Estimating Foliar Bacterial Spot Severity in a Peach Orchard from Hyperspectral Imagery with Laboratory-Grounded Neurosymbolic Learning

Bacterial spot, caused by Xanthomonas arboricola pv. pruni, is among the most damaging diseases of peach and other stone fruit. The disease severity is evaluated through visual severity ratings, which are time-consuming, subjective, and limited to coarse categorical classes. Hyperspectral imaging (HSI) offers an objective alternative to visual assessment; however, previous studies have focused primarily on detached leaves or potted plants under controlled lighting. In orchard canopies, tree-level ratings encompass thousands of spectral measurements, leaving individual sampling points without lesion-level annotations. To address this, a neurosymbolic model, SpotNeSy, was developed to predict tree-level foliar bacterial spot severity from ground-based HSI of peach orchards. Spectral sampling points fused from neighboring frames were represented as a spatial graph, in which message passing refined the embedding of each point with its local canopy context. A tissue head, supervised by fuzzy rules calibrated on a laboratory library of healthy and lesioned tissue, estimated the lesion probability of each point, which weighted an attention-based multiple instance learning (MIL) aggregation of the tree. Temporal attention then combined the imaging dates of each tree before a regression head predicted severity, while rule-gated pseudo-labels from unrated guard trees extended supervision. The model was trained and tested on two seasons of a replicated spray-program trial, with cross-season validation in both directions. SpotNeSy achieved a mean absolute error (MAE) of 0.13 and 0.11 rating units and a coefficient of determination (R²) of 0.89 and 0.94 in the two directions. The performance was compared with a spectral-spatial Mamba network (MHSI), a 3D convolutional neural network (3D-CNN), a transformer-based model (TBM), random forest (RF), and partial least squares discriminant analysis (PLSDA), whose MAE ranged from 0.26 to 0.68 in one direction and from 0.21 to 0.61 in the other. SpotNeSy also matched the rating exactly on 91% of observations and reproduced the rated ranking of spray programs with a Kendall tau of 0.81 and 1.00. These objective and repeatable severity estimates could support the evaluation of spray programs in peach orchards and complement visual scouting.