基于深度学习方法预测缺资料区域深层土壤水分  

Deep learning method-based prediction of deep soil moisture in data-deficient watershed

作　　者：张自豪 王国强[1] 薛宝林[1] 阿膺兰 张晓婧 王立波 ZHANG Zihao;WANG Guoqiang;XUE Baolin;A Yinglan;ZHANG Xiaojing;WANG Libo(College of Water Science,Beijing Normal University,Beijing 100875,China;State Key Laboratory of Watershed Water Cycle Simulation and Regulation,China Institute of Water Resources and Hydropower Research,Beijing 100038,China)

机构地区：[1]北京师范大学水科学研究院,北京100875 [2]中国水利水电科学研究院流域水循环模拟与调控国家重点实验室,北京100038

出　　处：《水利水电技术（中英文）》2022年第11期25-36,共12页Water Resources and Hydropower Engineering

基　　金：内蒙自治区科技重大专项(2020ZD0009);国家杰出青年科学基金项目(52125901)。

摘　　要：土壤水分是陆地大气间水和能量交换的重要变量,特别是半干旱区土壤水分在植被恢复过程中扮演着关键作用。针对黄河流域土壤水分站网稀少、获取高质量时空连续的土壤水分信息较为困难的问题,为了在这种缺资料地区获取高精度的土壤水分信息,从而进一步开展黄河流域生态水文退化恢复机制解析,选取该区域内浑河流域作为试验区,同为我国北方半干旱地区呼伦贝尔草原的土壤水分实测数据集被用来驱动时空融合的卷积-循环神经网络(CNN-RNN)深度学习方法,集成多源遥感数据和表层土壤水分数据预测缺资料区域浑河流域时空连续的深层土壤水分信息,并采用多指标评估该方法的可行性。结果表明:基于深度学习方法预测的浑河流域5~80 cm土壤水分与实测数据相比存在低估现象,总体的相关系数R、均方根误差RMSE和平均绝对误差MAE分别可以达到0.67、0.029 cm^(3)·cm^(-3)和0.025 cm^(3)·cm^(-3),其中5~10 cm土壤水分预测效果表现最好,即使对于难以预测的深层土壤水分,最低的R、RMSE和MAE也可以达到0.58、0.031 cm^(3)·cm^(-3)和0.027 cm^(3)·cm^(-3)。研究成果表明基于深度学习预测缺资料区域深层土壤水分的方法具有一定的可行性,为缺资料地区获取高精度的土壤水分提供了另一种思路。Soil moisture is an important variable for water and energy exchange between the terrestrial atmosphere,especially in semi-arid areas where soil moisture plays a key role in vegetation recovery.Aiming at the problem that the networks of soil moisture monitoring station within Yellow River Basin are sparse and then more difficult to obtain the relevant high quality spatio-temporal continuous soil moisture information,Hunhe River Watershed within the region is taken as the study area,while the measured soil moisture data set of Hulunbuir Grassland-a similar semi-arid area in northern China is use to drive the spatio-temporally fused convolution cyclic neural network(CNN-RNN)deep learning method for integrating multi-source remote sensing data and surface soil moisture data to predict the spatio-temporal continuous deep soil moisture information within the data-deficient region-Hunhe River Watershed,so as to get highly accurate soil moisture in this kind of data-deficient region and further analyze the restoration mechanism of the eco-hydrological degradation in the Yellow River Basin.Moreover,the feasibility of this method is evaluated by multiple indexes.The result shows that the soil moisture of 5~80 cm in Hunhe River Watershed predicted based on the deep learning method is underestimated if compared with the measured data,while the overall correlation coefficient(R),root mean square error(RMSE)and mean absolute error(MAE)can reach 0.67,0.029 cm^(3)·cm^(-3)and 0.025 cm^(3)·cm^(-3)respectively,in which the predicting effect is the best for the soil moisture of 5~10 cm,while the lowest R,RMSE and MAE can also reach 0.58,0.031 cm^(3)·cm^(-3)and 0.027 cm^(3)·cm^(-3)even for the deep soil moisture that is difficult to be predicted.The study result indicates that the deep learning-based method for predicting the deep soil moisture in the data-deficient region has certain feasibility and provides another idea for obtaining highly accurate soil moisture within data-deficient region.

关 键 词：土壤水分 深度学习 缺资料流域 生态水文 神经网络 半干旱区 黄河流域 气候变化 

分 类 号：P339[天文地球—水文科学]