Southern Corn Rust (SCR) is a severe threat to maize production in China and worldwide. As an airborne disease, SCR can spread from localized infection to field-wide outbreak within a few days. Achieving incubation period diagnosis and rough prediction can extend the response window for agricultural producers. This study establishes a framework for roughly predicting the Days Before Symptom onset (DBS) of SCR using continuous hyperspectral monitoring under solar illumination. Unlike previous research reliant on controlled environments and artificial inoculation, our approach leverages natural infection dynamics in Sanya, China, shifting the result from binary classification (healthy or infected) to coarse temporal estimation of DBS. We evaluated various feature extraction methods, including traditional disease-related spectral indices (DIs) and wavelet coefficients (WLCs), coupled with machine learning algorithms (Partial Least Squares Discriminant Analysis, Random Forest) and a one-dimensional convolutional neural network (1D CNN). The results demonstrate that traditional DIs were insufficient for incubation period monitoring, achieving an Overall Accuracy (OA) of only 54.51%. In contrast, the 1D CNN model, utilizing the top 64 wavelet coefficients achieved optimal performance with an OA of 71.62%. Crucially, the model demonstrated robust diagnostic capability up to 7 days prior to visual symptom appearance, achieving an OA of 92.05% with high precision (89.41%) and recall (84.98%) for identifying healthy leaves. A 12-day lead time proved less reliable due to high false-positive rates. By providing a rough prediction of symptom outbreak, this method offers a practical, high-applicability tool for crop management, enabling timely fungicide applications and minimizing unnecessary chemical usage.
Jianmeng et al. (Wed,) studied this question.