MTSC-Net: A Semi-Supervised Counting Network for Estimating the Number of Slash pine New Shoots.

Area of Science:

• Forestry Science
• Computer Vision
• Machine Learning

Background:

• Slash pine new shoot density is crucial for growth and photosynthetic capacity assessment.
• Automated new shoot counting using deep learning is hindered by costly data collection and labeling.

Purpose of the Study:

• To propose MTSC-Net, a semi-supervised counting network for estimating slash pine new shoot numbers.
• To overcome limitations of traditional data-intensive deep learning methods in forestry applications.

Main Methods:

• Utilized a mean-teacher framework with improved VGG19 for multiscale feature extraction.
• Incorporated an attention feature fusion module for effective feature integration.
• Employed multiscale dilated convolution and masked image modeling for enhanced counting accuracy.

Main Results:

• MTSC-Net demonstrated superior performance compared to other semi-supervised models across various labeled data percentages (5%-50%).
• Achieved a mean absolute error of 17.71 and root mean square error of 25.49 at 5% labeled data.
• Validated the model's efficiency in reducing the need for extensive labeled datasets.

Conclusions:

• MTSC-Net offers an efficient semi-supervised approach for automated slash pine new shoot counting.
• The method provides valuable automated support for tree breeding and genetic utilization programs.
• Highlights the potential of semi-supervised learning in addressing data scarcity in ecological studies.