@Article{cmc.2026.084384,
AUTHOR = {Amit Pimpalkar, Kapil N. Vhatkar, Rachna K. Somkunwar, Shweta Koparde, Dalia H. Elkamchouchi, Ateeq Ur Rehman, Pooja Verma, Salil Bharany},
TITLE = {STHarDNet: A Statistically Validated Swin Transformer–HarDNet Framework for High-Precision Plant Disease Detection and Classification},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/27345},
ISSN = {1546-2226},
ABSTRACT = {Plants are fundamental to global food security; however, plant diseases significantly reduce agricultural productivity, making early and accurate detection essential. Traditional inspection approaches rely heavily on manual observation, which is labor-intensive, subjective, difficult to scale, and susceptible to human error. In contrast, artificial intelligence (AI) combined with computer vision (CV) offers an effective solution for early-stage disease detection, minimizing yield losses while overcoming the limitations of manual monitoring systems. In this study, a novel deep learning architecture, the Swin Transformer with Harmonic Densely Connected Network (STHarDNet), is proposed. The framework integrates a Swin Transformer (ST) as the initial skip connection within a HarDNet-based U-Net architecture to precisely localize diseased regions in leaf images. Subsequently, a modified ResNet-152 model is employed for disease classification. The ST component captures long-range dependencies at high resolution, enabling enhanced feature representation and more accurate boundary delineation. To ensure robustness and reliability, extensive statistical validation techniques are applied, including 5-fold cross-validation, bootstrap confidence intervals, Gelman-Rubin convergence diagnostics, Cohen’s <i>d</i> effect size, and paired <i>t</i>-tests with Bonferroni correction. These analyses confirm both the statistical stability and practical effectiveness of the proposed model. Experiments conducted on the hybrid PlantVillage dataset, comprising 20,798 images across 17 classes, demonstrate that STHarDNet achieves an outstanding classification accuracy of 99.81%, outperforming existing methods across multiple evaluation metrics. This research establishes a reproducible and statistically validated benchmark for automated plant disease detection (PDD), supporting its scalability in precision agriculture. Furthermore, the proposed system highlights the potential of intelligent, high-accuracy tools to assist non-expert users in identifying plant diseases at early stages, thereby enabling timely intervention and improved crop management.},
DOI = {10.32604/cmc.2026.084384}
}