以Sentinel-2A和GF-1 WFV为数据源,分别采用最邻近法(面向对象)及BP神经网络法(基于像元),提取兴化市油菜种植区,并对提取结果进行精度验证;同时,结合官方数据,比较各提取组合类型(数据+方法)提取的油菜种植面积相对误差.通过此,探讨多源中高空间分辨率遥感数据中,油菜作物的提取效果差异性及适用的提取方法,并对研究区油菜种植空间格局特征进行分析.结果表明:针对Sentinel-2A和GF-1 WF数据,最邻近法及BP神经网络法提取效果均较好,各提取结果均显示油菜种植区在缸顾乡、周奋乡、垛田镇等西部区域主要呈集中连片分布,其他区域呈零星状分布.相对于基于像元的分类法,面向对象分类法在精度评价中的各参数表现更佳,并能较为有效地避免复杂地物类型区像元错分及漏分问题.针对同一数据,采用最邻近法所提取Sentinel-2A数据的生产者精度、用户精度以及油菜面积精度比BP神经网络法分别多3.22%、0.43%、6.24%,采用最邻近法所提取GF-1 WFV的生产者精度、用户精度和油菜面积精度比BP神经网络法高3.74%、0.10%、9.58%.针对同一方法,由于Sentinel-2A数据具更高的空间分辨率及更丰富的光谱信息,以上2种方法提取该数据的精度均高于GF-1 WFV数据,Sentinel-2A数据更适用于地物结构复杂,地块细碎的小尺度地区的作物信息提取.
To explore the difference of rapeseed extraction effect in the data characteristics of multi-source medium-high spatial resolution remote sensing data and the suitable extraction methods,based on Sentinel-2A and GF-1 WFV remote sensing data, the planting area of rapeseed in Xinhua city was extracted by using the nearest neighbor method and the BP neural network method in this paper;the confusion matrix was constructed based on sample points to verify the accuracy of classification. At the same time, combined with the official data and compared the relative error of Oilseed Rape planting area extracted by four extraction combination types(data +method). In addition, the spatial distribution of rapeseed extracted would be analyzed. Results showed that the extraction effects of the four extraction methods were all better. The rapeseed planting areas were mainly concentrated and contiguous in the western regions such as Canggu Township, Zhoufen Township and Putian Town, and the distribution in other areas were scattered. The Object-oriented classification method was better than pixel-based classification method in each parameter of accuracy evaluation, and it was more suitable for avoiding misclassification and leakage of mixed pixels. For the same data, the producer accuracy, user accuracy and rapeseed area accuracy of Sentinel-2A data extracted by the nearest neighbor method were 3.22%, 0.43% and 6.24% higher than that of BP neural network method, respectively, the producer accuracy, user accuracy and rape area accuracy of GF-1 WFV data extracted by nearest neighbor method were 3.74%, 0.10% and 9.58% higher than those of BP neural network method. For the same methods, the accuracy of Sentinel-2 data extracted by the two methods mentioned above was higher than that of GF-1 WFV because of it had higher spatial resolution and richer spectral information. So, the Sentinel-2A data was more suitable for crop information extraction in small-scale areas with complex terrain structure and fragmented plots.
Journal of Yunnan University(Natural Sciences Edition)