Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.
Evaluation of the performance of machine learning methods in soybean segmentation for image-based high-throughput phenotyping in greenhouse
Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org
Citation: 2018 ASABE Annual International Meeting 1801362.(doi:10.13031/aim.201801362)
Authors: Jing Zhou, Xiuqing Fu, Shuiqin Zhou, Jianfeng Zhou
Keywords: Plant segmentation, supervised learning, unsupervised learning, soybean, high-throughput phenotyping.
Abstract. Image-based plant phenotyping is a growing application for the identification of phenotypes, stress tolerance classification and yield prediction in soybean. A key task in the pipeline of high-throughput phenotyping is the segmentation of individual plants from an orthomosaic image with a large number of plants. Machine learning is a promising approach as its strong ability in extraction of multiple-layer details from images and successful applications in plant leaf segmentation using 2D images. The aim of this paper was to evaluate the performance of three machine learning techniques, i.e. boosting, Support Vector Machine and K-means clustering, in segmentation of plants from the 3D imaging data of 75 plants. Sequential images collected by an image-based high-throughput phenotyping platform on 75 soybean plants at two different growing stages (non-overlapped and overlapped) were used to reconstruct 3D dense point cloud using the Structure from Motion (SfM) method. Ten features including location coordinates (x, y and z), colors (Red, Green, and Blue in RGB color space, hue and saturation in HSV color space and TGI greenness) and HOG descriptors were used in segmentation of background and individual plants. The misclassification error rate was evaluated to compare the performance of the three methods. Results showed that K-means had the least misclassification error rates (0.36% and 0.20%) for background segmentation and non-overlapped plants segmentation. After using HOG descriptor as a training set, the least misclassification error rate for overlapped plants was 2.57% using SVM.
(Download PDF) (Export to EndNotes)