Estimation of Off-Target Dicamba Damage on Soybean Using UAV Imagery and Deep Learning

Area of Science:

• Agricultural Science
• Plant Breeding
• Remote Sensing

Background:

• Dicamba herbicide drift causes significant yield losses in non-tolerant crops.
• Developing non-genetically engineered dicamba-tolerant (DT) soybeans is crucial for agricultural sustainability.
• Efficient phenotyping is essential for accelerating crop breeding programs.

Purpose of the Study:

• To evaluate the efficacy of unmanned aerial vehicle (UAV) imagery and deep learning for quantifying off-target dicamba damage in diverse soybean genotypes.
• To develop and validate deep learning models for high-throughput phenotyping of dicamba-induced crop damage.

Main Methods:

• Collected UAV RGB imagery of 463 soybean genotypes exposed to off-target dicamba over two years.
• Segmented soybean plots and applied deep learning models (DenseNet121, ResNet50, VGG16, Xception) to classify damage levels.
• Assessed model performance based on classification accuracy and confidence intervals.

Main Results:

• The DenseNet121 model achieved the highest classification accuracy of 82% (95% CI: 79-84%) for dicamba damage.
• No extreme misclassifications occurred between tolerant and susceptible soybean genotypes.
• The developed method demonstrated high potential for identifying soybean genotypes with extreme tolerance phenotypes.

Conclusions:

• UAV imagery combined with deep learning offers a powerful tool for high-throughput phenotyping of dicamba damage in soybeans.
• This approach can significantly enhance the efficiency of breeding programs selecting for dicamba-tolerant soybean varieties.
• The findings support the development of improved soybean cultivars resilient to herbicide drift.