Fractal-Based Image Analysis for Multi-Stage Detection of Tomato Late Blight Using a Laboratory Image Dataset of Greenhouse-Grown Tomato Plants
Abstract
This paper considers the problem of early detection of late blight (Phytophthora infestans) in tomatoes based on computer vision and machine learning methods. The main purpose of the study was to develop a representative dataset of images of tomato leaves and an approach to extracting informative features for classifying the stages of disease development. A new dataset was generated using tomato plants grown under greenhouse conditions, with leaf images subsequently captured under controlled laboratory conditions, including five stages of late blight progression with variability in imaging devices, lighting conditions, and temporal disease dynamics. To improve the quality of image analysis, a preprocessing stage was applied, including conversion to grayscale, median filtering, and binarization using the Otsu method. In addition to the traditional textural features, fractal analysis was used to quantify the structural complexity of the affected leaf areas. To verify the information content of the selected features, classification experiments were conducted using Random Forest, XGBoost, and Support Vector Machine models, and the quality was evaluated using accuracy, precision, recall, and F1-score metrics. The results showed that the combination of textural and fractal features contributes to a more accurate distinction between the stages of disease. The developed dataset and the proposed approach can be used in further research on plant disease diagnosis, agricultural monitoring, and precision farming systems although it should be acknowledged that the dataset is limited to greenhouse settings, and field-scale generalizability requires further validation.