基于Sentinel多源数据的农田地表土壤水分反演模型构建与验证

土壤水分是影响水文、生态和气候等环境过程的重要参数,而微波遥感是农田地表土壤水分测量的重要手段之一。针对微波遥感反演农田地表土壤水分受植被覆盖影响较大的问题,该文基于Sentinel-1和Sentinel-2多源遥感数据,利用Oh模型、支持向量回归(support vector regression,SVR)和广义神经网络(generalized regression neural Network,GRNN)模型对土壤水分进行定量反演,以减小植被影响,提高反演精度。结果表明:通过水云模型去除植被影响后的Oh模型反演精度有所提高。加入不同植被指数的SVR和GRNN模型的反演效果总体优于Oh模型,基于SVR模型的多特征参数组合(双极化雷达后向散射系数、海拔高度、局部入射角、修改型土壤调整植被指数)反演效果最优,其测试集相关系数和均方根误差分别达到了0.903和0.015 cm~3/cm~3,为利用多源遥感数据反演农田地表土壤水分提供了参考。

Construction and validation of soil moisture retrieval model in farmland based on Sentinel multi-source data

As an important component of the earth ecosystem, soil moisture is of great significance in the fields of crop growth monitoring, crop yield estimation, variable irrigation and other related agricultural applications. With the rapid development of the technology and theory of microwave remote sensing, soil moisture retrieval with remote sensing data has been widely used at home and abroad. The multi-source remote sensing data used in this study was acquired from Sentinel-1 radar and Sentinel-2 optical satellites which belong to ESA''s Sentinel series and there are great advantages in space, time and data registration in monitoring soil moisture. The study area is located in Yangling Demonstration Zone, Shanxi Province and 45 sampling sites were selected and measured to validate the soil moisture retrieval model. Firstly, to deal with the problem that soil moisture retrieval was greatly affected by surface vegetation covers, this study applied Oh model to retrieve soil moisture after removing the influence of vegetation by water cloud model. Secondly, taking the great advantages of machine learning algorithms into account, the study selected support vector regression (SVR) and generalized regression neural network (GRNN) models to retrieve soil moisture, and the models were constructed with different combinations of characteristic parameters including VH polarization radar backward scattering coefficient and VV polarization radar backward scattering coefficient altitude (H0), local incident angle (LIA) which were calculated out with Sentinel-1 radar remote sensing data and vegetation indexes (normalized difference vegetation index, NDVI; modified soil adjusted vegetation index, MSAVI and difference vegetation index, DVI) which were calculated out with Sentinel-2 optical remote sensing data. Finally, this study defined the equivalent number of occurrences to evaluate the quantitative influence of each characteristic parameter because different parameters had different effect on farmland soil moisture retrieval. The results showed that the soil moisture retrieval accuracy of Oh model was increased after removing vegetation influence by water cloud model. The retrieval accuracies of SVR and GRNN models with MSAVI and NDVI were higher than that of Oh model. The optimal input combination of SVR model composed of five characteristic parameters, including VH polarization radar backward scattering coefficient, VV polarization radar backward scattering coefficient, H0, LIA, and MSAVI had the best retrieval accuracy with correlation coefficient of 0.903 and root mean square error of 0.014cm3/cm3 respectively. The optimal SVR model was used to retrieve the soil moisture in study area and the results were consistent with local rainfall events. The equivalent numbers of occurrences of characteristic parameters from high to low were VH polarization radar backward scattering coefficient, H0, VV polarization radar backward scattering coefficient, LIA, MSAVI, NDVI, DVI. For radar backward scattering coefficients from different polarized channel, VH polarization radar backward scattering coefficient is more sensitive to soil moisture than VV polarization radar backward scattering coefficient. Among the three vegetation indexes, the counting results indicated MSAVI had the strongest correlation with soil moisture content, followed by NDVI and DVI was the weakest. The experimental results showed that the fusion of radar and optical data had great potential in soil moisture retrieval in farmlands. The performances of the constructed model in other farmland types would be further investigated in the future.