基于无人机多光谱影像的完熟期玉米倒伏面积提取

由于土壤、地形、水分以及耕作方式等存在的时空变异性,致使灾后完熟期玉米地块存在4类作物形态,包括叶片呈绿色的未倒伏玉米、叶片淡黄的未倒伏玉米、叶片淡黄的倒伏玉米、黑色阴影区域。为进一步提高现有倒伏玉米面积提取方法的精度,该文以黑龙江省国营农场典型玉米倒伏地块为研究区,获取无人机多光谱数据,对比4类作物形态的光谱、植被指数以及纹理特征差异,经特征筛选后,首先面向倒伏玉米提取构建了5种典型特征组合。然后针对植被指数特征、光谱和纹理特征组合采用最大似然法分类,最后对提取结果的精度进行评价和分析。结果表明:反射光谱特征或植被指数特征无法准确区分4类作物形态,提取的倒伏玉米面积偏差较大;多类纹理特征法所得结果最优,4类典型作物形态的识别平均误差为9.82%,倒伏面积提取的误差为3.40%,Kappa系数为0.84。该研究延展了纹理特征在倒伏玉米面积提取中的应用并对完熟期倒伏玉米识别具有重要的借鉴意义。

Extraction of maize lodging area in mature period based on UAV multispectral image

Abstract: Lodging has been regard as one of the major destructive factors for crop quality and yield, resulting in an increasing need to develop cost-efficient and accurate methods for detecting crop lodging in a routine manner. Nowadays, rapid evolvement in unmanned aerial vehicle (UAV) and sensor technology has allowed high accurate and more accessible in monitoring crop development and health status with adequate temporal, spatial, and spectral resolutions. Compared with satellite and airborne photogrammetry, UAV with proper sensors offer a flexible, convenient, and cost-effective way to provide desired and customized observations on crop fields. Previous studies have extensively examined and verified the potential of UAV-based lodging recognition by leveraging photogrammetric algorithms, geospatial computing analysis, as well as pertinent agricultural expertise. As a substantive extension of previous published proceeding papers, this work presents a complete UAV-based survey methodology for monitoring lodging maize. Multispectral images of lodging mature maize in Youyi farm of Heilongjiang Province were collected to extract the lodging area. There were 4 crop forms in the research area: not-lodging maize with green leaves, not-lodging maize with yellowish leaves, lodging maize with yellowish leaves, and black shadows, based on the multispectral image. The 2 vegetation indexes and 8 co-occurrence measures texture features were calculated, and the feature sets of maize lodging area extraction was constructed on the basis of the above 2 kinds of predictors and spectral reflectivity features. 5 types of maize lodging identification feature sets were sifted, which included spectral feature set, normalized difference vegetation index (NDVI) feature set, red edge normalized difference vegetation index (NDVIR-edge) feature set, single-class texture feature set and multi-class texture feature set. The maximum likelihood method was used to identify maize lodging for all feature sets. Finally, we analyzed the classification error of 4 crop morphology, extraction error and Kappa coefficient of lodging area under different features. The results showed that maize lodging area extracted by spectral feature set and NDVI feature set was larger than measured lodging area, which mainly because the wrong classification of not-lodging B pixels into lodging pixels, while the main reason for the inaccurate maize lodging area obtained by NDVIR-edge feature was that the not-lodging B maize and the not-lodging maize affected by edge effect were classified into lodging. Extraction area of lodging maize by single texture feature set was smaller because some of the not-lodging B maize pixels were classified as lodging pixels, but more lodging pixels was misclassified as not-lodging pixels. Single and multi-class texture feature sets could remove the shadow of blade gap well with appropriate texture filtering window selected, but multi-class feature set had higher extraction accuracy. It was difficult to distinguish the lodging maize from the not-lodging maize with yellowish leaves in mature period, and there was no significant difference in spectral reflectance feature between the 2 crop morphology. Therefore, when we identified these 2 types of crop morphology, a large number of misclassification pixels would be generated. The multi-class texture features extracted from UAV multispectral images could accurately extract maize lodging area. The average error of 4 crop morphology was 9.82%, the extraction error of lodging area was 3.40%, and the Kappa coefficient was 0.84.