2020
In-field soybean seed pod analysis on harvest stocks using 3D imaging and machine learning
Home
Contributor/Checkoff:
Iowa Soybean Research Center
Category:
Sustainable Production
Keywords:
Data ManagementDrone/UAS
Parent Project:
This is the first year of this project.
Lead Principal Investigator:
Lie Tang, Iowa State University
Co-Principal Investigators:
Project Code:
GR-023053-00004
Contributing Organization (Checkoff):
Iowa Soybean Association
$30,000
Institution Funded:
Iowa State University
$30,000
Brief Project Summary:

Soybean seed pods directly contribute to the yield and represent an important trait for soybean breeding. Traits such as total pods, seeds per pod and the corresponding pod grouping, and the distribution of pods over the plant are all of great interest to soybean breeders and plant scientists, but have been difficult collect in an automated and high-throughput fashion, particularly under field conditions. With the advancement of 3D sensing technologies and the deep convolutional neural networks, the Agricultural Robotics lab at Iowa State University has made some breakthroughs in field-based plant phenotyping for plants. This project explores the great potential to extend these technologies and innovations into field-based soybean plant phenotyping.

Key Benefactors:
farmers, agronomists, Extension agents, soybean breeders, seed companies

Information And Results
Final Project Results

Updated October 5, 2023:

View uploaded report PDF file

In this project, we proposed an automated soybean seed and pod counting system consisting of a robotic platform and a set of deep learning based 3D point cloud processing algorithms for high throughput operations using images captured from two sides of the soybean plant. The results demonstrate that the proposed soybean pod and seed counting methods produced better accuracies than counting them using images captured from only one side of soybean rows. The proposed system can greatly reduce human effort. In the future, the counting and classification accuracies of the proposed system can be further improved by using more image samples to train the deep learning model as some highly overlapping pods were not detected. Besides, the accuracy of pod identification could be improved by combining multiple features like distance, inclination angle. Also, when there were overexposures, the quality of images was decreased, causing degraded 3D reconstruction. Improving the illumination uniformity of the strobe lights will alleviate this problem.

We presented our work on this project at the 2022 and 2023 ASABE conferences.

1. Liu, X., L. Xiang, L. Tang. 2022. In-field soybean seed pod phenotyping on harvest stocks using 3D imaging and deep learning. 2022 ASABE Annual International Meeting. Houston, TX, July 17-20, 2022. Paper No. 2201222.

2. Liu, X., L. Xiang, L. Tang, A. Raj, N. Butler. 2023. In-field soybean seed pod phenotyping on harvest stocks using 3D imaging and deep learning. 2023 ASABE Annual International Meeting. Omaha, NL. July 9-12, 2023. Paper No. 2301517.

The United Soybean Research Retention policy will display final reports with the project once completed but working files will be purged after three years. And financial information after seven years. All pertinent information is in the final report or if you want more information, please contact the project lead at your state soybean organization or principal investigator listed on the project.