一种基于多源多时相遥感信息的城市水稻田分布提取方法

A Method of Extracting Urban Rice Information by Combining Multi-source and Multi-temporal Remote Sensing Images

农作物空间监测与提取是城市群区域农业生产与城市发展的重要内容。尽管遥感技术在农业生产中已得到了广泛应用,但长三角地区水稻田监测受多云多雨条件限制,难以获取连续的高质量长时间序列多光谱影像,如何利用多时相多光谱影像及雷达影像高效地提取水稻田空间信息是一个亟须解决的问题。该文提出了一种融合多源多时相遥感影像的水稻田提取方法,以南京市江宁区为研究区域,以GF-1号卫星多光谱影像以及Sentinel-1雷达影像为数据源,选取覆盖水稻生长周期的3个时相雷达影像和3个时相多光谱影像,分别构建多光谱影像特征集(多波段反射率、NDVI、NDWI、EVI等)和雷达影像的反向散射系数特征集,综合实地调查数据选取训练样本,利用粒子群优化的SVM分类器进行分类。结果表明,在多光谱所有特征信息复合中表现最佳的多波段反射率与NDWI、WDRVI组合,再与雷达影像特征组合后,水稻田分类用户精度最高,达到93.42%,平均精度、总体精度和Kappa系数也分别达94.52%、94.31%和0.93。该方法能有效地提取水稻田空间分布信息,同时可为其他作物分类提供参考。

spatial monitoring and extraction of crops is an important part of agricultural production and urban development under the background of urban agriculture.The current remote sensing technology has been widely used in agricultural production,however,it is difficult to obtain continuous high-quality long-term series images due to the limits restrained by cloudy and rainy conditions,so that how to use a short time series image data to extract the paddy field spatial information effectively is a problem that needed to be solved.In this paper,a method of extracting paddy field by combining multi-source and multi-temporal remote sensing images was developed.The multi-temporal GF-1 and Sentinel-1 radar images of Jiangning District in Nanjing were used as the study data.Then the datasets for the rice extraction were constructed with two parts covering the growth cycle of rice,one is the multiband reflectance （MBR） with their derivative features including the normalized difference vegetation index （NDVI）,normalized difference water index （NDWI）,enhanced vegetation index （EVI）,wide dynamic range vegetation index （WDRVI） and perpendicular vegetation index （PVI） of GF-1 images and the other is the backscattering coefficient and their difference index （BCDI） of Sentinel-1 radar images.Besides,the particle swarm optimization SVM classification approach （PSO-SVM） was introduced to extract paddy field and a land survey data was used to validate the classification accuracy.The results show that the combination of MBR,NDWI,WDRVI and BCDI is the best performance in which the rice extraction accuracy reach 93.42%,the average accuracy,overall accuracy and Kappa coefficient reach 94.52%,94.31% and 0.93 respectively.