Answers are grounded in this paper and cite their sources.
Banana diseases cause significant crop losses worldwide and threaten food security in many tropical and subtropical regions. Accurate disease diagnosis remains challenging under real-world field conditions characterized by cluttered backgrounds, conditions far removed from laboratory-controlled environments. This paper presents Social-AgriNet, an explainable multimodal deep learning framework for robust banana disease detection from field-acquired imagery. The proposed architecture integrates a Vision Transformer (ViT-Base/16) backbone, an EfficientNet-B2 auxiliary branch, an image-derived metadata encoder, and a gated multimodal fusion mechanism. Evaluated on a geographically independent Healthy-only test set from Tanzania (5,628 images), Social-AgriNet achieves 94.42% ± 2.12% binary healthy recognition accuracy under the default training configuration and 96.58% ± 0.59% under the optimal configuration (omitting label smoothing). Because the Tanzanian test set contains no diseased samples, multi-class disease detection is reported separately on the Asian validation split (n = 936, all nine classes represented), where the model achieves 92.74% accuracy and macro-F1 = 0.9304. Multi-class field generalization to geographically distinct regions remains unproven. Ablation studies confirm that MixUp augmentation (+5.68 pp) and class-weighted sampling (+3.70 pp) are the critical regularization components, while label smoothing is detrimental on this dataset. Fusion gate analysis reveals a continuous modality-weighting mechanism, and Grad-CAM visualizations confirm alignment with agronomically meaningful lesion regions.
No takes yet. Share an insight, caveat, or question.
Moïse Kambale Kasambya (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: