Spatio-Temporal Graph Autoencoder for Sensor Data Reconstruction in Vineyard Microclimate Monitoring

Area of Science:

• Agricultural Science
• Data Science
• Sensor Networks

Background:

• Precision viticulture relies on continuous climatic data for decision support.
• Agricultural sensor networks often suffer from missing data due to technical issues.

Purpose of the Study:

• To develop a spatio-temporal graph-based autoencoder for reconstructing missing temperature and humidity time series.
• To address data gaps in vineyard sensor networks for improved data reliability.

Main Methods:

• A GRU-based temporal encoder with time-decay imputation was combined with a GraphSAGE spatial module.
• The model jointly utilized temporal dynamics and inter-node spatial correlations from a two-year vineyard dataset.

Main Results:

• Accurate reconstruction of temperature and humidity data was achieved under various missing-data scenarios.
• For moderate data loss (p=0.3), MAE values were below 0.03 °C for temperature and 0.1% for humidity.
• Performance remained stable even at higher corruption levels (p=0.7), with temperature reconstructed more accurately than humidity.

Conclusions:

• The proposed spatio-temporal graph autoencoder effectively reconstructs missing climatic data in vineyard sensor networks.
• The study demonstrates a trade-off between temporal window size and reconstruction accuracy based on data loss levels.
• The findings support enhanced data-driven decision-making in precision viticulture.