Rice is a primary staple food for more than half of the world’s population, so even moderate yield losses due to disease can have serious economic and food‑security consequences. Rice plants are vulnerable to a range of leaf and panicle diseases, including bacterial leaf blight, bacterial leaf streak, bacterial panicle blight, blast, brown spot, dead heart, downy mildew, hispa and tungro, all of which can severely reduce photosynthetic efficiency, grain quality and overall yield. Traditionally, diagnosis in farmers’ fields relies on visual inspection by agronomists or extension workers, which is time‑consuming, subjective and difficult to scale across large production areas, especially where expert availability is limited.
Over the past decade, deep learning has emerged as a powerful approach for automated, image‑based plant disease detection. Convolutional neural networks (CNNs) and, more recently, vision transformers have achieved high accuracy on benchmark datasets, particularly when combined with transfer learning from large‑scale corpora such as ImageNet. In rice disease detection, CNN backbones including VGG, ResNet, DenseNet, EfficientNet and MobileNet have demonstrated strong performance, with DenseNet121 reported to achieve around 97.5% test accuracy on the 10,407‑image Paddy Disease Classification dataset comprising nine disease classes and one healthy class. At the same time, hierarchical vision transformers such as Swin have shown competitive or superior results on generic image‑recognition tasks and are increasingly explored in agricultural imaging.
In this context, the Paddy Disease Classification dataset provides a suitable basis for training and evaluating deep learning models for rice leaf disease detection and for benchmarking new architectures against established CNN baselines. This dissertation investigates an ensemble‑based approach that combines DenseNet121, a densely connected CNN, with Swin‑Tiny, a compact hierarchical vision transformer, for automated classification of nine rice leaf disease and healthy categories from RGB images. Both models are fine‑tuned using transfer learning, strong spatial and colour data augmentation, mixed‑precision training and modern optimisation techniques, and their outputs are fused through a simple probability‑averaging ensemble designed to exploit complementary feature representations.
The overarching goal is to design, implement and evaluate a deep‑learning system that achieves high classification accuracy on the Paddy Disease Classification dataset and can serve as the core component of a practical decision‑support tool for precision agriculture and crop‑health monitoring. Within this goal, the main objectives are to: (i) prepare a robust training and validation pipeline for the dataset; (ii) fine‑tune DenseNet121 (Huang et al., 2017) and Swin‑Tiny under comparable training conditions (Liu et al., 2021); (iii) construct and assess a CNN–transformer ensemble; and (iv) critically compare the obtained results with previously reported DenseNet121 performance and discuss implications for real‑world deployment.
