Spectrophotometric and Machine Learning Approaches for Early Detection of Apple Fire Blight

Early detection of fire blight in apple orchards is critical for effective disease management and sustainable crop production. This study explores the use of a portable spectral device combined with machine learning techniques to classify disease status and identify key spectral features associated with biotic stress. Vegetation health was assessed using NDVI values derived from spectral data, which showed a significant decrease from the healthy stage (mean NDVI = 0.7518) to the presymptomatic stage (mean NDVI = 0.7307, p < 0.001). Random Forest classification achieved an overall accuracy of 92% in distinguishing healthy and infected samples, with an area under the receiver operating characteristic curve (AUC-ROC) of 0.94. Models trained within specific treatment groups demonstrated consistent performance, with accuracies ranging from 81% to 90% depending on the algorithm and treatment type. Key spectral features identified by Random Forest models primarily spanned the near-infrared region (770–830 nm), highlighting their importance in capturing physiological changes associated with disease progression. However, unsupervised learning suggested minimal impact of chemical treatments on spectral patterns, with spectral differences largely driven by inoculation status. While temporal analysis using LSTM networks demonstrated potential for disease progression tracking, results were limited due to the lack of laboratory-confirmed infection data. Future work will focus on validating the identified spectral features for fire blight detection, integrating laboratory-confirmed data for enhanced temporal modeling, and exploring regression-based approaches to quantify disease severity. This research establishes a foundation for portable, scalable, and machine learning-driven solutions for early disease detection in precision agriculture.