多源哨兵数据解译农田区土壤湿度算法研究

Research on algorithms for interpreting soil moisture in farmland by multi-source Sentinel data

针对土壤湿度受植被影响较大且缺少实测地表粗糙度的问题,该文结合Sentinel-1雷达卫星和Sentinel-2光学卫星多源遥感数据,使用水云模型去除植被影响,并通过高级积分方程模型(AIEM)和Dobson介电模型建立查找表获取每个采样点的有效粗糙度,利用经验方程多元回归、广义神经网络(GRNN)模型、随机森林回归模型和一种顾及植被影响的变化检测方法定量反演稀疏植被覆盖下的农田区地表土壤湿度。实验结果表明,随机森林回归模型的最优特征参数组合(垂直极化雷达后向散射系数、高程、局部入射角、增强型植被指数、有效粗糙度)反演精度最高,测试样本的相关系数为0.936,偏差和均方根误差分别达到了0.011 cm^(3)/cm^(3)和0.020 cm^(3)/cm^(3),而变化检测方法对于研究区土壤湿度存在整体高估的情况。

Aiming at the problem that soil moisture is greatly affected by vegetation and there is no measured surface roughness,based on Sentinel-1 radar satellite and Sentinel-2 optical satellite multi-source remote sensing data,water cloud model was used to remove vegetation influence and a lookup table was established to obtain the effective roughness of each sampling point by advance integrated equation model(AIEM)and Dobson dielectric model in this paper.Then,empirical equation multiple regression,generalized regression neural network(GRNN)model,random forest regression model and a change detection method considering vegetation influence were used to quantitatively retrieve surface soil moisture on sparsely vegetated farmland.Experimental results showed that the optimal combination of characteristic parameters of random forest regression model(VV polarization radar backscattering coefficient,elevation,local incident Angle,enhanced vegetation index,effective roughness)had the highest retrieval accuracy.The correlation coefficient of test samples was 0.936 and the deviation and root mean square error reached 0.011 cm^(3)/cm^(3)and 0.020 cm^(3)/cm^(3)respectively,while the change detection method overestimated the soil moisture in the study area.