Qualification of Soybean Responses to Flooding Stress Using UAV-Based Imagery and Deep Learning

Area of Science:

• Agricultural Science
• Plant Breeding
• Remote Sensing

Background:

• Flooding stress significantly reduces soybean yield and seed quality.
• Conventional flood tolerance evaluation in soybean is labor-intensive and subjective.
• High-throughput phenotyping technologies offer advanced crop stress monitoring.

Purpose of the Study:

• To assess Unmanned Aerial Vehicle (UAV)-based image features for estimating flood-induced soybean injury.
• To investigate the impact of different flight heights on image feature extraction for soybean stress phenotyping.
• To develop a deep learning model for classifying soybean flooding injury scores (FIS) using aerial imagery.

Main Methods:

• Collected visual flooding injury scores (FIS) for 724 soybean plots.
• Acquired multispectral and thermal infrared imagery from UAVs at 20, 50, and 80m.
• Extracted image features including canopy temperature, NDVI, area, width, and length.
• Utilized a deep learning model to classify FIS based on extracted image features.

Main Results:

• Image features varied significantly across the three flight heights.
• The deep learning model achieved the highest classification performance (0.9) for FIS at a 20m flight height.
• Canopy temperature, NDVI, and other geometric features were effective indicators of flooding stress.

Conclusions:

• UAV-based remote sensing is a promising tool for objective and efficient phenotyping of soybean flooding tolerance.
• Lower flight altitudes (e.g., 20m) provide optimal data for estimating soybean flooding injury.
• This approach can accelerate breeding programs for flood-tolerant soybean cultivars.