基于多源遥感数据的锡尔河中下游农田土壤水分反演

机器学习结合多源遥感数据反演土壤水分含量（SMC）是目前SMC研究的热点,因较少考虑温度、蒸散等重要SMC影响因子,反演结果存在一定的不确定性。利用Sentinel-1影像、MODIS产品和SRTM数据,提取雷达后向散射系数等32个SMC影响因子,经相关分析选择27个显著的SMC影响因子（P<0.05）作为反演因子,并设计三组因子组合。这三组因子组合分别与随机森林、支持向量回归、BP神经网络三种机器学习方法结合,发现基于随机森林结合所有因子的方案,其SMC反演精度最高,该组合均方根误差RMSE为0.039 m³/m³,将该方案被用于反演2017年生长季锡尔河流域中下游平原区农田SMC。结果表明：从上部至下部SMC总体呈逐渐增加的态势,但存在显著时空差异,春季和秋季SMC较高而夏季较低。SMC差异主要由土壤质地、热量条件和地表植被状况差异引起。春季平原区下部农田SMC要高于上部,SMC的主控因子是土壤质地和地表植被状况;在夏季,土壤水分的主控因子是热量条件,农田灌溉弥补了热量条件差异对土壤水分的影响,导致空间上平原上部和下部土壤SMC空间差异不显著;秋季SMC的主控因子植被状况抵消地表温度和土壤质地差异对SMC的影响,使得秋季SMC空间差异不显著。本文采用的研究方法在一定程度上克服了因考虑SMC影响因子不足而获取更高SMC精度的限制。

Inversion of soil moisture content in the farmland in middle and lower reaches of Syr Darya River Basin based on multi-source remotely sensed data

The use of machine learning method to estimate Soil Moisture Content (SMC) from multi-source remotely sensed data is a hot topic in the SMC inversion research. However, taking no account of the important variables of SMC in the ML method makes the SMC results uncertain. The Sentinel-1 and MODIS image products and the STRM data were obtained and used for extracting 32 SMC variables, such as backscattering coefficient, vegetation index, surface temperature and evapotranspiration. A total of 27 significant (P<0.05) SMC variables were selected as input parameters referring to the correlation analysis result, and the input parameters were assigned to 3 groups. Random forest, Support vector regression and Back Propagation Neural Network were tested with 3 groups parameters. The Random forest with the group with all input parameters showed the best estimation accuracy, with the RMSE being 0.039 m³/m³, and it was used for the inversion of SMC in the farmland in the middle and lower reaches of Syr Darya River Basin during the growing season of 2017. The retrieved SMC gradually increased in the middle to the lower reaches during the growing season, but there were significant temporal and spatial differences: SMC in spring and autumn was higher than that in summer. These differences were mainly caused by seasonal or spatial differences in soil texture, heat conditions (temperature) and vegetation cover. In spring, SMC in the lower part of the plain is higher than that in the upper part, and the main SMC controlling factors were soil texture and vegetation cover. In summer, the main SMC controlling factors were heat condition. Irrigation compensated for the influence of heat condition difference, resulting in no significant spatial difference of SMC between upper and lower parts of the plain. The main SMC controlling factors in autumn were soil texture and heat conditions, the influence of surface temperature compensated for the influence of soil texture on SMC, as a result, there was no significant spatial difference of SMC in autumn. With regard to overcoming the limitation of taking no account of the important variables in estimating SMC, the research method adopted in this study improves the retrieved SMC accuracy to a large extent.