Sun Sashuang, Liang Ning, Zuo Zhiyu, Parsons David, Morel Julien, Shi Jiang, Wang Zhao, Luo Letan, Zhao Lin, Fang Hui, He Yong, Zhou Zhenjiang
College of Biosystems Engineering and Food Science, Zhejiang University, Hangzhou, China.
School of Agricultural Engineering, Jiangsu University, Zhenjiang, China.
Front Plant Sci. 2021 Feb 11;12:622429. doi: 10.3389/fpls.2021.622429. eCollection 2021.
This study aims to provide an effective image analysis method for clover detection and botanical composition (BC) estimation in clover-grass mixture fields. Three transfer learning methods, namely, fine-tuned DeepLab V3+, SegNet, and fully convolutional network-8s (FCN-8s), were utilized to detect clover fractions (on an area basis). The detected clover fraction ( ), together with auxiliary variables, viz., measured clover height ( ) and grass height ( ), were used to build multiple linear regression (MLR) and back propagation neural network (BPNN) models for BC estimation. A total of 347 clover-grass images were used to build the estimation model on clover fraction and BC. Of the 347 samples, 226 images were augmented to 904 images for training, 25 were selected for validation, and the remaining 96 samples were used as an independent dataset for testing. Testing results showed that the intersection-over-union () values based on the DeepLab V3+, SegNet, and FCN-8s were 0.73, 0.57, and 0.60, respectively. The root mean square error () values for the three transfer learning methods were 8.5, 10.6, and 10.0%. Subsequently, models based on BPNN and MLR were built to estimate BC, by using either only or , grass height, and clover height all together. Results showed that BPNN was generally superior to MLR in terms of estimating BC. The BPNN model only using had a of 8.7%. In contrast, the BPNN model using all three variables ( , , and ) as inputs had an of 6.6%, implying that DeepLab V3+ together with BPNN can provide good estimation of BC and can offer a promising method for improving forage management.
本研究旨在为三叶草-禾本科牧草混合田中的三叶草检测和植物组成（BC）估计提供一种有效的图像分析方法。利用三种迁移学习方法，即微调的深度卷积神经网络（DeepLab）V3+、SegNet和全卷积网络8s（FCN-8s）来检测三叶草比例（基于面积）。检测到的三叶草比例（ ），连同辅助变量，即实测的三叶草高度（ ）和禾本科牧草高度（ ），用于建立多元线性回归（MLR）和反向传播神经网络（BPNN）模型，以估计BC。总共347张三叶草-禾本科牧草图像用于建立三叶草比例和BC的估计模型。在这347个样本中，226张图像扩充为904张用于训练，25张用于验证，其余96个样本用作独立数据集进行测试。测试结果表明，基于DeepLab V3+、SegNet和FCN-8s的交并比（IoU）值分别为0.73、0.57和0.60。三种迁移学习方法的均方根误差（RMSE）值分别为8.5%、10.6%和10.0%。随后，建立基于BPNN和MLR的模型来估计BC，分别仅使用 或同时使用 、禾本科牧草高度和三叶草高度。结果表明，在估计BC方面，BPNN总体上优于MLR。仅使用 的BPNN模型的RMSE为8.7%。相比之下，使用所有三个变量（ 、 和 ）作为输入的BPNN模型的RMSE为6.6%，这意味着DeepLab V3+与BPNN相结合可以很好地估计BC，并可为改善饲料管理提供一种有前景的方法。
