TreeNetAI – a tool for gap-filling time series data

TreeNetAI enables the filling of data gaps in time series with machine learning. The Python-based tool is freely available.

by Mirko Lukovic et al.

Dendrometer and environmental data are sent wirelessly via the LoRaWAN protocol every 10 minutes in TreeNet, without relying on a data logger, thus keeping energy consumption to a minimum. A disruption in the transmission process or defective hardware components lead to data loss and gaps in the time series. Unfortunately, data gaps can never be completely avoided in field measurements. However, it is essential to be able to replace missing data adequately, particularly in dendrometer time series because the annual growth is calculated as the sum over time.

To solve this problem, we have developed an application for gap filling of time series data (deepT). Although the code was developed for the TreeNet dendrometer sensors, it can be used with any time series data (with the respective training of the model). The core of the application is a deep neural network that has been trained on multi-channel time series consisting of stem radius changes, temperature, relative air humidity, vapor pressure deficit (VPD), solar radiation, soil water potential, and total precipitation.

We now provide several pre-trained machine learning models for gap-filling dendrometer data. A first model works with a single incomplete dendrometer time series. A second, more adequate model works with a dendrometer time series, a VPD time series and a soil water potential time series. None of these variables need to be complete. In each case, TreeNetAI provides fully gap-filled time series of all input variables.

All code, including the underlying training and evaluation procedures, is publicly available and can be found in the TreeNetAI repository on GitHub (https://github.com/treenet/treenetai) or via the repository EnviDat.

The work was supported by the Open Research Data Program of the ETH Board (https://open-research-data-portal.ch/projects/ai-module-for-gap-filling-treenet-time-series/). The host institution of the developer  was the Swiss Federal Institute for Forest, Snow and Landscape Research WSL.

Related publication: Mirko Luković, Roman Zweifel et al. (2022), Reconstructing radial stem size changes of trees with machine learning, J. R. Soc. Interface, 1920220349. http://doi.org/10.1098/rsif.2022.0349