Transfer learning models for wheat ear detection on multi-source dataset.

Area of Science:

• Agricultural Science
• Computer Vision
• Data Science

Background:

• Accurate wheat yield forecasting is vital for global food security, yet current methods are often biased or labor-intensive.
• Manual counting of wheat ears is a precise but time-consuming method for yield prediction.
• Automated wheat ear detection can improve the accuracy and efficiency of yield estimation.

Purpose of the Study:

• To introduce the BioS-Wheat dataset, a novel, high-quality RGB smartphone image dataset for wheat ear detection.
• To evaluate the performance of six deep learning models for automated wheat ear detection.
• To provide a baseline for future research in automated wheat yield prediction.

Main Methods:

• Creation of the BioS-Wheat dataset: 5,696 annotated RGB images across four wheat varieties with high sowing density and minimal row spacing.
• Evaluation of six deep learning models, including RetinaNet, YOLOv8, and RT-DETR, for wheat ear detection.
• Performance assessment using mean Average Precision (mAP@50).

Main Results:

• RetinaNet, YOLOv8, and RT-DETR achieved the highest mAP@50 of 91% for wheat ear detection.
• These top-performing models exhibited significantly higher computational complexity.
• The BioS-Wheat dataset's complexity, due to dense arrangements and occlusion, challenges model robustness.

Conclusions:

• The BioS-Wheat dataset offers a valuable resource for developing robust automated wheat ear detection models.
• Deep learning models show high potential for accurate, early-stage wheat yield prediction.
• Agronomic diversity in datasets is crucial for enhancing model performance across various conditions.