To download article.

DOI: 10.15507/2658-4123.036.202601.097-113

UDK 632.3:635.64

A Hybrid Explainable Deep Learning and Spectral-Texture Ensemble Approach for Tomato Leaf Disease Diagnosis

Pratik Buchke
Research Scholar, VIT Bhopal University (Kotrikalan, Sehore, Madhya Pradesh, India), ORCID: https://orcid.org/0000-0002-7575-5083,

Arul Vinayakam Rajasimman Mayuri
Senior Assistant Professor, VIT Bhopal University (Kotrikalan, Sehore, Madhya Pradesh, India), ORCID: https://orcid.org/0000-0002-4010-4600,

Abstract
Introduction. Early and reliable detection of tomato leaf diseases is critical for reducing yield loss and enabling precision agriculture. Recent advances in deep learning have improved classification performance; however, challenges remain in interpretability and robustness under real-field conditions.
Aim of the Study. This study aims to develop an accurate and explainable hybrid framework that integrates handcrafted spectral–texture descriptors with deep convolutional features to achieve high-performance multi-class classification of tomato leaf diseases across ten categories.
Materials and Methods. A three-stage pipeline is proposed. Spectral features including Excess Green (ExG), Excess Red (ExR), HSV color channels, and vegetation indices are extracted from RGB images to simulate multispectral responses. Texture features derived from Gray Level Co-occurrence Matrix (GLCM), Tamura descriptors, and FFT-based energy and entropy capture lesion morphology and frequency-domain patterns. These features are classified using a Random Forest model. In parallel, an EfficientNetB0-based CNN is fine-tuned on augmented images to learn deep spatial representations. Model interpretability is achieved using SHAP for feature-level analysis and Grad-CAM for visual localization. A late-fusion ensemble strategy integrates both models.
Results. The handcrafted feature-based Random Forest model achieves a baseline classification accuracy of 89.2%, while the fine-tuned EfficientNetB0 CNN attains 94% accuracy. The ensemble framework further improves overall performance to 96%, demonstrating enhanced robustness and generalization across all ten disease classes.
Discussion and Conclusion. The proposed hybrid and explainable framework effectively combines domain-driven features with deep learning representations, delivering high accuracy and transparent decision-making. Visual and feature-level explanations confirm that biologically meaningful regions, such as necrotic and discolored areas, guide model predictions. This approach provides a scalable and reliable solution for automated tomato disease diagnosis, supporting real-world deployment in smart farming and precision agriculture systems.

Keywords: tomato leaf disease detection, spectral and texture features, efficientnet, shap explainability, ensemble learning

Conflict of interest: The authors declare that there is no conflict of interest.

For citation: Buchke P., Mayuri A.V.R. A Hybrid Explainable Deep Learning and Spectral-Texture Ensemble Approach for Tomato Leaf Disease Diagnosis. Engineering Technologies and Systems. 2026;36(1):97–113. https://doi.org/10.15507/2658-4123.036.202601.097-113

Authors contribution:
P. Buchke – development or design of methodology; creation of models, software, verification, whether as a part of the activity or separate, of the overall replication / reproducibility of results / experiments and other research outputs, application of statistical, mathematical, computational, or other formal techniques to analyse or synthesize study data, investigation, resources, data curation, writing original draft preparation, writing review and editing.
A.V.R. Mayuri – development or design of methodology; creation of models, verification, whether as a part of the activity or separate, of the overall replication / reproducibility of results / experiments and other research outputs, application of statistical, mathematical, computational, or other formal techniques to analyze or synthesize study data, conducting a research and investigation process, specifically performing the experiments, or data/evidence collection, data curation, writing review and editing.

All authors have read and approved the final manuscript.

Submitted 22.07.2025;
revised 03.10.2025;
accepted 09.10.2025

REFERENCES

This work is licensed under a Creative Commons Attribution 4.0 License.