbox contains only a gap-filling module. More modules are planned for the near future.
We are open and happy to collaborate with others on this project. If you are intrested, please contact Mirko Lukovic.
The lead authors of the TreeNetAI project are Mirko Lukovic and Roman Zwifel. The idea behind the project is published in the following manuscirpt: Reconstructing radial stem size changes of trees with machine learning. The initial development and deployment of the code was made possible by the Open Research Data Program of the ETH Board (link). The host institution of the developers was the Siwss Federal Institute for Forest, Snow and Landscape Research, WSL.
If you use this software in your work, please cite it and the accompanying scientific article.
Software: You can find the citation details in the top right corner of this repository, under the "About" section ("Cite this repository") or copy the text below.\
@software{Lukovic_TreeNetAI_2025,
author = {Luković, Mirko and Zweifel, Roman},
month = March,
title = {{TreeNetAI}},
url = {https://github.com/treenet/treenetai},
version = {1.0.0},
year = {2025},
doi = {https://www.doi.org/10.16904/envidat.446}
}
Article: Mirko Luković, Roman Zweifel et al. (2022), Reconstructing radial stem size changes of trees with machine learning, J. R. Soc. Interface, 1920220349. Citation: BibTex or ris or endnote.
Link to article.
@article{doi:10.1098/rsif.2022.0349,
author = {Luković, Mirko and Zweifel, Roman and Thiry, Guillaume and Zhang, Ce and Schubert, Mark },
title = {Reconstructing radial stem size changes of trees with machine learning},
journal = {Journal of The Royal Society Interface},
volume = {19},
number = {194},
pages = {20220349},
year = {2022},
doi = {10.1098/rsif.2022.0349}
}
All of the software in this repository is licensed under the GNU General Public License v3.0 (GPLv3). Permissions of this strong copyleft license are conditioned on making available complete source code of licensed works and modifications, which include larger works using a licensed work, under the same license. Copyright and license notices must be preserved. Contributors provide an express grant of patent rights.
Active development