High-Throughput Rice Density Estimation from Transplantation to Tillering Stages Using Deep Networks

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Accurate rice density estimation is crucial for yield prediction, growth diagnosis, and agricultural management.
• Current manual counting methods are inefficient and prone to inaccuracies.
• Computer vision (CV) offers automation potential but faces challenges in variable field conditions.

Purpose of the Study:

• To develop an automated, accurate, and efficient method for rice density estimation using deep learning.
• To address challenges in CV-based counting, including illumination, scale, and appearance variations.

Main Methods:

• Proposed the Scale-Fusion Counting Classification Network (SFC²Net), a deep learning model.
• Employed a multicolumn pretrained network and multilayer feature fusion to enhance feature representation.
• Utilized a blockwise classification strategy to handle scale-induced sample imbalance.

Main Results:

• SFC²Net achieved high accuracy on the Rice Plant Counting (RPC) dataset with MAE of 25.51 and R² of 0.98.
• Demonstrated a 48.2% relative improvement in Mean Absolute Error (MAE) compared to the CSRNet approach.
• Achieved high-throughput processing at 16.7 frames per second for 1024x1024 images.

Conclusions:

• SFC²Net effectively automates rice density estimation, offering a reliable alternative to manual counting.
• The model's performance suggests its suitability for early-stage rice growth monitoring and management.
• The developed approach shows significant potential for improving agricultural data collection efficiency and accuracy.