Transfer Learning for Black Pepper Disease Detection Using Entropy Based Segmentation
Shahana I.L. *
Department of Computer Science, Kerala Agricultural University, CoA Padannakkad, Kerala-671314, India.
Hima V.M.
Department of Plant Pathology, Kerala Agricultural University, CTI Mannuthy, Vellanikkara, Kerala-680656, India.
Manju Mary Paul
Department of Agricultural Statistics, Kerala Agricultural University, CoA Vellayani, Kerala-695522, India.
Anjana Krishna B J
Department of Agricultural Statistics, Kerala Agricultural University, CoA Vellanikkara Kerala- 680656, India.
*Author to whom correspondence should be addressed.
Abstract
Black pepper encounters significant production challenges due to major foliar diseases such as foot rot, pollu, and mottle virus. Pre-trained CNNs has been demonstrated to be effective for disease classification, but the accuracy of these models is suboptimal for black pepper disease classification due to similarity in symptoms. Furthermore, Preliminary interpretability analysis using LIME reveals that even the better-performing models do not focus on key symptomatic regions and were often influenced by irrelevant factors such as shadows and background regions, indicating their unreliability and a lack of true symptom recognition. This indicates the necessity of symptom-based segmentation prior to model training. The objective of this research is to improve the performance and reliability of pre-trained CNNs for black pepper disease classification through entropy-based symptom-aware segmentation. A dataset of 2,881 leaf images comprising five classes, Healthy, Foot rot, Pollu (initial), Pollu (advanced), and Viral disease, was collected from major pepper-growing regions of Kerala. Then entropy-based segmentation was done on the collected dataset to effectively segment disease-specific regions. Subsequently, the performance of the 15 state-of-the-art CNN architectures were evaluated on the segmented dataset. Experimental results revealed that, the use of the segmented dataset improved the accuracy of all 15 pre-trained CNN models, and LIME results indicated that the predictions were based on the relevant symptoms. Moreover, entropy-based segmentation enabled the lightweight architectures like MobileNetV2 to achieve comparable performance to deeper and more computationally intensive models. Specifically, DenseNet201, VGG19, and MobileNetV2 outperformed on the segmented dataset with accuracy above 99%. These results suggest that, combining CNNs with symptom-based segmentation can create a reliable and practical tool for monitoring black pepper diseases in real-world conditions.
Keywords: Machine Learning (ML), Deep Learning (DL), LIME (Local Interpretable Model-Agnostic Explanations), Convolutional Neural Network (CNN), Global Average Pooling (GAP), Histogram of Oriented Gradients (HOG), Gray-Level Co-Occurrence Matrix (GLCM)