[Paper Review] Dynamic Structure Learning through Graph Neural Network for Forecasting Soil Moisture in Precision Agriculture

Anoushka Vyas, Sambaran Bandyopadhyay|arXiv (Cornell University)|Dec 7, 2020

Computational Physics and Python Applications4 citations

TL;DR

This paper proposes a dynamic graph neural network (DGLR) that jointly learns temporal graph structures and forecasts soil moisture in precision agriculture, enabling robust performance on real-world data with missing ground truth values. The method achieves state-of-the-art accuracy by integrating feature and target smoothness regularizers to dynamically refine graph connectivity during training.

ABSTRACT

Soil moisture is an important component of precision agriculture as it directly impacts the growth and quality of vegetation. Forecasting soil moisture is essential to schedule the irrigation and optimize the use of water. Physics based soil moisture models need rich features and heavy computation which is not scalable. In recent literature, conventional machine learning models have been applied for this problem. These models are fast and simple, but they often fail to capture the spatio-temporal correlation that soil moisture exhibits over a region. In this work, we propose a novel graph neural network based solution that learns temporal graph structures and forecast soil moisture in an end-to-end framework. Our solution is able to handle the problem of missing ground truth soil moisture which is common in practice. We show the merit of our algorithm on real-world soil moisture data.

Motivation & Objective

• To address the challenge of accurate, scalable soil moisture forecasting in precision agriculture using data-driven methods.
• To model spatio-temporal correlations in soil moisture across regions, which conventional machine learning models often fail to capture.
• To develop a semi-supervised learning framework that can effectively utilize both labeled (with ground truth) and unlabeled (missing ground truth) soil moisture data.
• To learn dynamic graph structures that evolve over time to reflect changing spatial dependencies in soil moisture patterns.
• To improve robustness to sensor and communication failures by training on incomplete datasets with missing soil moisture values.

Proposed method

• The method employs a dynamic graph neural network (DGLR) that iteratively updates graph structures based on node features and predicted soil moisture values.
• It introduces two key regularizers: feature smoothness (FS) to encourage connections between nodes with similar features, and target smoothness (TS) to promote connections between nodes with similar soil moisture values.
• The model uses non-shared GRU parameters across locations to capture location-specific temporal dynamics in soil moisture.
• Graph structure learning is performed end-to-end via optimization of a loss function that combines prediction loss with FS and TS regularizers.
• The framework is trained in a semi-supervised manner, allowing it to leverage both labeled and unlabeled data points during training.
• The initial graph is constructed using a heuristic based on spatial proximity and correlation thresholds, which is then refined during training.

Experimental results

Research questions

• RQ1 Can a graph neural network framework effectively model spatio-temporal dependencies in soil moisture without relying on fixed or pre-defined graph structures?
• RQ2 How does the performance of the proposed dynamic graph learning method compare to static graph-based and sequence-only models in soil moisture forecasting?
• RQ3 To what extent can the model handle missing ground truth soil moisture values in real-world datasets?
• RQ4 How sensitive is the model’s performance to the initial graph structure constructed via heuristic methods?
• RQ5 How do the feature and target smoothness regularizers contribute to the evolution of the graph structure and prediction accuracy?

Key findings

• DGLR achieves state-of-the-art performance on real-world soil moisture datasets, outperforming both static graph-based models and sequence-only models like RNNs.
• The model shows minimal performance degradation even with up to 30% missing soil moisture values in training data, demonstrating strong robustness to data incompleteness.
• Performance remains stable across different initial graph construction thresholds, indicating low sensitivity to heuristic initialization.
• The feature smoothness and target smoothness losses decrease significantly during training and stabilize, confirming that the model learns meaningful, coherent graph structures.
• Ablation studies show that all components of DGLR—non-shared GRUs, graph structure learning, and regularizers—are essential for optimal performance.
• DGLR's performance with 20% missing values surpasses most baselines trained on fully labeled data, highlighting its effectiveness in semi-supervised settings.