Reconstructing radial stem size changes of trees with machine learning.

Area of Science:

• Ecology
• Data Science
• Machine Learning

Background:

• Ecological studies require continuous data, but field conditions often cause data loss.
• Environmental factors like weather and biological interference corrupt sensor data.
• Missing data hinders accurate ecological analysis and understanding.

Purpose of the Study:

• To apply advanced machine learning techniques for reconstructing multi-channel time-series data.
• To address data gaps in ecological datasets, specifically focusing on tree growth data.
• To evaluate the effectiveness of deep learning models in ecological data restoration.

Main Methods:

• Tested five neural network architectures, including encoders, recurrent neural networks (RNNs), and convolutional neural networks (CNNs).
• Utilized a deep neural network combining Long Short-Term Memory (LSTM) with 1D convolutional layers for optimal performance.
• Applied the best-performing model to reconstruct the TreeNet dataset, which contains tree stem growth and environmental data.

Main Results:

• A deep neural network integrating LSTM and 1D CNNs demonstrated superior performance in data reconstruction.
• The model successfully reconstructed missing segments in the TreeNet dataset.
• The reconstructed data improved performance in a subsequent ecological classification task.

Conclusions:

• Machine learning, particularly deep neural networks, offers a powerful solution for ecological data gap-filling.
• The proposed method enhances the integrity and usability of ecological time-series data.
• Accurate data reconstruction facilitates more reliable ecological modeling and analysis.