Epsilon-Insensitive Primal Loss With Support Vector Machine for Robust Cotton Leaf Disease Classification

Abstract

Cotton diseases significantly reduce both yield and quality of crops, making early and accurate identification of the disease crucial for effective prevention and control. However, achieving a balance between lightweight models and precise classification of cotton diseases remains challenging. This difficulty arises primarily from the dispersed nature of disease symptoms and interference from noisy backgrounds, which complicate the detection process. To address this issue, this research proposes an Epsilon-Insensitive Primal Loss Function-based Support Vector Machine (EI-PLF-SVM) for cotton leaf disease classification. The EI-PLF improves the SVM’s capability to differentiate between the predicted probability and true distribution of classes, resulting in better accuracy. In the Feature extraction process, a Local Binary Pattern (LBP) and Grey-Level Co-Occurrence Matrix (GLCM) are utilized for features extraction to improve the model performance. The proposed EI-PLF-SVM incorporates an ε-insensitive primal loss formulation through dual texture descriptors using LBP and GLCM, facilitating robust discrimination under noisy and visually complex leaf patterns. The experimental results illustrate that the proposed EI-PLF-SVM method reaches a higher accuracy of 99.46% and 92.48% on Kaggle Cotton Disease and Plant Village datasets respectively as compared to existing works such as lesion-aware visual transformer and Custom VGG-16. These results confirm that the EI-PLF-SVM provides an effective and lightweight alternative for practical cotton disease diagnosis.