Dynamic Structure Learning through Graph Neural Network for Forecasting Soil Moisture in Precision Agriculture

Anoushka Vyas, Sambaran BandyopadhyayIJCAI 2022 · DOI

TL;DR

This paper introduces DGLR, a novel dynamic graph neural network (GNN) that learns to forecast soil moisture by jointly modeling spatial and temporal dependencies. Unlike previous methods, DGLR dynamically learns the graph structure representing relationships between locations, making it robust to missing data and adaptable to real-world agricultural challenges.

What Problem Does This Paper Address?

Precision agriculture relies heavily on accurate soil moisture forecasting to optimize irrigation, improve crop yields, and conserve water. Traditional physics-based models are computationally intensive and require extensive data, while conventional machine learning models often fail to capture the complex spatio-temporal correlations inherent in soil moisture across different locations. Existing GNN approaches, while capable of handling graph-structured data, typically assume a static graph structure or require explicit ground truth for graph construction, which is often unavailable or expensive to obtain in real-world scenarios due to sensor costs and communication failures. This paper addresses the challenge of forecasting soil moisture in a dynamic environment with potentially missing data, by developing a method that can learn both the temporal patterns and the evolving spatial relationships between agricultural locations.

Methods

The core of the proposed solution is the Dynamic Graph Learning through Recurrent graph neural network (DGLR). DGLR integrates several key components:

  • Initial Graph Construction: An initial graph is formed based on the geographic proximity of locations, where nodes are connected if their distance is below a certain threshold. This provides a starting point for the GNN.
  • Temporal Graph Neural Network (GNN+GRU): A GNN layer captures spatial dependencies by learning node embeddings based on neighboring nodes' features and their learned importance. These spatial embeddings are then fed into a Gated Recurrent Unit (GRU) to capture temporal dependencies across time steps. This GNN+GRU layer is applied iteratively.
  • Dynamic Graph Structure Learning: Crucially, DGLR doesn't rely on a fixed graph. It learns and updates the graph structure over time. This is achieved by reconstructing an adjacency matrix based on the similarity of learned node embeddings (using dot products) and incorporating regularization terms:
  • Graph Closeness: Encourages the learned graph to stay close to the initial, heuristic-based graph.
  • Feature Smoothness: Promotes connections between locations with similar input features (e.g., weather, NDVI), even if they are geographically distant.
  • Target Smoothness: Encourages connections between locations with similar soil moisture values, leveraging ground truth data where available.
  • Semi-Supervised Learning: DGLR is designed to be semi-supervised, meaning it can learn from both labeled (ground truth soil moisture) and unlabeled (missing soil moisture) data points. This is vital for real-world applications where complete data is rare.
  • Joint Optimization: All components, including GNN/RNN parameters and graph structure learning, are trained end-to-end using an optimizer like ADAM.

Datasets

The paper utilizes two real-world soil moisture datasets:

  • Spain: Features include NDVI, SAR backscattering coefficients, and weather parameters (temperature, humidity, precipitation). Soil moisture data is from the REMEDHUS network.
  • USA: Features include soil temperature and weather data. Soil moisture data is from the SCAN network.

The authors note that these datasets were curated from multiple sources and that publicly available code and datasets are provided.

Key Insights and Ideas

  • Dynamic Graph Learning is Crucial: The ability to learn and adapt the graph structure over time, rather than relying on a static or pre-defined graph, significantly improves soil moisture forecasting, especially in dynamic environments.
  • Semi-Supervised Learning is a Game-Changer: DGLR's robustness to missing soil moisture data is a major advantage. By leveraging unlabeled data, it can maintain performance even when ground truth is scarce, making it more practical for real-world deployment.
  • Feature and Target Smoothness Regularizers are Key: These regularizers effectively guide the graph learning process, ensuring that the learned connections reflect meaningful relationships based on input features and target soil moisture values, leading to better predictions.
  • Location-Specific GRU Parameters: Not sharing GRU parameters across different locations allows the model to capture unique temporal trends in soil moisture for each site, boosting performance.

Key Results

  • DGLR consistently outperforms various baseline algorithms (SVR, ARIMA, RNN, DCRNN, EvolveGCN) on both the Spain and USA datasets across metrics like RMSE, SMAPE, and Correlation Coefficient.
  • The performance of DGLR drops very slowly even with up to 30% of missing soil moisture values in the training set, demonstrating its robustness. Its performance with 20% missing values was still better than most baselines with no missing data.
  • DGLR shows insensitivity to the initial graph construction threshold, indicating that its learned graph structure is robust and adapts well to the data.
  • The feature and target smoothness losses decrease significantly during training, indicating that the model successfully learns meaningful graph structures.

Limitations and Future Work

  • Computational Complexity: The paper acknowledges that the computation of GNNs and GRUs, along with graph updates, can be computationally intensive. They suggest potential optimizations like restricting graph connections for distant nodes.
  • Generalizability: While DGLR shows promise for soil moisture, the authors express interest in applying it to other spatio-temporal data types with different domain-specific challenges.
  • Sustainability Impact: Future work aims to explore the impact of DGLR on broader sustainability aspects, such as improving crop production, quality, and natural resource preservation.

Summary updated

BibTeX
@inproceedings{vyas2022dynamic,
  author = {Anoushka Vyas and Sambaran Bandyopadhyay},
  title = {{Dynamic Structure Learning through Graph Neural Network for Forecasting Soil Moisture in Precision Agriculture}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/720},
}
Figure