Open Access

ARTICLE

An ROI-Guided Optimized Machine Learning Framework for Orange Disease Recognition with Feature Selection and Explainability

Israt Jahan Munny¹, Anup Majumder², Bibhas Roy Chowdhury Piyas^3,*, Fahmid Al Farid^4,5,*, Md. Rafsan Jani², Fatama Jannat Tisha³, Israt Jahan³, Abu Saleh Musa Miah⁶, Hezerul Abdul Karim^4,*

1 Department of Computer Science and Engineering, City University, Dhaka, Bangladesh
2 Department of Computer Science and Engineering, Jahangirnagar University, Dhaka, Bangladesh
3 Department of Software Engineering, Daffodil International University, Dhaka, Bangladesh
4 Centre for Image and Vision Computing (CIVC), COE for Artificial Intelligence, Faculty of Artificial Intelligence and Engineering (FAIE), Multimedia University, Cyberjaya, Malaysia
5 Faculty of Computer Science and Informatics, Berlin School of Business and Innovation, Karl-Marx-Straße 97–99, Berlin, Germany
6 Department of Computer Science and Engineering, Rajshahi University (RU), Rajshahi, Bangladesh

* Corresponding Authors: Bibhas Roy Chowdhury Piyas. Email: ; Fahmid Al Farid. Email: ; Hezerul Abdul Karim. Email:

Computers, Materials & Continua 2026, 89(1), 65 https://doi.org/10.32604/cmc.2026.083167

Received 01 April 2026; Accepted 26 May 2026; Issue published 13 August 2026

Abstract

Orange is one of the most economically significant citrus crops worldwide, which is essential for the global food distribution network and supports rural livelihoods. However, its high susceptibility to destructive diseases results in substantial yield losses and long-term economic damage. Despite recent advances in smart agriculture, early and precise disease diagnosis remains challenging due to visual resemblance among disease symptoms, high computational cost, and limited model interpretability. To overcome these difficulties, we introduce a novel lightweight and Region of Interest (ROI)-guided explainable machine learning framework to identify orange disease that integrates a strategic feature selection method with Adaptive Step-Controlled Gorilla Troops Optimizer (ASC-GTO). The proposed method starts with Contrast Limited Adaptive Histogram Equalization (CLAHE)-based image enhancement, followed by K-means clustering to accurately segment and separate the diseased part, which is labelled as ROI. To extract discriminative features from the ROI, Gray-Level Co-occurrence Matrix (GLCM)-based texture and color features are first extracted. Least Absolute Shrinkage and Selection Operator (LASSO) is then used for ranking the features and finding the most discriminative features for each class. Finally, the proposed feature selection method integrates the union and intersection of top features identified in the class-wise scenario using LASSO with globally dominant features found by feature ranking to get a compact and discriminative feature subset for better multi-class classification. Model hyperparameter optimization was performed using the proposed ASC-GTO. Experimental findings indicate that the proposed method outperforms existing techniques with an accuracy of 99.57% on the widely adopted orange disease dataset from Kaggle. Furthermore, it significantly reduces computational complexity, with reductions of 22.2%, 34.76%, and 7.5% in training time, model size, and inference time, respectively, compared to models trained on unprocessed raw input images. Model explainability is further analyzed using SHAP and LIME to identify the most influential features contributing to the prediction outcomes. Overall, the proposed method supports early disease intervention, precision agriculture, and sustainable farming.

---

Keywords

---