Automatic counting of rapeseed inflorescences using deep learning method and UAV RGB imagery

Area of Science:

• Agricultural Science
• Computer Vision
• Remote Sensing

Background:

• Rapeseed (Brassica napus L.) seed yield is critically dependent on the total number of flowers produced.
• Accurate quantification of inflorescences is essential for yield prediction and crop management.
• Manual counting of rapeseed inflorescences is labor-intensive and prone to errors.

Purpose of the Study:

• To develop an automated, low-cost method for counting rapeseed inflorescences using Unmanned Aerial Vehicle (UAV) imagery.
• To establish a benchmark dataset for rapeseed inflorescence detection and counting.
• To evaluate the model's performance against existing object detection methods and assess its correlation with crop yield.

Main Methods:

• Developed a YOLOv5-based object detection model incorporating the Convolutional Block Attention Module (CBAM).
• Utilized Red-Green-Blue (RGB) imagery captured by a DJI Phantom 4 Pro V2.0 UAV.
• Constructed the Rapeseed Inflorescence Benchmark (RIB) dataset with 165 plot images and 60,000 manual labels.

Main Results:

• The proposed model achieved a counting accuracy indicator (R) over 0.96 and a mean Average Precision (mAP) for location exceeding 92% on the RIB dataset.
• Demonstrated state-of-the-art counting performance and superior location accuracy compared to Faster R-CNN, YOLOv4, CenterNet, and TasselNetV2+.
• Identified a significant positive correlation between automatically counted inflorescences and actual crop yield at the field plot level.

Conclusions:

• The developed UAV-assisted method provides an effective and accurate approach for quantifying rapeseed inflorescences.
• This automated system supports dynamic monitoring of flower richness and aids in yield prediction.
• The findings facilitate the breeding of high-yield rapeseed varieties by providing robust data for selection.