基于无人机可见光影像的绿色植被提取

在分析健康绿色植被光谱特性及无人机可见光影像典型地物各波段像元值差异的基础上，提出一种综合利用红、绿、蓝3个可见光波段信息的新型绿色植被指数——差异增强植被指数（DEVI）.利用该指数及其他8种常见可见光植被指数结合不同阈值方法提取研究区域绿色植被信息，并采用地表真实感兴趣区和基于SVM的监督分类方法进行精度量化评价.结果表明：由DEVI计算的植被指数灰度影像直方图具有良好双峰形态，可利用双峰直方图阈值法快速确定阈值，且阈值一般位于0.9~1之间；同时，DEVI提取精度明显优于其余8种植被指数，且采用双峰直方图阈值法时，总体精度为98.98%，Kappa系数为0.9791，相对误差为1/83.为验证DEVI是否具有良好的可适用性及可靠性，选取3种典型植被覆盖区域进行可行性验证，结果表明：利用DEVI可高精度提取建筑密集区域和植被零散分布区域的绿色植被信息，总体精度分别为98.42%和98.56%，Kappa系数分别为0.9610和0.9635，相对误差分别为1/125和1/91；而植被集中分布区域提取精度略低于上述2种典型区域，总体精度为97.40%，Kappa系数为0.9371，相对误差为1/53.因此，提出的差异增强植被指数——DEVI可以有效、高精度、低成本提取不同植被覆盖典型区域无人机可见光影像中的绿色植被信息，为陆地生态系统中的绿色植被监测研究提供一种可行性方法.

Green vegetation extraction based on visible light image of UAV

Using visible light images taken by unmanned aerial vehicle (UAV) as data source, a new green vegetation index named as Difference Enhanced Vegetation Index (DEVI) was proposed based on the analysis of healthy green vegetation spectral characteristics and the differences of pixel values among different bands of typical ground objects in visible light images of UAVs. DEVI utilized the information of red, green and blue visible bands, which can not only eliminate the interference caused by the difference of pixel values in a single band of green wave of different ground objects, but also enhance the characteristic that the reflectivity of green wave of green vegetation is greater than that of red and blue bands. This new index and 8 other common visible light vegetation indexes were used to extract the green vegetation by the threshold method in the study area, and then the support-vector machine (SVM)-based supervised classification method and the ground truth area of interest (ROIs) were used to evaluate the extraction accuracy. The results showed that the extraction accuracy of DEVI was significantly better than the other eight vegetation indexes. When the threshold method of image histogram visual detection was adopted, the overall accuracy was 98.98%, the Kappa coefficient was 0.9791, and the relative error was 1/83. Meanwhile, the gray image histogram of vegetation index calculated by DEVI had a good bimodal shape, which could quickly determine the threshold value, and the threshold value was generally located between 0.9 and 1. To verify whether DEVI has good applicability and reliability, this study chose three typical areas to conduct the feasibility verification analysis:area with high vegetation coverage, area with dense regions of buildings, and area with discretely distributed vegetation. The results showed that the green vegetation information in regions with dense buildings and discretely distributed vegetation could be extracted with high precisions by DEVI. The overall accuracy was 98.42% and 98.56%, the Kappa coefficient was 0.9610 and 0.9635, and the relative error was 1/125 and 1/91, respectively. However, the extraction accuracy in areas with high vegetation coverages was slightly less accurate with the overall accuracy of 97.40%, a Kappa coefficient of 0.9371, and a relative error of 1/53. Therefore, the new DEVI could extract the green vegetation information from UAV visible light images in typical vegetation covered areas in an effective, high-precision and low-cost way. Therefore, DEVI is a feasible method for the green vegetation monitoring research in terrestrial ecosystems.