Accurate, scalable estimation of soil organic carbon (SOC) and clay content from remote sensing data is crucial for monitoring soil health at large continental scales. Standard baseline methods such as PLSR, SVR, Random Forest, Gradient Boosting etc. struggle to exploit the inherent relationship between spectral signatures and spatial neighborhoods of the corresponding samples. In order to overcome this limitation, we propose MVHGCN, a Multi-View Hypergraph Convolutional Network developed for joint SOC and clay content prediction from Sentinel-2 imagery.

The proposed architecture addresses three main drawbacks of standard approaches via the following contributions. First, hyperedges can capture higher-order relationships between soil samples, connecting pixels through both spectral or geographic similarity within a single cluster, instead of decomposing those relationships into pairwise links. Second, a multi-view formulation processes six spectral transforms of each sample (continuum-removal, absorbance, Savitzky-Golay derivatives, SNV-normalized variants), encoded via a 1D-CNN, as parallel views fused through a learned View-Based Attention Fusion module that adaptively weights each transform via an attention mechanism. Third, the model operates on two parallel branches, a global cosine-similarity branch and local multi-scale spatial branch (within-pixel and within-tile), which are fused at every block via a Cross-Talk blending unit, so local geographic proximity and global spectral similarity can be simultaneously exploited. Finally, we introduce Adaptive Hypergraph Learning (AHL), an additional attention mechanism that refines incidence matrices during training, and a multi-task head that predicts SOC and clay jointly under an uncertainty-weighted loss with learnable per-task variance.

We evaluate our approach on data from the LUCAS 2015 Topsoil survey matched to satellite-derived reflectance composites, generated from multi-temporal Sentinel-2 imagery. The proposed model is compared against a series of 1) Standard machine learning methods (PLSR, SVR, Random Forest, Gradient Boosted Trees), a state-of-the-art 1D-CNN developed specifically for estimation of soil-properties from remote sensing data, and several baseline graph and hypergraph networks. The results are reported on the original target scale, after inverting the log1p and z-score transforms. The proposed MVHGCN is consistently found to outperform all standard baselines, the respective graph and hypergraph base models, as well as the state-of-the-art 1D-CNN model.