International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Simulation-Based Data Augmentation for Apple Leaf Disease Using Statistical Moments and HSV Color Features

Authors :

Seyedeh Maryam Moosavi¹ Morteza Gholipour² Yasser Baleghi³

1- Department of Electrical and Computer Engineering Babol Noshirvani University of Technology 2- Department of Electrical and Computer Engineering Babol Noshirvani University of Technology 3- Department of Electrical and Computer Engineering Babol Noshirvani University of Technology

Keywords :

Precision Agriculture،Data Augmentation،Simulation-based Augmentation،Plant Disease

Abstract :

Traditionally, plant disease detection relied on labor-intensive visual inspection. With the rise of precision agriculture, machine learning methods have gained traction for automated detection but require large labeled datasets, which are often difficult to collect due to disease rarity and annotation costs. To overcome this limitation, this study introduces a realistic data augmentation method based on statistical analysis of apple leaf lesions infected by apple scab, using the publicly available PlantVillage dataset. The method reconstructs the color texture of disease lesions by analyzing their statistical and color space properties. Lesion distribution masks are created at various scales to simulate a range of lesion sizes and spatial patterns. Smooth transitions and natural appearances are achieved by gradually blurring lesion boundaries using a Gaussian function. The leaf skeleton is extracted through a median-based approach to place lesions at anatomically plausible locations, resulting in visually realistic synthetic images. These synthetic lesions are then fused with healthy leaf images by pixel-wise multiplication to generate realistic diseased samples. The effectiveness of the proposed approach was evaluated using the SegNet architecture and the F1-score as the evaluation metric. Experimental results show that pretraining the model on synthetic images and fine-tuning with real data improves the F1-score by 25% compared to training solely on real data. This method offers a practical solution to the challenge of data scarcity in plant health monitoring and smart agriculture.