EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Lightweight Convolutional Neural Networks for Wheat Disease Detection and Classification: A Review and Comparative Study From Field to Storage

Lynda Djakhdjakha, Siham Amrouch, Hiba Abdelmoumene, Imane Bouacida, Brahim Farou

IEEE Access · 2026 · 14 · 125087-125111

DOI: 10.1109/access.2026.3721950 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Lynda Djakhdjakha und 4 weiteren Autor:innen. Die Arbeit nennt 116 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The increasing adoption of Lightweight Convolutional Neural Networks (LCNNs) for agricultural applications has highlighted their potential for automated wheat disease detection and classification across both field (pre-harvest) and post-harvest storage environments. However, the effectiveness and deployability of such systems depend not only on model architecture but also critically on the quality and suitability of the datasets used for training and evaluation. In this paper, two complementary reviews are presented to address these intertwined challenges. First, a systematic review of LCNN architectures is conducted, focusing on models applied to wheat-related image analysis across both field and storage scenarios, with particular emphasis on classification performance, model compactness, computational efficiency, and compatibility with edge and embedded platforms, especially in the context of deployment under storage conditions. Second, a critical review of existing wheat grain image datasets is provided, examining dataset scale, acquisition conditions, annotation strategies, class diversity, and accessibility, with an emphasis on their relevance to storage-based disease detection tasks. The integrated analysis reveals a strong interdependence between architectural performance and dataset characteristics. Reported gains in LCNN accuracy are often influenced by dataset bias, controlled imaging conditions, and class imbalance, which can obscure real-world performance in storage environments. Lightweight models, while advantageous for deployment, are shown to be especially sensitive to limitations in data realism and annotation quality. Moreover, a mismatch is identified between current dataset acquisition practices and the operational conditions encountered in grain storage, including bulk aggregation, occlusion, dust, and uneven illumination. The findings highlight a fragmentation between model development and dataset construction in existing research. This study emphasizes the need for a co-design paradigm in which LCNN architectures and storage-realistic datasets are developed in tandem.

Zitieren

APA

Djakhdjakha, L., Amrouch, S., Abdelmoumene, H., Bouacida, I., & Farou, B. (2026). Lightweight Convolutional Neural Networks for Wheat Disease Detection and Classification: A Review and Comparative Study From Field to Storage. IEEE Access, 14, 125087-125111. https://doi.org/10.1109/access.2026.3721950

Vancouver

Djakhdjakha L, Amrouch S, Abdelmoumene H, Bouacida I, Farou B. Lightweight Convolutional Neural Networks for Wheat Disease Detection and Classification: A Review and Comparative Study From Field to Storage. IEEE Access. 2026;14:125087-125111. doi:10.1109/access.2026.3721950

BibTeX
@article{Djakhdjakha2026,
  author = {Djakhdjakha, Lynda and Amrouch, Siham and Abdelmoumene, Hiba and Bouacida, Imane and Farou, Brahim},
  title = {Lightweight Convolutional Neural Networks for Wheat Disease Detection and Classification: A Review and Comparative Study From Field to Storage},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {125087--125111},
  doi = {10.1109/access.2026.3721950}
}
Kostenlos starten →