基于2D Triple-Collocation和机器学习方法的多源遥感土壤湿度融合与降尺度研究  

Fusion and Downscaling of Multi-source Remote Sensing Soil Moisture based on 2D Triple-Collocation and Machine Learning Methods

作　　者：刘兴文 李巧玲[1] 宋淑红 LIU Xingwen;LI Qiaoling;SONG Shuhong(College of Hydrology and Water Resources,Hohai University,Nanjing 210098,China;Hydrology and Water Resources Survey Center of Shanxi Province,Xi’an 710068,China)

机构地区：[1]河海大学水文水资源学院,江苏南京210098 [2]陕西省水文水资源勘测中心,陕西西安710068

出　　处：《遥感技术与应用》2023年第6期1317-1327,共11页Remote Sensing Technology and Application

基　　金：国家重点研发计划项目(2018YFC1508100)。

摘　　要：为提高卫星遥感土壤湿度产品精度及空间分辨率,支撑水文过程的高精度模拟与预报,首先基于SMOS、SMAP、ASCAT 3种遥感土壤湿度产品,采用2D Triple-Collocation(2D TC)从时空两方面分析不同产品精度并基于最小二乘框架融合。而后分析支持向量机(SVM)、随机森林(RF)、极端梯度提升机(XGBoost)3种机器学习方法的降尺度效果及适用性,并对融合成果降尺度。研究结果表明:多源遥感融合成果较单一卫星遥感产品RMSE更小(RMSE=0.04 m3/m3),且优化了单一卫星监测数据缺失及高估的问题。3种基于融合成果的降尺度模型均有较好的抗过拟合能力,其中以XGBoost表现最优,其测试集R2较RF、SVM测试集分别提高了4.5%、36.6%,RMSE分别降低了15%、46.9%。以CLDAS产品为参考,XGBoost降尺度成果较融合成果R值提高了16.53%,RMSE降低了17.50%。融合与降尺度方法组合能有效提高遥感土壤湿度产品空间分辨率及精度,适用性更好。In order to improve the accuracy and spatial resolution of satellite remote sensing soil moisture prod⁃ucts and support the high-precision simulation and prediction of hydrological processes,firstly,based on three remote sensing soil moisture products of SMOS,SMAP and ASCAT,2D Triple-Collocation(2D TC)was used to analyze the accuracy of different products from both time and space,and the fusion was conducted based on the least squares framework.Then,the downscaling effect and applicability of support vector machine(SVM),Random Rorest(RF)and extreme gradient boosting(XGBoost)machine learning methods were ana⁃lyzed,and the fusion results were downscaled.The results show that the RMSE of multi-source remote sensing fusion results is smaller than that of single satellite remote sensing products(RMSE=0.04 m3/m3),and the problem of missing and overestimation of single satellite monitoring data is optimized.The three downscaling models based on fusion results all have good anti-overfitting ability,among which XGBoost performs best,and its R2 and RMSE(4.5%and 36.6%)are higher than those of RF and SVM for the test sets,which are reduced by 15%and 46.9%respectively.Taking the CLDAS product as a reference,the XGBoost downscaling result is 16.53%higher than the fusion result R value,and the RMSE is reduced by 17.50%.The fusion and down⁃scaling technique can effectively enhance the accuracy of remote sensing soil moisture products,thereby exhibit⁃ing superior applicability.

关 键 词：2D Triple-Collocation 机器学习 土壤湿度 多源融合 降尺度 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]