Cross-crop generalization in multi-species leaf disease recognition

Introduction Plant diseases are major biotic stress factors whose visual symptoms may be conserved across taxonomic families, though the extent remains unquantified. Cross-crop leaf disease recognition requires architectures that generalize to species absent from training, but existing methods lack a theoretically grounded criterion for ensuring cross-species transferability. Methods This study proposes functional-domain separation as such a criterion, derived from the disentangled representation learning hypothesis, and implements it as the leaf-disease composite feature interaction network (LCFIN), a configurable multi-branch feature enhancement framework. LCFIN assigns a two-level Squeeze-and-Excitation mechanism to a channel calibration domain and an MLP-Mixer head to a spatial interaction domain on an EfficientNet-B3 backbone, with dual-weighted sampling addressing cross-level class imbalance. It is trained for main fruit and vegetable leaf disease identification (MFVLDI) on three families (Rosaceae, Vitaceae, Solanaceae) and evaluated under a frozen-representation linear-probe protocol on two unseen crops, maize of Poaceae and potato of Solanaceae. Results LCFIN reaches 88.99% accuracy on maize and 95.23% on potato under matched acquisition conditions. Across five independent group-aware partitions the in-domain accuracy is 99.93% ± 0.05%, with 19 misclassified test images among 27,937. A duplicate and near-duplicate audit bounds the accuracy inflation attributable to data redundancy at a few hundredths of a percentage point. Ablation across the five partitions shows the two domains to be functionally complementary rather than redundant, but the interaction is sub-additive, with a residual of 0.04 percentage points below the additive prediction. Discussion These findings are consistent with the learned features capturing phytopathological regularities, although this observation remains descriptive rather than causal. They support the hypothesis that functional-domain separation induces implicit feature disentanglement and enables capture of phytopathological signatures shared across the tested families, without demonstrating that they transcend taxonomic boundaries. Because the extractor is frozen and only a linear head is fitted, extending the model to a further crop costs far less than training a network from scratch.