基于光谱特征和颜色特征的油菜提取研究

Extraction for Oilseed Rape Based on Spectral Feature and Color Feature

大区域油菜空间分布的准确提取是油菜估产、食用油保障及农业管理的基础。花期油菜不仅光谱特征发生变化,其黄色花朵与同时期植被在视觉上的差异也相当显著。本文基于2016年油菜盛花期的湖北省GF-1 WFV影像,利用NGVI表征光谱特征,HSV变换后的H、S、V分量表征颜色特征,按NGVI、H、S、V顺序逐级确定油菜和非油菜分离阈值,实现油菜提取;对提取结果首先用混淆矩阵进行精度评定,并与支持向量机方法进行精度对比,然后用农业统计数据对油菜提取面积进行验证。基于本文方法提取的油菜总体精度为94.51%,Kappa系数为0.89,分别比支持向量机方法提高约4个百分点和0.1;与统计面积相比,省级尺度油菜提取面积相对误差为-14.14%,市级、县级尺度决定系数分别为0.837(n=17)、0.738(n=83)。此外,将本文方法应用到GF-2 PMS影像上,其结果与油菜参考图相比,油菜提取面积相对误差为-8.33%,空间一致度为91.67%。本研究方法简单有效,可以为大区域油菜制图提供一种全新、高效的解决方案。

Large scale management of spatial distribution of oilseed rape is essential for grain yield estimations,ensuring edible oil supply and sustainable agricultural management. Flowering period is the special growth stage of oilseed rape. Spectral feature of oilseed rape in this period changes largely.Furthermore,the sense of sight for oilseed rape also has a big difference against with other vegetation types during flowering period. Thus,spectral feature and color feature in the flowering stage can be set as the unique features for identifying oilseed rape as well as the basis of oilseed rape extraction. NGVI,a flowering-contained detecting indicator,was used to represent spectral feature of oilseed rape in the flowering period. H,S and V components were conducted as color feature of oilseed rape after processing colorimetric transformation from RGB color space to HSV color space. And then,the samples of oilseed rape and non-oilseed rape,which were interpreted on wide field view（ WFV） images from Gaofen satellite no. 1（ GF-1） combined Google Earth images and field investigation,were analyzed to determine the thresholds of NGVI,H,S,and V successively. Afterwards,oilseed rape in Hubei Province of China in 2016 was extracted based on GF-1 WFV images that obtained in full-flowering stage,which was evaluated by confusion matrix and compared with traditional support vector machine（ SVM） method.Meanwhile,the GF-1 WFV-estimated planting acreage of oilseed rape was validated against agricultural census data. As a result,the sample evaluation achieved 94. 51% of overall accuracy and 0. 89 of Kappa coefficient,which improved four percentage points and 0. 1 compared with SVM method,respectively.The result against statistical data had -14. 14% of relative error at provincial level as well as 0. 837（ n = 17） and 0. 738（ n = 83） of decision coefficients at municipal level and county level. Moreover,the method was applied on panchromatic and multi-spectral（ PMS） image from GF-2 and validated byreference oilseed rape map. The relative error of oilseed rape extraction was -8. 33% and spatial consistency was 91. 67%. Therefore,this study proposed a simple,effective and robust oilseed rape extraction strategy in large-regional scale based on satellite imagery of full-flowering period.