Counting Canola: Toward Generalizable Aerial Plant Detection Models

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Manual plant counting is labor-intensive and relies on sampling.
• Deep learning offers automated plant population estimation from aerial imagery.
• Current models struggle with diverse or unseen image conditions.

Purpose of the Study:

• To investigate how training dataset characteristics influence deep learning model generalizability for plant detection.
• To determine the impact of training data size, diversity, and quality on model performance.
• To introduce a new tool and dataset for remote-sensed aerial plant detection.

Main Methods:

• Utilized deep learning object detection models on aerial canola field imagery.
• Experimented with varying training set sizes, diversity, and annotation quality.
• Developed and used the 'Canola Counter' web tool for dataset preparation and annotation.

Main Results:

• Increasing training set size alone does not close the performance gap for unseen data.
• Training set diversity significantly improves model generalizability.
• Different types of annotation noise lead to varied model behaviors on out-of-distribution data.

Conclusions:

• Model generalizability in aerial plant detection is highly dependent on training data diversity and quality.
• The 'Canola Counter' tool and associated dataset facilitate advancements in remote-sensed crop monitoring.
• Future work should focus on creating diverse and high-quality training datasets for robust AI in agriculture.