🌴 SODA: Saudi Organic Date Fruits Dataset

Welcome to SODA (Saudi Organic Date fruits dataset), a comprehensive, class-balanced dataset designed to advance fine-grained image classification, precision agriculture, and localized computer vision applications.

While Saudi Arabia is a leading global producer of date palm fruits, high-quality, well-curated, and publicly available datasets representing local varieties have historically been limited. SODA bridges this gap by offering a unique dual-modality resource—combining raw imagery, AI-processed backgrounds, extracted anatomical parts (pits), and deep-learning-generated tabular features—making it perfect for both deep learning and classical machine learning experimentation.

---

💡 Behind the Name: Why "SODA"?

Linguistic Note: While SODA serves as an acronym for this dataset, the word SODA (سوداء) literally means "black" in the Arabic language. It is traditionally used to refer to date fruits because most varieties mature into rich, deep, blackish colors as they ripen. It is a nod to local heritage and visual characteristics, completely unrelated to soda carbonated drinks!

---

🚀 Key Features & Dataset Composition

SODA is uniquely engineered to overcome the limitations of existing datasets by focusing strictly on individual fruits and their corresponding pits, enabling highly detailed visual and feature-based analysis.

• 100% Class-Balanced: Contains exactly 2,613 high-quality images uniformly distributed across 13 distinct organic Saudi date varieties (201 images per class).
• Dual-Anatomy Insights: The first publicly available dataset to include images of both the individual whole fruits and their corresponding inner pits, unlocking richer multi-view structural analysis.
• Background-Removed Variants: Includes professionally curated raw images alongside background-removed counterparts processed via rembg, ensuring clean, distraction-free feature mapping.
• Pre-computed Tabular Embeddings: To accelerate experimentation, the dataset includes 5 ready-to-use tabular CSV files containing feature embeddings extracted using top pretrained CNN architectures: MobileNetV3, EfficientNetV2, ResNet50, ResNetV2, and InceptionV3.
• Expanded Botanical Diversity: Features four prominent Saudi varieties (Raziz, Shalabi, Anbara, and Wanan) completely absent from previously published academic datasets.

---

📊 Covered Varieties & Regional Origins

The dataset captures the rich agricultural heritage of Saudi Arabia, spanning multiple producing regions:

Variety (English)	Variety (Arabic)	Farm Source	Region in Saudi Arabia
Ajwa	عجوة	Yala	Al-Madinah (Western)
Medjool	مجدول	Yala	Al-Qassim (Central)
Khalas	خلاص	Al Yasmeen	Al-Ahsa (Eastern)
Raziz	رزیز	Al Yasmeen	Al-Ahsa (Eastern)
Sukari Galaxy	جالكسي سكري	Al Dka	Al-Ghat (Central)
Sukari Rutab	رطب سكري	Yala	Al-Qassim (Central)
Sukari Mofattal	مقتل سكري	Yala	Al-Qassim (Central)
Shalabi	شلبي	Al Mohsineyah	Al-Madinah (Western)
Shaishi	شيشي	Al Yasmeen	Al-Ahsa (Eastern)
Safawi	صفاوي	Yala	Al-Madinah (Western)
Sagai	صقعي	Yala	Al-Qassim (Central)
Anbara	عنبرة	Al Mohsineyah	Al-Madinah (Western)
Wanan	ونان	Al Dka	Al-Ghat (Central)

---

📈 Baseline Machine Learning Benchmarks

SODA comes out of the box with proven baseline benchmarks evaluated using stratified 10-fold cross-validation within the Orange Data Mining platform:

• Top Performing Model: A shallow Neural Network paired with MobileNetV3 deep embeddings achieved a peak AUC of 95.7% (Classification Accuracy: 70.9%, F1-Score: 71.0%).
• Alternative Efficient Configurations: Logistic Regression paired with ResNet50 features yielded comparable high-tier results, establishing both as highly robust baselines for quick, lower-resource deployment.
• No-Code Workflow Included: The complete, reproducible Orange Data Mining pipelines are packed with the dataset, allowing beginners and experts alike to quickly train and compare shallow vs. deep models.

---

🎯 Potential Applications

• 🌾 Precision Agriculture & Smart Sorting: Train edge-AI systems (TinyML) and robotic sorters for post-harvest grading, packaging, and automatic quality control.
• 🎓 Contextual AI Education: Ideal as a culturally relevant, real-world educational alternative to generic standard datasets (like Iris) for teaching complete end-to-end ML pipelines.
• 🗺️ Cultural & Tourism AI: Power mobile apps that help visitors dynamically identify date varieties, learn about their taste profiles, regional origins, and cultural significance.

---

📜 Citation & Acknowledgments

If you use this dataset in your research or educational projects, please cite the official paper in: https://doi.org/10.21608/jesaun.2026.420750.1727

@article{SODA2026,
  title={SODA: A Balanced Dataset for Saudi Arabian Organic Date Fruits},
  author={Bati, Ghassan F. and Kurdi, Ghader R.},
  journal={Journal of Engineering Sciences (JES), Faculty of Engineering, Assiut University},
  doi={10.21608/jesaun.2026.420750.1727},
  year={2026}
}

A Note of Gratitude: The authors owe a great debt of gratitude to their beloved students (computer engineers to be) who participated passionately in collecting the SODA dataset.