Introduction
Precision agriculture technologies play a crucial role in optimizing crop management by enabling site-specific interventions, particularly in weed control. Traditional weed management methods often result in excessive herbicide use, environmental damage, and crop damage. Automated weed detection using deep learning offers a promising solution, accurately distinguishing between crops and weeds, thereby facilitating targeted removal. Saffron, a high-value crop, faces competition from invasive weeds such as Flixweed and Hoary Cress, which reduce yield and quality. This study leverages computer vision and deep learning to classify saffron and these two common weeds under natural field conditions. Convolutional Neural Networks (CNNs), were employed due to their proven effectiveness in image classification tasks. Transfer learning was applied to enhance model performance by utilizing pre-trained weights from ImageNet. The research aims to develop a robust classification model that can support precision agriculture tools, such as robotic weeders, by accurately identifying weeds while preserving the main crop. The success of this approach could significantly reduce herbicide use, lower production costs, and improve saffron yield through automated, site-specific weed management.
Method
A dataset of 291 field images of saffron, Flixweed, and Hoary Cress was collected under natural lighting and environmental conditions. Each image was resized to 150×150 pixels, and data augmentation techniques (e.g., rotation, flipping, scaling) were applied to expand the dataset and improve model generalization artificially. The study utilized the VGG16 CNN architecture, fine-tuned via transfer learning with ImageNet weights. The model was trained to classify the three plant categories, and its performance was evaluated using test data. Key metrics included accuracy, F1-score, and precision to assess classification effectiveness.

Results
The proposed model achieved an overall accuracy of 91% on unseen test data, with a loss of 0.3759. The F1-scores for Saffron, Hoary Cress, and Flixweed were 85%, 100%, and 86%, respectively. Notably, the model demonstrated perfect precision (100%) in distinguishing saffron from Hoary Cress, indicating no false positives for these classes. Flixweed recognition was slightly less precise but still highly effective (86% F1-score). The high classification accuracy suggests that deep learning, combined with transfer learning, is a viable approach for weed detection in precision agriculture. The model's ability to differentiate saffron from invasive weeds under real-world conditions supports its potential integration into automated weeding systems. These results suggest that robotic weeders equipped with such AI models can selectively target weeds while minimizing crop damage, thereby reducing reliance on broad-spectrum herbicides.

Conclusions
This study demonstrates the effectiveness of deep learning in distinguishing saffron from Flixweed and Hoary Cress under natural field conditions. The improved VGG16 model achieved high accuracy (91%) and near-perfect precision for certain weed classes, validating its potential for real-world agricultural applications. The findings provide a foundation for developing AI-driven weed removal robots, which could enhance precision farming by enabling targeted, sustainable weed management. Future research should focus on optimizing models for real-time processing and integration with robotic systems. Expanding the dataset to include more weed species and varying environmental conditions could further improve robustness. Overall, this work contributes to advancing precision agriculture technologies, offering a scalable solution for automated weed control in saffron fields and similar high-value crops.

Author Contributions
S.I. Saedi: Software, Methodology, Modeling, Writing of the initial draft, Final review, and editing.
H.Makarian: Review and editing of the text, Research, Data collection, Validation.
Data Availability Statement
Not applicable.

Acknowledgements
Not applicable.
Ethical Considerations
This section states ethical approval details (e.g., Ethics Committee, ethical code) and confirms adherence to ethical standards, including avoidance of data fabrication, falsification, plagiarism, and misconduct.

Conflict of Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper
Funding Statement
The author(s) received no specific funding for this research