Detecting Septoria Leaf-Spot Disease in Tomato Plants Using Shifted Window Transformer Model.

Abstract

Tomato (Solanum lycopersicum) cultivation plays a critical role in ensuring food security and economic stability in many parts of the world. However, productivity is often hindered by fungal infections, notably Septoria Leaf Spot, caused by Septoria lycopersici. This disease significantly reduces crop yields by damaging the foliage and accelerating premature defoliation. Timely detection and diagnosis are essential for effective intervention. In this study, we present a deep learning-based approach for automated identification of Septoria Leaf Spot using the Shifted Window Transformer (Swin Transformer), a hierarchical vision transformer architecture known for its balance of computational efficiency and high accuracy in image classification tasks.
The research followed a structured quantitative methodology encompassing model design, data acquisition, training, and performance evaluation. A publicly available dataset sourced from Kaggle, comprising annotated images of healthy and infected tomato leaves, was used for model development. Preprocessing steps included image resizing, normalization, denoising, and data augmentation techniques such as flipping, brightness adjustment, and rotation.
The Swin Transformer model achieved an accuracy of 93.18%, a precision score of 0.92, and an AUC-ROC of 90.91% on the test set, outperforming conventional CNN models like ResNet-50 and VGG16. These results validate the model’s strong generalization capability and its potential use in smart agricultural applications. The study also emphasizes the architectural advantages of the Swin Transformer in extracting both local and global features critical to plant disease identification.
Nonetheless, the exclusive use of a Kaggle dataset introduces limitations, particularly in representing real-world variability. The paper recommends additional validation using field-captured images under diverse conditions to improve robustness. The findings highlight the promise of attention-based models for early and accurate plant disease detection, potentially contributing to increased agricultural productivity and sustainable farming practices.