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

Continuous monitoring of climatic variables is essential for precision viticulture and data-driven decision support systems. However, agricultural sensor networks are frequently affected by missing data due to hardware failures, communication issues, or maintenance interruptions. In this work, we propose a spatio-temporal graph-based autoencoder for reconstructing missing temperature and relative humidity time series collected from a five-node vineyard sensor network over a two-year period. The model combines a GRU-based temporal encoder, augmented with a time-decay imputation mechanism applied to the input data, with a GraphSAGE spatial module, enabling the joint exploitation of temporal dynamics and inter-node spatial correlations. Experimental results on real-world data show that the proposed approach achieves accurate reconstruction under controlled missing-data scenarios generated through structured artificial masking. For moderate corruption levels (p=0.3), the model attains reconstruction losses of 0.003 for temperature and 0.005 for humidity using short temporal windows (L = 36~3 h), corresponding to MAE values below 0.03 °C and 0.1%, respectively. Even at higher corruption levels (p=0.7), performance remains stable, with losses below 0.008 and 0.011, and MAE values within 0.05 °C and 0.17%. The results highlight a trade-off between temporal context and reconstruction accuracy: shorter windows yield lower absolute errors under moderate corruption whereas, under extreme data loss (p=0.9), the longer windows (L = 144~12 h) reduce the composite temperature reconstruction loss from 0.027 to 0.021. Additionally, temperature is consistently reconstructed more accurately than humidity, reflecting its smoother dynamics and stronger spatial coherence.