An Edge AI and Mobile Sensing Framework for Real-Time Multi-Crop Disease Detection and Geospatial Surveillance in Smallholder Agricultural Systems

Abstract

Abstract Maize and cassava are staple crops in Nigeria, but their productivity is limited by viral and fungal diseases. This study created a mobile system based on lightweight CNN for smartphone portable real-time detection of cassava and maize diseases. In the 2025 planting season, 10,800 leaf images were gathered from a 10-acre experimental farm in Araromi Area, Bakatari Farm, Ido Local Government, Ibadan that cut across the seven classes of healthy cassava (1,620, 15%), cassava mosaic disease (1,540, 14.3%), cassava brown streak disease (1,410, 13.1%), healthy maize (1,880, 17.4%), maize leaf blight (1,540, 14.3%), maize rust (1,360, 12.6%) and maize streak virus (1,450, 13.4%). In order to make the dataset more diverse, data augmentation was done in the form of rotation by +/-30°, flipping, brightness by +/-20% and random cropping in the range of 80-100%. Lightweight CNN architectures MobileNetV2, EfficientNet-Lite, ShuffleNet, and custom CNN were trained in an 80:20 ratio for train and test. Out of 11 models tested, EfficientNet-Lite model forecast the highest where it achieved an accuracy of 94.6%, precision of 0.95, recall of 0.94, F1-score of 0.94, and an ROC-AUC of 0.97. As for MobileNetV2, it achieved an accuracy of 93.8% while ShuffleNet was estimated to achieve the fastest mobile inference at 65 ms. As for the class-wise analysis, it can be seen that the maize leaf blight (95.2%) and cassava mosaic disease (94.1%) had the most accurate predictions. The offline prediction from mobile deployment showed that EfficientNet-Lite occupied 92 ms and 135 MB. The findings show that low-cost and practical smartphone-based disease diagnosis requires the use of lightweight CNN models that can deliver accuracy and alertness that can allow farmers to manage the crop and food security on their own. Smallholder farmers in a resource-poor rural environment will be able to benefit from these models.

Plant phenotyping relevance

植物葉画像から病害状態を推定するCNNベースのモバイル画像解析手法を開発・比較・実装しており、植物状態の取得と技術性能評価が研究の中心である。

Code and data availability

The paper describes a paper-specific leaf image dataset (10,800 cassava/maize images with GPS-referenced sampling) and trained lightweight CNN models, but no public deposit, repository, or authors' URL is provided. The Data Availability Statement says the datasets are available only from the corresponding author on合理e,

No evidence-backed public reproduction asset is currently recorded.