A unified YOLOv11 framework for multi-crop detection of diseases, pests and nutrient deficiencies, with preliminary field validation and mobile decision support
Abstract
Diseases, insect pests and nutrient deficiencies account for substantial annual crop losses worldwide, yet field diagnosis remains largely dependent on expert visual inspection. Although deep learning has shown considerable promise, most existing studies focus on a single crop or stress category under controlled conditions, with limited attention to nutrient deficiencies and few progressing to real-world deployment. This study presents a unified crop-stress detection framework covering 53 classes (41 biotic stresses, five nutrient deficiencies and seven healthy classes) across seven economically important crops of the Oriental region of Morocco. A dual-device dataset of 159,000 images was assembled, 90.4% acquired in the field with a DSLR camera and a smartphone and 9.6% from open-access sources used for training only. A lightweight YOLOv11n detector was trained at 832 × 832 resolution, with all settings selected on a separate field-acquired validation partition. On an independent test partition not used for model selection, it achieved Precision = 0.977, Recall = 0.972, F1-score = 0.974, mAP@50 = 0.975 (95% CI 0.973–0.977) and mAP@50–95 = 0.953, outperforming YOLOv8n trained under the same protocol on every metric (mAP@50–95 = 0.934) whilst being lighter. In the complete 53-class confusion matrix, all residual misclassifications (3.2%) occurred between visually similar classes of the same stress category, a pattern requiring confirmation on externally acquired data. Performance did not depend on crop recognition alone, as multi-host classes reached a mean mAP@50 comparable to that of single-crop classes (0.968 vs 0.976). Preliminary independent validation on 150 farmer-captured field images covering five stress classes yielded an overall accuracy of 98.7% (148/150; 95% CI 95.3–99.6%). The exported ONNX model reproduced all class predictions of the original PyTorch model on these 150 images. The model was integrated into a multilingual, offline-capable mobile application developed with Flutter and ONNX Runtime Mobile. A preliminary usability assessment with the System Usability Scale, completed by the same 50 farmers, gave a mean score of 93.7 (SD 3.2), a very high value that probably also reflects social desirability and a ceiling effect. This work represents a step towards translating AI-based crop-stress detection from experimental research to practical decision support for farmers.
Related
- Translation, Cross-Cultural Adaptation and Preliminary Validation of the Italian Version of the Early Cognitive and Listening Links (EarlyCaLL) Framework.
- Wildlife disease surveillance under uncertainty: an adaptive search-theoretic framework for early detection of transboundary animal diseases
- A representation-level wavelet–EMD fusion framework for spectral analysis and detection of epileptiform patterns in pediatric EEG: computational validation with a roadmap for external benchmarking
- Development and preliminary validation of an omp25-based RPA-CRISPR/Cas12a visual assay for rapid screening of brucellosis-associated Brucella DNA
- Preliminary Evaluation of Two iiPCR Platforms for Qualitative Field Detection of Cannabis sativa
BibTeX
@article{hamdaoui2026unified,
title = {{A unified YOLOv11 framework for multi-crop detection of diseases, pests and nutrient deficiencies, with preliminary field validation and mobile decision support}},
author = {Hajar Hamdaoui and Hind Mesbah and Mashail N. Alzain and Yassine Zarrouk and Hanae Al Kaddouri and Omar Noman and Noureddine Kouddane},
journal = {Frontiers in Plant Science},
year = {2026},
month = nov,
doi = {10.3389/fpls.2026.1989847},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1989847},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A unified YOLOv11 framework for multi-crop detection of diseases, pests and nutrient deficiencies, with preliminary field validation and mobile decision support
AU - Hajar Hamdaoui
AU - Hind Mesbah
AU - Mashail N. Alzain
AU - Yassine Zarrouk
AU - Hanae Al Kaddouri
AU - Omar Noman
AU - Noureddine Kouddane
PY - 2026
DA - 2026/11/13
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1989847
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1989847
AB - Diseases, insect pests and nutrient deficiencies account for substantial annual crop losses worldwide, yet field diagnosis remains largely dependent on expert visual inspection. Although deep learning has shown considerable promise, most existing studies focus on a single crop or stress category under controlled conditions, with limited attention to nutrient deficiencies and few progressing to real-world deployment. This study presents a unified crop-stress detection framework covering 53 classes (41 biotic stresses, five nutrient deficiencies and seven healthy classes) across seven economically important crops of the Oriental region of Morocco. A dual-device dataset of 159,000 images was assembled, 90.4% acquired in the field with a DSLR camera and a smartphone and 9.6% from open-access sources used for training only. A lightweight YOLOv11n detector was trained at 832 × 832 resolution, with all settings selected on a separate field-acquired validation partition. On an independent test partition not used for model selection, it achieved Precision = 0.977, Recall = 0.972, F1-score = 0.974, mAP@50 = 0.975 (95% CI 0.973–0.977) and mAP@50–95 = 0.953, outperforming YOLOv8n trained under the same protocol on every metric (mAP@50–95 = 0.934) whilst being lighter. In the complete 53-class confusion matrix, all residual misclassifications (3.2%) occurred between visually similar classes of the same stress category, a pattern requiring confirmation on externally acquired data. Performance did not depend on crop recognition alone, as multi-host classes reached a mean mAP@50 comparable to that of single-crop classes (0.968 vs 0.976). Preliminary independent validation on 150 farmer-captured field images covering five stress classes yielded an overall accuracy of 98.7% (148/150; 95% CI 95.3–99.6%). The exported ONNX model reproduced all class predictions of the original PyTorch model on these 150 images. The model was integrated into a multilingual, offline-capable mobile application developed with Flutter and ONNX Runtime Mobile. A preliminary usability assessment with the System Usability Scale, completed by the same 50 farmers, gave a mean score of 93.7 (SD 3.2), a very high value that probably also reflects social desirability and a ceiling effect. This work represents a step towards translating AI-based crop-stress detection from experimental research to practical decision support for farmers.
SN - 1664-462X
ER -