Early Detection of Sugar Beet Cercospora Leaf Spot Disease Using Machine Learning-Assisted Thermal Image Processing Method

Abstract Early diagnosis of diseases in agriculture is an important factor in reducing the negative environmental impacts by effectively and economically managing the losses caused by these diseases and reducing the use of chemicals. There are different options within the scope of remote sensing for the early detection of diseases. Among these, choosing a method that can detect diseases accurately without harming the plant and the environment is important. Today, positive developments have been made toward non-invasive and effective detection of diseases with thermal camera-based image processing techniques. In this context, there is potential for disease detection with data collection, image processing, and the determination of the characteristics of disease agents through thermal imaging. The research was based on Cercospora leaf spot ( Cercospora beticola Sacc. ) diseases which have significant economic loss potential in sugar beet. The effectiveness of the proposed method was evaluated in experiments involving Cercospora beticola , utilizing a climate station early warning system and UAV-based thermal images across three subjects and six replicate field trial plots. Analyses were made for the early detection of diseases by comparing thermal images taken from the field with multispectral images taken simultaneously. It was investigated whether it was possible to diagnose the disease early before physical symptoms were seen using image processing and machine learning methods. The variability of leaves was analyzed using field images, thermal images, and machine learning algorithms. Thermal imaging enables the rapid detection of potential disease development by measuring increases in leaf temperature in infrared wavelengths. However, a significant limitation of this method in practice is its sensitivity to climate factors such as air temperature and humidity, which can cause rapid fluctuations in the index. This study compared five machine learning algorithms based on four key metrics. MS imaging achieved about 25% higher accuracy in predicting early disease than TE imaging. This study indicates that thermal imaging provides valuable information but is not as effective as multispectral imaging in detecting early-stage stress factors related to diseases.