Li Tiansheng, Wang Haijiang, Cui Jing, Wang Weiju, Li Wenruiyu, Jiang Menghao, Shi Xiaoyan, Song Jianghui, Wang Jingang, Lv Xin, Zhang Lifu
College of Agriculture, Shihezi University, Shihezi, China.
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China.
Front Plant Sci. 2024 Mar 27;15:1333089. doi: 10.3389/fpls.2024.1333089. eCollection 2024.
Timely and accurate estimation of cotton seedling emergence rate is of great significance to cotton production. This study explored the feasibility of drone-based remote sensing in monitoring cotton seedling emergence. The visible and multispectral images of cotton seedlings with 2 - 4 leaves in 30 plots were synchronously obtained by drones. The acquired images included cotton seedlings, bare soil, mulching films, and PE drip tapes. After constructing 17 visible VIs and 14 multispectral VIs, three strategies were used to separate cotton seedlings from the images: (1) Otsu's thresholding was performed on each vegetation index (VI); (2) Key VIs were extracted based on results of (1), and the Otsu-intersection method and three machine learning methods were used to classify cotton seedlings, bare soil, mulching films, and PE drip tapes in the images; (3) Machine learning models were constructed using all VIs and validated. Finally, the models constructed based on two modeling strategies [Otsu-intersection (OI) and machine learning (Support Vector Machine (SVM), Random Forest (RF), and K-nearest neighbor (KNN)] showed a higher accuracy. Therefore, these models were selected to estimate cotton seedling emergence rate, and the estimates were compared with the manually measured emergence rate. The results showed that multispectral VIs, especially NDVI, RVI, SAVI, EVI2, OSAVI, and MCARI, had higher crop seedling extraction accuracy than visible VIs. After fusing all VIs or key VIs extracted based on Otsu's thresholding, the binary image purity was greatly improved. Among the fusion methods, the Key VIs-OI and All VIs-KNN methods yielded less noises and small errors, with a RMSE (root mean squared error) as low as 2.69% and a MAE (mean absolute error) as low as 2.15%. Therefore, fusing multiple VIs can increase crop image segmentation accuracy. This study provides a new method for rapidly monitoring crop seedling emergence rate in the field, which is of great significance for the development of modern agriculture.
及时、准确地估算棉花出苗率对棉花生产具有重要意义。本研究探讨了基于无人机遥感监测棉花出苗情况的可行性。通过无人机同步获取了30个地块中2 - 4片真叶棉花幼苗的可见光和多光谱图像。采集的图像包括棉花幼苗、裸土、地膜和PE滴灌带。构建17个可见光植被指数（VI）和14个多光谱植被指数后，采用三种策略从图像中分离棉花幼苗：（1）对每个植被指数进行大津阈值处理；（2）根据（1）的结果提取关键植被指数，并采用大津交叉法和三种机器学习方法对图像中的棉花幼苗、裸土、地膜和PE滴灌带进行分类；（3）使用所有植被指数构建机器学习模型并进行验证。最后，基于两种建模策略[大津交叉法（OI）和机器学习（支持向量机（SVM）、随机森林（RF）和K近邻（KNN））]构建的模型显示出更高的准确率。因此，选择这些模型来估算棉花出苗率，并将估算结果与人工测量的出苗率进行比较。结果表明，多光谱植被指数，尤其是归一化差异植被指数（NDVI）、比值植被指数（RVI）、土壤调整植被指数（SAVI）、增强型植被指数2（EVI2）、优化土壤调整植被指数（OSAVI）和修正型叶绿素吸收反射指数（MCARI），在作物幼苗提取精度上高于可见光植被指数。融合基于大津阈值提取的所有植被指数或关键植被指数后，二值图像纯度得到极大提高。在融合方法中，关键植被指数 - 大津交叉法（Key VIs - OI）和所有植被指数 - K近邻法（All VIs - KNN）产生的噪声较少且误差较小，均方根误差（RMSE）低至2.69%，平均绝对误差（MAE）低至2.15%。因此，融合多个植被指数可以提高作物图像分割精度。本研究为田间快速监测作物出苗率提供了一种新方法，对现代农业发展具有重要意义。
