Noise-Tolerant Detection of Cucumber and Grape Leaf Diseases Using Median and Gaussian Filters with Advanced Machine Learning Classifiers

Abstract

The study is a design and development of a strong disease detection system of cucumber and grape leaves with noisy image data, focusing on the ability to withstand salt-and-pepper and Gaussian noises. The image datasets used in agriculture are usually affected by noise because of changes in light, sensor defects, and environmental conditions, which may lead to lower diagnostic accuracy. In order to address this, the proposed system incorporates high noise reduction methods whereby a median filter and a Gaussian filter are used to restore the image quality without compromising on the important leaf texture information. After processing, colour, texture, and shape are used to extract features, which are effective in extracting disease-specific visual representations. These fine features are then trained on various optimized machine learning models, such as Light Gradient Boosted Machine (LGBM), Quantum Support Vector Machine (QSVM), a Modified Random Forest (MRF) with adaptive weighted features, and a Multi-SVM classifier with a custom kernel to map nonlinear features. Through experimental analyses, the proposed ensemble framework is shown to be highly accurate, robust, and noise-tolerant as opposed to the traditional frameworks. The hybrid method is effective in recognizing the significant cucumber diseases and grapes, including powdery mildew, downy mildew, and anthracnose, which will be utilized in the noisy agricultural conditions in the real world. In general, this system offers a noise-resistant, reliable, and computationally efficient system to detect early signs of plant diseases, which can be used in sustainable crop monitoring and precision farming.

Plant phenotyping relevance

植物葉の画像から病徴を推定する画像処理・特徴抽出・機械学習システムの設計開発が中心であり、植物病害状態のフェノタイピング手法に該当する。

Code and data availability

The paper uses two public Kaggle leaf-image datasets (cucumber and grape) as its phenotyping inputs; both are publicly accessible with URLs given in the references. No author code or model checkpoints are reported.