In-Field Nondestructive Detection of Nitrogen Status on 'Yotsuboshi' Strawberry Using Deep Learning Algorithm

Area of Science:

• Agricultural Science
• Computer Science
• Plant Science

Background:

• Optimizing nitrogen (N) management is crucial for strawberry yield and quality in open-field cultivation.
• Current methods like visual diagnosis are subjective, while tissue analysis is destructive and costly.
• Existing non-destructive imaging techniques (e.g., NDVI) require expensive multispectral systems.

Purpose of the Study:

• To develop an affordable, non-destructive method for in-field nitrogen status detection in strawberries.
• To leverage deep learning on standard RGB images for accurate nutrient assessment.
• To overcome limitations of existing subjective and costly N assessment techniques.

Main Methods:

• A deep learning model (YOLOv11) was trained on a novel dataset of strawberry leaf images classified into NormalN, LowN, and AdvancedLowN categories.
• A low-cost phenotyping cylinder was developed to standardize smartphone image acquisition under varying light conditions.
• The methodology was validated using the phenotyping cylinder and in-field tests under different lighting.

Main Results:

• The YOLOv11 model achieved over 99% mAP50 during training.
• Validation with the phenotyping cylinder yielded an 87% mAP50, with in-field tests showing 82.7% under cloudy and 79% under direct sunlight.
• Model classifications for 'NormalN' and 'LowN' strongly correlated with NDVI measurements.

Conclusions:

• Combining deep learning with RGB imaging and a phenotyping cylinder provides a rapid, cost-effective, and reliable tool for in-field nitrogen detection.
• This approach has potential applications for nitrogen management in various crops and environments.
• The study demonstrates a significant advancement over traditional, less reliable N assessment methods.