Integrating Automated Labeling Framework for Enhancing Deep Learning Models to Count Corn Plants Using UAS Imagery.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate plant counting is crucial for crop management, impacting yield and quality assessment.
• Traditional methods are labor-intensive, and current automated systems often require extensive manual data labeling.

Purpose of the Study:

• To develop a robust corn counting model by integrating an automatic image annotation framework.
• To enhance the accuracy and efficiency of plant counting using Unmanned Aerial System (UAS) imagery and deep learning (DL).

Main Methods:

• Collected high-spatial-resolution UAS images of corn at the V2-V4 growth stage.
• Developed an automated image annotation process by extracting corn rows and applying image enhancement.
• Trained four DL models (InceptionV3, VGG16, VGG19, Vision Transformer) using the automatically annotated images.

Main Results:

• The automated annotation achieved 80% accuracy in identifying corn plants.
• VGG16 demonstrated superior performance among the DL models, with an F1 score of 0.955.
• The VGG16 model achieved an R² of 0.94 and an RMSE of 9.95 when compared to ground truth data.

Conclusions:

• The integrated automated image annotation framework significantly improves the scalability and consistency of DL model training for plant counting.
• This approach streamlines the development and deployment of accurate corn counting models, benefiting large-scale agricultural data management.