融合多源数据的高分辨率土壤水分模拟模型构建及应用  

Construction and application of a high-resolution soil moisture simulation model integrating multi-source data

作　　者：付平凡 杨晓静 姜波 苏志诚[1,2] 孙东亚[1,2] FU Pingfan;YANG Xiaojing;JIANG Bo;SU Zhicheng;SUN Dongya(China Institute of Water Resources and Hydropower Research,Beijing 100038,China;Research Center of Flood and Drought Disaster Reduction of the Ministry of Water Resources,Beijing 100038,China;Soil Moisture Monitoring Center of Jilin Province,Changchun 130033,China)

机构地区：[1]中国水利水电科学研究院,北京100038 [2]水利部防洪抗旱减灾工程技术研究中心,北京100038 [3]吉林省墒情监测中心,长春130033

出　　处：《农业工程学报》2025年第5期96-106,共11页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家重点研发计划项目(2023YFC3206001);江西省重点研发计划项目(20232BBG70029);江西省“科技+水利”联合计划项目(2022KSG01002)。

摘　　要：实时动态高分辨率土壤水分产品可为区域农业生产安全保障提供重要支撑,目前常用的土壤水分遥感产品存在空间分辨率较低及时间序列不连续等问题。为了生成时空连续的高分辨率土壤水分结果,该研究引入集成学习中的随机森林(random forest,RF)和梯度提升机(grandient boosting machine,GBM)算法,构建了融合多源数据的高分辨率土壤水分模拟(high-resolution soil moisture simulation,HRSMS)模型。2017—2022年SMAP微波土壤水分、植被指数、地表温度等遥感数据和墒情站点实测数据为模型输入和输出,利用Savitzky-Golay滤波方法和多元回归方法填补缺失的植被指数和地表温度数据,基于RF和GBM算法实现SMAP表层(0~5 cm)土壤水分数据分辨率提升(从9 km提高至1 km)。以吉林省为例验证模型可行性,结果表明:1)HRSMS模型相较于常用的多项式回归拟合法精度显著提升。均方根误差(root mean square error,RMSE)、平均绝对误差(mean absolute error,MAE)较多项式回归拟合法精度降低了22.2%、43.9%,决定系数(R^(2))提高了0.270,西北部粮食主产区的误差减少了33.2%;2)HRSMS模型中,RF与GBM算法计算效能相近,在吉林省开展相关研究时可结合数据条件任选其一进行模型构建。HRSMS模型有效提升了土壤水分遥感数据产品的分辨率和精度,对进一步提升土壤水分精准监测能力具有重要意义。Soil moisture is one of the most critical hydrologic indicators in the land-atmosphere heat exchange and global climate dynamics.The high-resolution products of soil moisture are greatly contributed to the precise monitoring of agricultural droughts.However,the existing datasets of soil moisture are limited to the coarse spatial resolution(typically>9 km)and temporal discontinuity.In this study,a high-resolution soil moisture simulation(hrsms)framework was developed to incorporate an ensemble learning approach,particularly for multisource data fusion.Spatially continuous estimates of soil moisture were then captured at 1 km resolution with temporal consistency.The accuracy of estimation was improved significantly,compared with the conventional approaches.Three computational procedures are included in the framework.Firstly,the high-resolution ancillary datasets(e.g.,vegetation indices and land surface temperature)were spatiotemporally reconstructed using Savitzky-Golay filtering with multivariate regression.Data gaps were also determined to preserve the temporal dynamics.Secondly,the spatial downscaling was performed on the soil moisture active passive(smap)observations(2017-2022,0-5 cm depth)from 9 km to 1 km resolution.A systematic investigation was also made to clarify the synergistic relationships among vegetation indices,land surface temperature,soil properties,and topographic parameters.In situ measurements were then implemented using ensemble machine learning,including random forest(rf)and gradient boosting machine(gbm).Thirdly,the multi-scale assessments were selected to compare with the original moderate resolution imaging spectroradiometer land surface temperature(modis lst)products.The point-scale evaluation of in-situ networks was also carried out in Jilin Province,China.A systematic quantification was then performed on the computational efficiency and accuracy metrics,including the root mean square error(rmse),mean absolute error(mae),and coefficient of determination(R^(2)).Finally,the polynomial regr

关 键 词：土壤水分 随机森林 梯度提升机 SMAP SSM 降尺度 点面数据融合 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] S152.7[自动化与计算机技术—控制科学与工程]