Palm Leaf Disease Detection and Classification Using Deep Learning with Stable Diffusion-Based Data Augmentation

Abstract

In this study, we are addressing the important issue of leaf disease detection and classification for date palm plants, which is a major crop for the economy of arid and semi-arid countries. The height of the trees and the commonality of disease symptoms often hamper the use of standard inspection methods, resulting in delayed or wrong identification. To address these problems, we present a deep learning-based architecture using DenseNet201 as a teacher model and EfficientNetV2-S as a student model in a knowledge distillation setting. The initial image data was provided by Kaggle. This data was diversified using Stable Diffusion inpainting. Manual and automatic mask construction approaches were used for accurate data augmentation. The full workflow includes systematic preprocessing, robust augmentation, and gradual fine-tuning. Experimental results reveal that the suggested system is quite accurate for the classification of four major types.