Supervised and Weakly Supervised Deep Learning for Segmentation and Counting of Cotton Bolls Using Proximal Imagery

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Breeding

Background:

• Accurate cotton boll counting is crucial for crop yield estimation and breeding.
• Deep learning offers potential for automated phenotypic trait measurement.
• A bottleneck exists in supervised learning due to extensive data annotation requirements for cotton boll counting.

Purpose of the Study:

• To develop and compare fully supervised and weakly supervised deep learning models for segmenting and counting cotton bolls from proximal RGB images.
• To evaluate the accuracy and annotation cost-efficiency of different deep learning approaches for cotton boll counting.

Main Methods:

• Collected 290 RGB images of cotton plants in various settings (potted, in-field).
• Processed images into 4350 tiles for training and testing.
• Implemented and compared two supervised models (Mask R-CNN, S-Count) and two weakly supervised models (WS-Count, CountSeg).

Main Results:

• Fully supervised models (S-Count, Mask R-CNN) achieved lower RMSE (1.181, 1.175) for counts <10 bolls.
• Weakly supervised models (WS-Count, CountSeg) showed competitive performance with RMSEs of 1.826 and 1.284.
• Weakly supervised approaches demonstrated at least 10x greater cost efficiency in data annotation.

Conclusions:

• Weakly supervised deep learning models provide a viable and cost-effective solution for cotton boll counting.
• These models can be extended to other plant organs and benefit breeders, physiologists, and growers.
• Automated boll counting using low-cost RGB imagery enhances cotton breeding and yield prediction.