基于机器学习模型的多层土壤湿度反演

为了获取深层土壤湿度缺测值,采用支持向量机、BP神经网络和随机森林3种机器学习算法,在表层至深层土壤中利用主成分分析法选择与土壤湿度相关性显著的气象因子作为输入数据,建立多层土壤湿度反演模型反演了不同深度的土壤湿度。结果表明:随机森林模型模拟结果更加稳定,反演效果更佳；受气象因子驱动的影响,3种机器学习模型对地表0~10.cm深度内土壤湿度的反演效果更佳,对深层土壤湿度的反演效果随着深度增加而变差；增加表层土壤湿度及不同深度土壤温度作为驱动因子可以有效提高机器学习模型对深层土壤湿度的反演能力。

Multi-layer soil moisture inversion based on machine learning models

In order to obtain soil moisture data at deep soil layers and to complete the missing soil moisture measurements at different depths, machine learning technique is used to perform the soil moisture inversion at multi-layers from surface to root zone. Three machine learning algorithms, support vector machine (SVM), back propagation (BP) neural network and random forest (RF), are applied to construct the inversion models for the train and inversion of soil moisture data at different soil layers, and the meteorological factors, which have high correlation coefficient with the soil moisture, are selected as the input factors based on the principal component analysis (PCA) method in each layer of soil. The major findings are: the simulation results of random forest are more stable and the inversion effect is the best; due to the impact of meteorological variables, the three machine learning models perform best for the inversion of surface soil moisture at the surface soil layer within 0 to 10 cm depth; however, the inversion effect for the soil moisture at the deep zone is gradually reduced with the depth of soil layer; the addition of surface soil moisture and soil temperature at different soil layers as driven factors could improve the inversion capacity of machine learning model for the soil moisture at the deep zone.