DELTA-SoyStage: A Lightweight Detection Architecture for Full-Cycle Soybean Growth Stage Monitoring

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Accurate soybean phenotyping is vital for optimizing agricultural practices and preventing significant yield losses.
• Current deep learning models often focus on limited growth stages, lacking comprehensive phenological coverage.

Purpose of the Study:

• To develop a novel, computationally efficient deep learning architecture for comprehensive soybean growth stage classification.
• To introduce a new dataset and a specialized detection head for improved soybean phenotyping.

Main Methods:

• Developed DELTA-SoyStage, an object detection architecture using EfficientNet, ChannelMapper neck, and a novel DELTA detection head.
• Created and utilized a dataset of 17,204 labeled RGB images covering nine soybean growth stages (VE to R8).
• Evaluated model performance based on average precision (AP) and computational cost (GFLOPs).

Main Results:

• DELTA-SoyStage achieved 73.9% AP with 24.4 GFLOPs, significantly outperforming baselines in computational efficiency.
• The model demonstrated a 4.2× reduction in FLOPs compared to the best baseline (DINO-Swin) with a minimal accuracy difference.
• The DELTA head and ChannelMapper neck resulted in a 43.5% parameter reduction while maintaining competitive accuracy.

Conclusions:

• The proposed DELTA-SoyStage model offers a computationally efficient solution for accurate soybean growth stage identification.
• Its lightweight design and competitive performance make it suitable for deployment on resource-constrained edge devices in precision agriculture.
• This facilitates timely decision-making for optimized crop management without cloud dependency.