A novel data-centric AI approach based on sensitivity and correlation analyses for multi-organ plant disease classification

Hammad Saleem, Muhammad; Hammad, Fakhia; Taha, Muhammad; Palaiahnakote, Shivakumara; ur Rehman, Sadaqat; Saraee, Mohamad

Abstract

With advancements in deep learning (DL), most research on classification problems has focused on developing or modifying DL models, known as model-centric artificial intelligence (AI) approaches. However, this approach is time-consuming and overlooks the exploration of the available resources and expertise required to address industrial problems. This study proposes a new data-centric AI-based approach by thoroughly investigating dataset complexities, using multi-organ plant disease classification as a case study. To the best of our knowledge, this study is the first to perform comprehensive sensitivity and correlation analyses to evaluate the relationship between dataset complexity exclusion and the accuracy of DL classifier. In contrast to conventional sensitivity analyses which only evaluate changes in model output with respect to input changes, this study introduces a novel Sensitivity Correlation Score (SC-score). The SC-score combines sensitivity and correlation analyses into a single metric formulated as the product of the Absolute Sensitivity Function and Pearson Correlation Coefficient which is normalised for interpretability. This formulation rewards positive sensitivity and strong correlation while neutralising the effects of negative correlation. The SC-score successfully evaluated both the responsiveness and consistency of the performance enhancement of the DL model owing to the elimination of dataset complexities. To demonstrate the robustness of this study, the proposed data-centric DL-based approach was validated on an external testing dataset from diverse agricultural environments and achieved an accuracy improvement of 10.94%. This study demonstrates the strength of data-centric AI in solving industry-oriented problems in real-world applications

• Journal Article Type: Article
• Acceptance Date: May 26, 2025
• Online Publication Date: May 29, 2025
• Deposit Date: May 30, 2025
• Publicly Available Date: Jun 2, 2025
• Journal: Expert Systems with Applications
• Print ISSN: 0957-4174
• Publisher: Elsevier
• Peer Reviewed: Peer Reviewed
• Article Number: 128365
• DOI: https://doi.org/10.1016/j.eswa.2025.128365