Instance Segmentation to Estimate Consumption of Corn Ears by Wild Animals for GMO Preference Tests

Area of Science:

• Agricultural Science
• Computer Science
• Ecology

Background:

• Wild animals' feeding preferences between GMO and non-GMO corn are not fully understood.
• Assessing corn consumption in wildlife requires accurate, scalable methods for analyzing image data.

Purpose of the Study:

• To develop and validate a deep learning-based image processing pipeline for estimating corn consumption by wild animals.
• To apply this pipeline to analyze image data from a Genetically Modified (GMO) Corn Experiment.

Main Methods:

• Instance segmentation using Mask Regional Convolutional Neural Network (Mask R-CNN) to identify corn and bare cob in images.
• Comparison of two segmentation approaches: whole corn ears vs. bare cob without kernels.
• Ablation studies to optimize the Mask R-CNN model for corn consumption estimation.

Main Results:

• The approach segmenting bare cob without kernels demonstrated superior performance and accuracy.
• The optimized Mask R-CNN model achieved high accuracy in corn consumption estimation (R² = 0.99) compared to manual labeling.
• The developed pipeline effectively processes complex image data from the GMO Corn Experiment.

Conclusions:

• Deep learning-based instance segmentation with Mask R-CNN provides a highly accurate method for estimating corn consumption.
• This approach can be reliably applied to analyze extensive image datasets for wildlife feeding preference studies.
• The methodology holds potential for broader applications in plant phenotyping, including yield estimation and stress quantification.