基于Sentinel-1和Sentinel-2的不同物候期农作物识别研究

Crop Identification Study Based on Sentinel-1 and Sentinel-2 for Different Phenological Periods

为减少农作物提取过程中受光学数据成像质量的影响,基于Google Earth Engine平台,采用Sentinel-1和Sentinel-2数据,分别对小麦越冬期、返青期、孕穗期、成熟期四个物候期进行小麦和油菜的识别。使用随机森林方法对构建的光谱特征、植被指数特征、红边指数特征、纹理特征和极化特征共34个特征进行优选,构建特征集;并对比最小距离、决策树、支持向量机、随机森林四种分类器在四个物候期的识别结果,确定最优的分类器;同时还验证了极化特征在四个物候期对识别结果的影响。研究结果表明,在四个物候期中最优的分类器均为随机森林,其中识别精度从高到低的物候期分别为小麦的孕穗期、成熟期、返青期、越冬期,总体精度(OA)分别为92.91%、91.93%、90.24%、87.69%,Kappa系数分别为91.00%、89.92%、87.61%、84.53%。在四个物候期中加入极化特征均能提高识别的精度,其中在小麦的返青期和成熟期更为明显。

In order to reduce the influence of optical data imaging quality in the process of crop extraction,Senti-nel-1 and Sentinel-2 data were used based on Google Earth Engine platform to identify wheat and rape respective-ly in four phenological periods:overwintering period,regreening period,booting period and maturity period.A feature set is constructed by using a random forest method to preferentially select the optimal features from a set of 34 features consisting of spectral features,vegetation index features,red edge index features,texture features and polarisation features.It also compared the recognition results of four classifiers,namely,minimum distance,decision tree,support vector machine and random forest,in the four phenological periods to determine the opti-mal classifier;and also verified the influence of polarization features on the recognition results in the four pheno-logical periods.The results showed that the optimal classifier was random forest in all four phenological periods;the phenological periods with the highest to lowest recognition accuracy were wheat booting,maturity,regreen-ing and overwintering,with OA of 92.91%,91.93%,90.24%and 87.69%,and Kappa coefficients of 91.00%,89.92%,87.61%and 84.53%;The inclusion of polarisation features in each of the four phenological periods im-proved the accuracy of identification,more so in the two phenological periods of wheat,greening and maturity.