基于Sentinel-2超分辨率影像的干旱区水体提取方法

针对干旱区复杂环境下水体光谱特性空间差异大、水体提取方法适用性差的问题,本研究基于Sentinel-2卫星多光谱数据,通过超分辨率算法重建10 m空间分辨率多光谱影像,将短波红外(Short-wave infrared, SWIR)重建波段、近红外(Near-infrared, NIR)重建波段作为水体识别特征波段,在此基础上采用超像素分割算法识别水体像元,基于24种光谱指数、支持向量机(Support vector machine, SVM)、神经网络(Neural network, NN)、K-means共构建60种水体提取方法,采用总体精度(Overall accuracy, OA)、准确率(Precision)、F1值、马修斯相关系数(Matthews correlation coefficient, MCC)等水体提取精度指标进行综合评价,以黑河流域为典型研究区,确定干旱区最佳水体提取方法。结果表明,基于Sentinel-2绿色波段(中心波长为560 nm)与超分辨率重建短波红外波段(中心波长为1 610 nm)构建的改进的归一化水体指数方法,显著增强水体提取时对干旱区细小水...

Water Body Extraction Method in Arid Area Based on Sentinel-2 Super-resolution Images

Aiming at the problems of large spatial differences in the spectral characteristics of water bodies in the complex environment of arid zones and the poor applicability of water body extraction methods, based on the multispectral data of Sentinel-2 satellite, 10m spatial resolution multispectral images were reconstructed by super-resolution algorithm. The short-wave infrared (SWIR) reconstruction band and the near-infrared (NIR) reconstruction band were used as the feature bands for water body identification, on the basis of which the super-pixel segmentation algorithm was used to determine the water body image elements, and a total of 60 water body extraction methods were constructed based on 24 kinds of spectral indices, support vector machine (SVM), neural network (NN) and K-means. Overall accuracy (OA), precision, F1-score, Matthews correlation coefficient (MCC) and other water body extraction accuracy indicators were used as for comprehensive evaluation, to determine the best water body extraction method in the Heihe Basin. The Heihe Basin was taken as typical study area to determine the best water body extraction method in arid areas. The results showed that the improved normalized water body index method constructed based on Sentinel-2 green band (center wavelength of 560nm) and super-resolution reconstruction of the short-wave infrared band (center wavelength of 1.610nm) significantly enhanced the ability to identify the fine water bodies, shadows, and cloud elements in the arid zone during the extraction of the water body. The overall accuracy of water extraction was 99.81%, the accuracy was 92.04%, the F1-score was 88.02%, and the G-mean and Mathews correlation coefficient were both greater than 0.88, which was better than other methods. The research results can quickly and accurately extract water bodies in arid zones and provide theoretical support for the application field of water bodies in arid zones.