Sesame Plant Segmentation Dataset

• Sesame Plant Segmentation Dataset is an annotated agricultural imagery collection providing pixel-level, polygonal masks for precise segmentation of sesame plants.
• It features YOLO-compatible labels derived from real-world farm images in Northern Nigeria, supporting robust deep learning workflows in precision agriculture.
• Baseline results using YOLOv8 show strong performance in detection and instance segmentation, underlining the dataset's potential for yield estimation and disease analysis.

The Sesame Plant Segmentation Dataset is an annotated collection of agricultural imagery tailored to the semantic segmentation of sesame plants for precision farming applications. Sourced from real-world farm environments in Northern Nigeria, the dataset provides YOLO-compatible, pixel-level annotations suitable for deep learning workflows targeting detection and fine-grained segmentation of sesame in challenging, naturally-lit conditions. Its design fills a regional and technical gap in publicly available agricultural vision resources, enabling the deployment of AI-based monitoring and automation tools directly in smallholder farm contexts (Muhammad et al., 12 Jan 2026).

1. Dataset Scope and Objectives

The primary aim of the Sesame Plant Segmentation Dataset is to facilitate the development and benchmarking of AI models capable of high-precision identification and delineation of sesame plants at an instance level. Unlike prior efforts limited to bounding boxes, this resource delivers fine-grained, polygonal masks, supporting:

• Precise plant localization even in dense agricultural fields
• Enhanced performance in downstream vision tasks such as automated yield estimation, disease/stress phenotype analysis, and real-time monitoring
• Customization for the agronomic and environmental conditions specific to the savanna zone of Northern Nigeria

This specificity is intended to produce models robust to local environmental variations and to bridge gaps for underrepresented crops and growing regions.

2. Data Acquisition and Annotation Workflow

Image data were gathered from farms in Jirdede, Daura Local Government Area, Katsina State, Nigeria, prioritizing rainy-season, daylight conditions to reflect typical operational scenarios. Acquisition leveraged an iPhone 11 (wide f/1.8 and ultra-wide f/2.4, 12 MP), capturing multiple perspectives (horizontal, top, side) of sesame plants during their early growth (45–85 days post-planting)—with 95% of sampled plants being healthy.

Annotation followed a model-assisted protocol:

• Initial segmentation was performed using Meta’s Segment Anything Model v2 (SAM-2), executed within the Roboflow platform.
• Local sesame farmers, acting as domain experts, manually corrected these masks for botanical fidelity, ensuring pixel-level accuracy and minimizing false or incomplete delineations.
• The joint pipeline yielded high-quality, instance-segmented masks suitable for research and production YOLO workflows.

3. Data Structure, Format, and Label Taxonomy

The dataset comprises 293 images, partitioned as follows:

Split	Images	Percentage
Training	206	70%
Validation	43	15%
Testing	43	15%

Directory structure is fully YOLO-compatible:

$[0,1]$0

Label files use the YOLO segmentation format, with a single class index (0 for “SESAME_PLANT”). Each object is specified on one line as: $[0,1]$1 where $(x_i, y_i)$ are the normalized polygon vertices of the instance mask, and all coordinates are in $[0,1]$.

All images are available as JPEG or PNG. Labels are plain-text .txt files. The entire resource is open-access via Kaggle.

4. Evaluation Metrics

Model evaluation strictly follows contemporary object detection and segmentation conventions, with explicit metric definitions:

• Intersection over Union (IoU):

$$\mathrm{IoU} = \frac{ |A_{\text{pred}} \cap A_{\text{gt}}| }{ |A_{\text{pred}} \cup A_{\text{gt}}| }$$

• Precision: $P = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP}}$
• Recall: $R = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FN}}$
• Average Precision at IoU $\tau$:

$\mathrm{AP}_\tau = \int_0^1 p(r) dr$

with $p(r)$ the precision as a function of recall.

• [email protected]: AP at a single 0.50 IoU threshold.
• [email protected]:0.95: Mean AP over ten thresholds, $\tau\in\{0.50,0.55,\dots,0.95\}$:

$$\mathrm{[email protected]:0.95} = \frac{1}{10} \sum_{k=0}^9 \mathrm{AP}_{\tau=0.50+0.05k}$$

These metrics enable rigorous, standardized reporting and facilitate direct comparison with prior segmentation and detection datasets.

5. Baseline Model Results and Benchmarking

Using Ultralytics YOLOv8 (default hyperparameters; 100 epochs; batch size 16; input 640×640):

• Bounding-Box Detection:
  • Precision: 79%
  • Recall: 79%
  • [email protected]: 84%
  • [email protected]:0.95: 58%
• Instance Segmentation:
  • Precision: 77%
  • Recall: 82%
  • [email protected]: 84%
  • [email protected]:0.95: 52%

These results confirm strong baseline performance for both detection and pixel-level segmentation tasks, validating dataset integrity and demonstrating its suitability for immediate deployment within state-of-the-art model frameworks.

6. Applications, Limitations, and Prospective Directions

Key Applications

• Precision Agriculture: Enables real-time crop monitoring, disease/stress detection, and autonomous treatment (smart spraying, irrigation) using drones or robotic platforms.
• Yield Estimation and Phenotyping: Fine-grain masks support advanced phenotypic trait extraction, enabling automated yield and health assessments.
• Agricultural Research: Supports computer vision studies targeting locally relevant crop species, closing the data gap for sesame in the Sub-Saharan context.

Limitations

• Coverage is limited to early growth stages (45–85 days) and typical rainy-season daylight, omitting phenological and illumination diversity seen across the full growing season.
• Only healthy plants of a single class (“SESAME_PLANT”) are annotated; no negative (weed, diseased, or other crop) samples are included, restricting the direct applicability to multi-class segmentation tasks.
• Data is geographically constrained to one Local Government Area, potentially restricting model generalization outside Northern Nigeria.

Future Enhancements

Planned developments include extending the dataset to later growth stages, adding images under diverse lighting/weather conditions (e.g., dawn, dusk, overcast), introducing multi-class annotations (weeds, stressed or diseased plants), and expanding the geographic sampling base to enable pan-regional generalization.

7. Access and Community Impact

The Sesame Plant Segmentation Dataset is available freely via Kaggle, facilitating broad adoption and benchmarking. It addresses the urgent need for regionally appropriate, pixel-level annotated data to support advanced computer vision solutions in smallholder-dominated agricultural systems across Nigeria and similar environments. This resource is positioned to catalyze research in plant monitoring, disease detection, and AI-driven agricultural automation, directly targeting productivity and sustainability challenges in emerging economies (Muhammad et al., 12 Jan 2026).