Explainable Deep Learning-Based Potato Leaf Disease Detection and Severity Assessment for Smart Agriculture in Bangladesh

Abstract

Abstract In Bangladesh, the potato (Solanum tuberosum L.) stands as an indispensable food and cash crop, deeply intertwined with national food security, rural livelihoods, and the broader agricultural economy. However, foliar diseases such as early blight and late blight frequently precipitate substantial yield losses and quality degradation when not identified and mitigated during the nascent stages of infection. Contemporary diagnostic paradigms remain predominantly manual and visual, relying heavily on agricultural professionals, which is often inefficient and inaccessible for remote farmers. While deep learning has demonstrated remarkable efficacy in automated plant disease recognition, existing methodologies frequently lack interpretability, disease severity quantification, and real-world field applicability. This paper introduces a comprehensive, interpretable deep learning-based framework utilizing EfficientNetV2-B0 for classifying potato leaf images into healthy, early blight, and late blight categories. By integrating Gradient-Weighted Class Activation Mapping (Grad-CAM), the model achieves high transparency, highlighting critical prediction regions. Furthermore, a severity assessment module estimates infection percentages, providing actionable treatment recommendations, ultimately enhancing agricultural decision-making.

Plant phenotyping relevance

ジャガイモ葉の画像から病害の種類と感染割合（重症度）を推定する深層学習手法が研究の中心であり、植物病害状態の表現型取得・定量化に該当する。

Code and data availability

The preprint describes a potato leaf disease detection framework using PlantVillage images plus field images from Bangladesh, but provides no public dataset deposit, no author code/model availability statement or URL, and no supplement. PlantVillage is a cited third-party dataset, not a paper-specific asset.

No evidence-backed public reproduction asset is currently recorded.