An improved deep learning plant doctor: a paradigm shift toward Zero Hunger

Abstract

Abstract Malnutrition and food insecurity have remained critical issues faced in Africa. According to the 2023 statistical data in Nigeria, for instance, approximately 87 million out of the country’s 220 million people (39.5 per cent) still live below the poverty line. From this trajectory, it is safe to say that the continent of Africa is not yet on the path to Zero Hunger (SDG 2) by 2030. While several successive government administrations have proposed various intervention programs such as Operation Feed the Nation, the Green Revolution, and Lower Niger River Basin Development Authority, fertiliser support, among others, in Nigeria, little attention has been given to plant diseases, one of the root causes of low productivity among smallholder farmers. Therefore, this research leveraged the inherent characteristics of deep learning models and developed an improved deep learning Plant Doctor based on MobileNetV3-Small architecture with a user-friendly interface that enables the drag and drop of plant images or direct upload. The developed system was tailored towards the computational demands of smallholder farmers’ low computing devices. The developed MobileNetV3-Small architecture uses a smart patch-based scanning process that focuses on the leaf regions, resizes the image to 224 × 224 as input size, and unfreezes 20 layers for feature learning from the patches. The patches are augmented using brightness, contrast, and slight rotation to ease the detection of tiny symptoms. This allows for detailed symptom analysis without overwhelming memory or processing power. The developed, improved MobileNetV3-based plant doctor easily detects plant diseases through their visual symptoms on their leaves, prescribes treatments, and broadcasts detected diseases to farmers within the same region for preventive control measures. The evaluation of the developed MobileNetV3-small showed that the system can detect plant disease with an accuracy of 99.85% on the merged PlantVillageDoc dataset, with the added advantage of broadcasting detected plant diseases to other farmers within the same cluster through their registered email. This system offers a paradigm shift in educating smallholder farmers by providing timely disease detection and expert guidance, thereby reducing crop losses, improving yields, and strengthening national food security.

Plant phenotyping relevance

葉の視覚症状から植物病害を検出する深層学習システムの開発と評価が中心であり、植物の病害状態を直接推定するフェノタイピング手法に該当する。

Code and data availability

The paper describes a modified MobileNetV3-Small plant disease detector trained on a merged PlantVillageDoc dataset built from the public PlantVillage and PlantDoc benchmarks. However, no author-deposited dataset, code, model checkpoint, or supplement with an explicit public availability URL is provided in the supplied

No evidence-backed public reproduction asset is currently recorded.