Spatially-aware unsupervised anomaly detection for soil stress monitoring using graph autoencoders in precision agriculture

Introduction Early detection of crop stress in precision agriculture is hindered by complex multivariate sensor interactions, spatial variability across fields, and the scarcity of labeled anomaly data. Existing feature-based methods process sensors independently, failing to exploit spatial correlations that characterise localised stress events such as drought. Methods Multivariate soil sensor data (volumetric water content and temperature at five depths) from 42 spatially distributed nodes at Cook Agronomy Farm were analysed. Three unsupervised reconstruction-based models were evaluated: a standard Autoencoder (AE), a Graph Autoencoder (GAE) with k-nearest-neighbour spatial connectivity, and a Graph Attention Autoencoder (GAT-GAE) with adaptive multi-head attention weighting. Detection capability was assessed via synthetic drought injection (Δ = 2.5, normalised units) into 7% of nodes, using Top-K ranking metrics and ROC-AUC analysis across 20 independent runs. Results All three models achieved Hit@10 = 1.00. The GAE yielded a substantially larger anomaly separation margin (8.75) than the AE baseline (2.53), while the GAT-GAE produced the sharpest spatial localisation. AUC values were 0.99 (AE), 0.95 (GAE), and 0.97 (GAT-GAE). Over 20 runs, mean Hit@10 was 0.98 ± 0.07 (Wilcoxon p < 0.0001), confirming statistical robustness. Discussion Graph-based spatial modelling substantially improves anomaly separation and localisation compared to feature-only reconstruction. Reconstruction scores are normalised into a Soil Stress Index (SSI ∈ [0, 1]), providing a threshold-calibrated environmental indicator for data-driven soil health monitoring aligned with SDG 2 (Zero Hunger) and SDG 15 (Life on Land).