Potato Research Agronomy and Crop Science

PotatoCapDS2026 and DAF-BLIP: A Real-World Multimodal Dataset and Disease-Aware Vision–Language Framework for Potato Tuber Health-Condition Captioning

Ganvir, Arvind, Thakre, Pankajkumar, Chakole, Jagdish
Springer Nature · 2026-10-23 · DOI 10.1007/s11540-026-10169-x

Abstract

Potato tuber diseases can adversely affect post-harvest quality, storage potential, and market value, creating a need for informative automated assessment. However, existing agricultural disease datasets are primarily designed for classification or detection and provide limited multimodal supervision for disease-oriented image captioning. We introduce two complementary contributions: (i) PotatoCapDS2026, a real-world multimodal dataset of 8016 potato tuber images covering nine health-condition classes, with structured symptom and severity attributes and five template-guided natural-language reference captions per image; and (ii) Disease-Aware Fusion BLIP (DAF-BLIP), a vision-language framework extending Bootstrapping Language–Image Pre-training (BLIP) by incorporating health-condition and attribute-aware representations with visual semantic features through a gated fusion mechanism for caption generation. DAF-BLIP is evaluated against convolutional neural network–long short-term memory (CNN+LSTM), Vision Transformer–long short-term memory (ViT+LSTM), BLIP-1, and BLIP-2 using BLEU, METEOR, ROUGE-L, CIDEr, and BERTScore. Across five repeated 80:10:10 holdout evaluations, DAF-BLIP achieved BLEU-4 of 0.7553, METEOR of 0.6993, ROUGE-L of 0.7417, CIDEr of 1.1801, and BERTScore-F1 of 0.9524. Statistical analysis identified significant differences in the evaluated performance measures, while expert evaluation provided additional assessment of caption quality and semantic relevance. The dataset and framework provide a resource for research on potato tuber health-condition image captioning and disease-oriented agricultural vision-language analysis.

Read at publisher →