中国南方山区油菜识别的遥感影像选择策略研究——以云南罗平县为例

REMOTE SENSING IMAGE SELECTION STRATEGY FOR RAPE RECOGNITION IN MOUNTAINOUS AREAS OF SOUTH CHINA——TAKING LUOPING COUNTY, YUNNAN PROVINCE AS AN EXAMPLE

关键词:  油菜识别精度  空间分辨率  时间特征  雷达特征  影像选择策略

Key words:  identification accuracy of rape  spatial resolution  time characteristics  radar characteristics  image selection strategy

摘要:

目的 油菜的精确识别对我国油菜估产和食用油供应至关重要。在多云多雨、地块破碎的南方山区，遥感影像数据源的选择直接影响油菜种植信息提取的精确度。为了进行业务化油菜精确测量，需要明确当前各类遥感数据在油菜识别中的适用性，探究各类遥感特征在油菜识别过程的精度贡献，以总结不同场景下南方山区油菜提取的影像选择策略。方法 文章利用GF-2 PMS、GF-6 PMS、Sentinel-2光学影像数据和Sentinel-1雷达数据，结合支持向量机分类算法对云南省罗平县研究区内的山区油菜进行识别提取和精度分析。在油菜的关键时间窗口（油菜开花期）和非关键时间窗口分别设计方案从空间分辨率特征、时间特征和时序雷达特征探究油菜的识别精度贡献，分析影响精度的原因。结果 （1）油菜破碎区对空间分辨率的降低更加敏感，在分辨率优于16 m的情况下才能有效识别油菜。（2）从时间特征来看，单时相关键时间窗口光学影像（制图精度99.69%，用户精度96.24%）>时间序列光学影像（制图精度98.59%，用户精度93.61%）>单时相非关键时间窗口光学影像（制图精度91.90%，用户精度87.99%）。（3）在雷达特征上，时间序列VV+VH极化特征雷达影像优于任一单极化、单时相雷达特征影像，能够在云雨天气替代光学影像达到理想的分类效果（制图精度95.60%，用户精度92.04%）。结论 研究结果能够针对复杂的影像数据获取情况提供合适的使用策略，可为南方山区业务化油菜测量和影像方案选择提供一定的参考。

Abstract:

The precise identification of rapeseed is crucial for the yield estimation and edible oil supply in China. In the southern mountainous areas with cloudy and rainy climates and fragmented fields, the choice of remote sensing image data sources directly affects the accuracy of rapeseed planting information extraction. To carry out professional precise measurement of rapeseed, it is necessary to clarify the applicability of various types of remote sensing data in rapeseed identification, explore the contribution of various remote sensing features to the accuracy of rapeseed identification process, and summarize the image selection strategy for extracting rapeseed in different scenarios in southern mountainous areas. This study used GF-2 PMS, GF-6 PMS, Sentinel-2 optical image data, and Sentinel-1 radar data, combined with support vector machine classification algorithm to identify and extract rapeseed in the mountainous area of Luoping county, Yunnan province, and analyze its accuracy. In the key time window (rapeseed flowering period) and non-key time window of rapeseed, design schemes were proposed to explore the contribution of rapeseed identification accuracy from spatial resolution features, time features, and time-series radar features, and analyze the reasons affecting the accuracy. The results were showed as follows. (1) The fragmented area of rapeseed was more sensitive to the reduction of spatial resolution, and could rapeseed be effectively recognized only when the resolution was better than 16 m. (2) From the standpoint of temporal features, single-time key time window optical images (Producer's Accuracy 99.69%, User's Accuracy 96.24) > time-series optical images (Producer's Accuracy: 98.59, User's Accuracy: 93.61) > single-time non-key time window optical images (Producer's Accuracy: 91.90, User's Accuracy: 87.99). (3) With respect to radar features, time-series VV+VH polarization feature radar images were superior to any single polarization, single-time radar feature images. These could substitute for optical images in overcast or rainy conditions to achieve optimal classification results (Producer's Accuracy: 95.60, User's Accuracy: 92.04). In summary, the findings from this study can offer suitable strategies for complex image data acquisition scenarios and serve as a reference for selecting image solutions for professional rapeseed measurement and application in southern mountainous areas.