Detection and analysis of wheat spikes using Convolutional Neural Networks.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Remote sensing and machine learning are increasingly used for high-throughput crop analysis.
• Accurate wheat spike detection is crucial for estimating grain production and selecting high-yielding varieties.

Purpose of the Study:

• To develop and validate a deep learning approach for accurate wheat spike detection, counting, and yield estimation.
• To create a labeled dataset (SPIKE) for training and evaluating Convolutional Neural Networks (CNNs).

Main Methods:

• Applied region-based Convolutional Neural Networks (R-CNN) to analyze high-definition RGB images of wheat fields.
• Trained four R-CNN models using the manually annotated SPIKE dataset, capturing images at various growth stages.
• Tested models under challenging field conditions including variable illumination and spike occlusion.

Main Results:

• Achieved average detection accuracy ranging from 88% to $$ across different R-CNN models and test image sets.
• Identified the most robust R-CNN model for analyzing variations in spike production across 10 wheat varieties and three fertilizer treatments.
• Demonstrated the effectiveness of the SPIKE dataset and trained CNNs in quantifying yield differences.

Conclusions:

• Deep learning techniques, supported by comprehensive datasets like SPIKE, can achieve high accuracy in wheat spike analysis.
• The developed R-CNN model is optimized for diverse field scenarios and aids in high-throughput selection of superior wheat varieties.
• The publicly available SPIKE dataset and trained CNN model offer significant potential to advance wheat breeding programs.