Smartphone Deep Learning for Counting Sorghum Grains
Author Information
Author(s): Gustavo N. Santiago, Cisdeli Magalhaes Pedro H., Carcedo Ana J. P., Marziotte Lucia, Mayor Laura, Ciampitti Ignacio A.
Primary Institution: Kansas State University
Hypothesis
Can deep learning methods using smartphone imagery improve the accuracy of sorghum grain counting?
Conclusion
The study developed a framework that accurately estimates the number of grains in sorghum panicles using smartphone images, achieving a mean absolute percentage error of 17%.
Supporting Evidence
- The proposed framework achieved an average precision of 75% and 89% for panicle detection and segmentation, respectively.
- The Sorghum-Net model outperformed other models with a mean absolute percentage error of 17%.
- Using smartphone imagery allows for a cost-effective solution for high-throughput phenotyping.
Takeaway
This study shows that we can use smartphone pictures and smart computer programs to count how many grains are in sorghum plants, making it easier for farmers to know how much they can harvest.
Methodology
The study used smartphone images of sorghum panicles, trained deep learning models for detection and segmentation, and developed a linear model to estimate grain counts.
Limitations
The overlap between background and foreground panicles made manual labeling difficult, leading to potential false positives.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website