Multi-Modal Spatio-Temporal Graph Neural Network with Mixture of Experts for Soil Organic Carbon Prediction
Researchers have developed SpTGNN, a novel multi-modal spatio-temporal graph neural network designed for predicting soil organic carbon (SOC). This model addresses limitations in existing methods by integrating spectral and temporal data through a heterogeneous graph structure and a fine-tuned TerraMind encoder. SpTGNN utilizes a Mixture-of-Experts module for feature fusion and incorporates advanced uncertainty quantification techniques, outperforming traditional XGBoost baselines in evaluations across Africa and Europe. AI
IMPACT This new framework integrates foundation-model feature extraction and advanced graph attention for improved soil organic carbon prediction, potentially aiding agricultural sustainability and land use planning.