Fungal infections: Classification performance and detectability with machine learning models.

Abstract

This study explores the use of AI-driven machine learning algorithms for the early detection of fungal diseases in tomato plants, a method that enhances diagnostic accuracy and enables more effective crop protection. The research was conducted in 2023 in Shandong Province, China, using two tomato cultivars - Dongfeng-1 and Gypsy. Ten greenhouse plots (2,000 m² each) and ten open-field plots (5,000 m² each) were studied, with image data collected via high-resolution cameras, multispectral sensors, and microclimate sensors, yielding approximately 20,000 annotated images. Five machine learning models were tested: convolutional neural networks (CNNs), random forests, gradient boosting, support vector machines (SVMs), and k-nearest neighbors (KNN). CNNs demonstrated superior accuracy in both greenhouse (95.2% ± 0.3) and open-field plots (92.5% ± 0.5), with corresponding AUC-ROC values of 0.96 and 0.93 (p = 0.001). The false positive rate for CNNs was 4.1% in greenhouses and 5.3% in open-field plots, while diagnostic time was shorter in greenhouses (8.3 s vs. 10.5 s). Compared to visual inspection, CNNs significantly improved diagnostic accuracy and reduced fungicide use. To ensure robustness, the models were evaluated under varying lighting and microclimate conditions. Assessments on both GPU and CPU platforms demonstrated the model's feasibility for deployment on edge devices and cloud-based systems.

Plant phenotyping relevance

トマト植物の真菌病を画像・センサー観測から推定する機械学習手法を開発・比較し、精度、頑健性、計算環境で評価しており、病害状態の表現型取得が中心である。

Code and data availability

The paper's ~20,000 annotated tomato fungal-disease images and sensor data are not publicly deposited; the authors state the dataset is available only upon request. No public code, models, or data URLs are provided.

No evidence-backed public reproduction asset is currently recorded.