Multi-crop disease classification using a Dilateformer-MobileNetV2 dual-branch fusion network

Multi-crop disease classification requires discrimination among visually similar symptoms. This study evaluates late feature fusion of Dilateformer and MobileNetV2 for 45-class image classification. The two pooled representations are concatenated before classification. The original experiment achieved 97.21% Top-1 accuracy and 96.63% macro F1 on a 59,905-image benchmark. During revision, PlantDoc-derived samples were added to the existing collection to broaden image diversity. On this expanded 68,682-sample dataset, the selected fusion checkpoint achieved 95.45% accuracy and 93.63% macro F1 on 6,550 test samples, compared with 95.48% accuracy for MobileNetV2. These results belong to distinct data and implementation versions. Dataset provenance, parameter accounting and branch-specific gradient-weighted class activation mapping (Grad-CAM) support interpretation of the revision-stage results. The findings demonstrate useful multi-crop classification performance, while showing that simple feature concatenation does not guarantee improvement over a single branch. Independent field generalization remains to be established.