Plant Detection in RGB Images from Unmanned Aerial Vehicles Using Segmentation by Deep Learning and an Impact of...

Area of Science:

• Agricultural Science
• Computer Vision
• Remote Sensing

Background:

• Precision agriculture relies on Unmanned Aerial Vehicle (UAV) monitoring for plant growth control.
• Plant stand counting is a key task in field monitoring, essential for yield prediction and seedling assessment.
• Current methods use computer vision and deep learning but are sensitive to image resolution and data quality.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for plant detection in UAV imagery.
• To assess the impact of training data diversity and quality on plant detection accuracy.
• To analyze how automatic plant detection accuracy affects crop pattern evaluation using texture features.

Main Methods:

• Utilized deep learning algorithms, specifically convolutional neural networks, for image segmentation-based plant detection.
• Trained and tested models using 12 orthomosaics from Russia and 17 external datasets from Roboflow.
• Compared texture features (GLCM mean, GLRM long run, GLRM run ratio) between manual and automated plant mask estimations.

Main Results:

• Incorporating varied image datasets, including lower resolution and quality, substantially enhanced plant stand counting performance.
• Higher accuracy in automatic plant detection correlated with better agreement in texture parameter estimations compared to manual methods.
• Certain texture features showed close agreement between manual and automated estimates, while others revealed significant differences.

Conclusions:

• Data augmentation, even with lower quality images, is vital for improving the robustness of plant detection models.
• The accuracy of automated plant detection directly influences the reliability of crop cropping pattern evaluations.
• While some texture features are robust to detection accuracy, others require highly accurate plant masks for dependable analysis.