Combining texture and spectral feature values for rice plant detection using unmanned aerial vehicle (UAV) imagery

Xiang Zhou, Xiaoqing Xu, Xiaokang Ge, Yu Zhang, Xia Yao, Tao Cheng, Yan Zhu, Weixing Cao, Yongchao Tian *

National Engineering and Technology Center for Information Agriculture, Nanjing Agricultural University, Nanjing, Jiangsu 210095, China

Abstract:

Exploringa fast and accurate classification method for remotely sensed images is an important step in crop quantitative remote sensing (RS). The unmanned aerial vehicle (UAV) enables the acquisition of high spatiotemporal resolution imagery at a regional scale and provides a novel approach for crop RS. In this study, a multispectral (MS) camera, digital camera, and modified normalized difference vegetation index (modified NDVI) camera were mounted on a multi-rotor UAV to obtain images of the experimental fields at critical rice growth stages. A combination of texture features [mean and coefficient of variation (C.V)] and spectral features were introduced into the decision tree based on texture and spectra (TS-Decision Tree) for rice plants detected using these UAV images. Primarily, the image was classified as either pure rice plants, background, and the corresponding mixed class with the C.V value, and then the background of water or soil was distinguished from the mixed class by comparing the mean value and spectral feature. This classification method was tested with three types of images, and the result showed that TS-Decision Tree resulted in high classification accuracies of 93.53%, 91.25%, and 92.88% for the three types of imagery with the optimal spectral features of 800 nm (MS image), excess green vegetation index (ExG, Digital image) and near infrared (NIR-G-B image), respectively. TS-Decision Tree also achieved more stable and higher mapping accuracy than other classification methods with the texture feature such as Parallelepiped, Minimum Distance, Mahalanobis Distance, Maximum Likelihood, Support Vector Machine, and Neural Network methods. The results from NIR-G-B images at different altitudes showed that spatial resolution had an obvious impact on the classification accuracy. Classification accuracy was significantly higher in images acquired at 30 m (93.15%) and 50 m (96.19%) than those acquired at 100 m (81.81%). Thus, the TS-Decision Tree method proposed in this study shows a high accuracy and good applicability, and is a suitable technical approach for vegetation fraction mapping using unmanned aerial systems imagery.

Keywords: Rice, UAV imagery, Texture feature, Spectral feature, Plant detection, TS-Decision Tree