AgroMM-GSF++: Confidence-Adaptive Lightweight Multimodal Fusion for Plant Disease Recognition with Repeated Multi-Seed Cross-Validation and External Benchmarking

Abstract

Abstract Recent multimodal agricultural research emphasizes large-scale vision-language systems, while lightweight reproducible approaches remain underexplored for constrained deployments. We present AgroMM-GSF++, a compact image-text fusion framework for plant disease recognition. The model combines a small visual backbone and text-prototype branch with confidence-adaptive gating. To strengthen empirical evidence, we run repeated multi-seed stratified cross-validation (three seeds, repeated folds) and include stronger pretrained transfer baselines (EfficientNet-B0 and ViT-B/16 fine-tunes). We further evaluate on an external PlantVillage subset to test cross-dataset robustness under the same protocol family. On beans, transfer baselines lead absolute macro-F1 (EfficientNet-B0-FT: 0.6355±0.0340), while AgroMM-GSF++ reaches 0.5891±0.0552 with much lower latency than transfer models (9.80 vs 205.66 ms/image for EfficientNet-B0-FT). On the external subset, AgroMM-GSF++ improves over fixed lightweight fusion (0.8719 vs 0.8608 macro-F1) while remaining below transfer-heavy baselines. We provide an end-to-end reproducible workspace including scripts, datasets, figures, and camera-ready tables.

Code and data availability

The paper's own code and processed split artifacts are only available on request ('Code used for data processing, model training, and evaluation is available from the corresponding author on reasonable request'). The public datasets cited (iBean/beans, PlantVillage/HF DScomp380) are third-party prior resources, not the

No evidence-backed public reproduction asset is currently recorded.