Answers are grounded in this paper and cite their sources.
Many large-scale cyberphysical systems, including smart agriculture, smart grid, etc., rely on spatiotemporal sensor data to achieve situational awareness and enable optimal control. Given the scale of agricultural systems, it is common practice to deploy a few sensors across a field and use the measured data to estimate soil parameters at unobserved locations. Such spatiotemporal sensor data imputation problems in the agriculture domain can be solved via kriging that has a high data requirement for good performance. Alternatively, tensor completion methods that use linear factorization techniques or rank minimization work well if the measured data exhibit linear dependencies. For more complex data with non-linear dependencies, neural networks have been explored. However, neural networks often suffer from overfitting due to the sparse nature of spatiotemporal sensor data that is common in the agricultural domain. This work incorporates the expressive nature of graph neural network (GNN) and self-adaptation of a non-parametric Bayesian framework via a Gaussian process (GP) to perform spatiotemporal soil moisture sensor data imputation in agricultural cropland. Results demonstrate that a graph-based Gaussian model outperforms a traditional Gaussian process, classical spatial baselines including inverse distance weighting (IDW), and ordinary kriging in mean absolute error (MAE) and achieves lower root mean squared error (RMSE) across missing-sensor scenarios in heterogeneous crop sites. Evidence from conducted experiments suggests that it is important to preserve GP’s assumption of smoothness to have good imputation performance. thness to have good imputation performance. The proposed GNN-GP framework also provides the uncertainty associated with the predicted values, which is crucial for downstream control and decision-making tasks such as smart irrigation.
No takes yet. Share an insight, caveat, or question.
Rashid et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: