EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Explainable Deep Learning for Prickly Pear Cladode Condition Classification: Reproducible Evaluation and Mobile Deployment

Jose Antonio Rivas-Navarrete, Rosa Janette Perez-Chimal, Rosalia Vergara-Esparza, Carlos Mauricio Lastre-Domínguez, Jorge Ulises Muñoz-Minjares

AgriEngineering · 15.11.2026 · 8 (11) · 491

DOI: 10.3390/agriengineering8110491 ↗

Kurz gefasst

Fachartikel in der Zeitschrift AgriEngineering, online erschienen am 15.11.2026. Verfasst von Jose Antonio Rivas-Navarrete und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Prickly pear (Opuntia ficus-indica) cladodes may exhibit physiological differences, abiotic injuries, and tissue deterioration that can affect crop management and product quality. However, their visual assessment is commonly based on manual inspection and may be subjective, particularly when different conditions exhibit similar surface characteristics. This study presents a deep learning framework for multiclass visual classification using RGB images, with emphasis on reproducible evaluation, mobile deployment, and Grad-CAM interpretability. A dataset comprising 1328 regions of interest (ROIs) extracted from 78 field-acquired photographs was organized into five categories: healthy young and mature cladodes (SANV and SANP), early and advanced fertilizer-associated injury (CAT and CAA), and rot and tissue deterioration (PUD). Four convolutional architectures were evaluated using photograph-disjoint partitions and five training repetitions, complemented by ablation experiments, hyperparameter optimization, and statistical comparisons. MobileNetV2 achieved the highest mean accuracy (81.20±1.57%), F1-score (75.84±1.61%), and balanced accuracy (76.15±1.48%). However, photograph-level bootstrap confidence intervals did not establish a clear F1-score advantage over MobileNetV3Small. Ablation experiments demonstrated the importance of pretrained representations and the benefits of online augmentation, whereas Hyperopt did not improve average classification performance. TensorFlow Lite FP16 reduced MobileNetV2 storage requirements to 4.31 MiB while largely preserving predictive performance. On a Samsung SM-A156U smartphone, MobileNetV2 FP16 achieved a median inference latency of 26.59 ms, compared with 8.21 ms for MobileNetV3Small FP16. Grad-CAM provided qualitative visualizations of correct and incorrect predictions. These results demonstrate the feasibility of mobile-assisted visual classification under the evaluated acquisition conditions, while further validation across independent plants, locations, and seasons remains necessary.

Zitieren

APA

Rivas-Navarrete, J. A., Perez-Chimal, R. J., Vergara-Esparza, R., Lastre-Domínguez, C. M., & Muñoz-Minjares, J. U. (2026). Explainable Deep Learning for Prickly Pear Cladode Condition Classification: Reproducible Evaluation and Mobile Deployment. AgriEngineering, 8(11), 491. https://doi.org/10.3390/agriengineering8110491

Vancouver

Rivas-Navarrete JA, Perez-Chimal RJ, Vergara-Esparza R, Lastre-Domínguez CM, Muñoz-Minjares JU. Explainable Deep Learning for Prickly Pear Cladode Condition Classification: Reproducible Evaluation and Mobile Deployment. AgriEngineering. 2026;8(11):491. doi:10.3390/agriengineering8110491

BibTeX
@article{RivasNavarrete2026,
  author = {Rivas-Navarrete, Jose Antonio and Perez-Chimal, Rosa Janette and Vergara-Esparza, Rosalia and Lastre-Domínguez, Carlos Mauricio and Muñoz-Minjares, Jorge Ulises},
  title = {Explainable Deep Learning for Prickly Pear Cladode Condition Classification: Reproducible Evaluation and Mobile Deployment},
  journal = {AgriEngineering},
  year = {2026},
  volume = {8},
  number = {11},
  pages = {491},
  doi = {10.3390/agriengineering8110491}
}
Kostenlos starten →