黄河景电灌区土壤盐渍化反演的多模型对比  

Multi-model comparison on soil salinization inversion in Jingdian irrigation area of the Yellow River

作　　者：蒋小芳 徐青霞 段翰晨 廖杰 郭平林 黄翠华 薛娴 Jiang Xiaofang;Xu Qingxia;Duan Hanchen;Liao Jie;Guo Pinglin;Huang Cuihua;Xue Xian(Key Laboratory of Desert and Desertification,Northwest Institute of Eco-Environment and Resources,Chinese Academy of Sciences,Lanzhou 730000,China;Drylands Salinization Research Station,Northwest Institute of Eco-Environment and Resources,Chinese Academy of Sciences,Lanzhou 730000,China;University of Chinese Academy of Sciences,Beijing 100049,China;Water Authority Bureau of Minqin County,Minqin 733300,Gansu,China)

机构地区：[1]中国科学院西北生态环境资源研究院沙漠与沙漠化重点实验室,甘肃兰州730000 [2]中国科学院西北生态环境资源研究院干旱区盐渍化研究站,甘肃兰州730000 [3]中国科学院大学,北京100049 [4]民勤县水务局,甘肃民勤733300

出　　处：《中国沙漠》2023年第5期18-30,共13页Journal of Desert Research

基　　金：第二次青藏高原综合科学考察研究项目(2019QZKK0305)。

摘　　要：位于中国西北干旱区东部的景电灌区是黄河景泰川电力提灌二期工程覆盖的重要地区。不合理的水资源利用和区内排水不畅导致该区成为次生盐渍化发生的重点区域。为更好地预测景电灌区的土壤盐渍化问题,服务盐渍化防治和盐渍土改良的国家需求,基于地表实测高光谱反射率和土壤电导率数据,从模型稳定性、噪声问题、共线性问题和准确度4个方面对比分析了深度神经网络(Deep neural network,DNN)、分布式随机森林(Distributed random forest,DRF)和梯度提升机(Gradient boosting machine,GBM)3个模型在景电灌区土壤盐分预测方面的适用性。结果表明:(1)实测高光谱反射率数据与土壤电导率之间存在较强的相关性,高光谱数据为土壤盐分预测研究提供了便利;(2)DNN模型的稳定性高,对噪声和共线性问题的处理能力更强,模拟准确度相对较高,而DRF和GBM模型模拟结果差别较小。DNN模型更适于景电灌区土壤盐分预测研究,这在模型适用性方面为该区域的土壤盐渍化研究提供了参考。Located in the eastern part of the arid area of northwest China,Jingdian irrigation area is an important region covered by the second phase of the Jingtaichuan electric power irrigation project of the Yellow River.Irrational water resources utilization and poor drainage in the area led to the occurrence of secondary salinization in the area.In order to better monitor the soil salinization problem in Jingdian irrigation area and serve the national demand for salinization prevention and improvement of saline soil,this paper compares and analyzes the deep neural network(DNN),distributed random forest(DRF),and gradient boosting machine(GBM)from four aspects:model stability,noise problem,collinearity problem,and accuracy based on the measured hyperspectral reflectance and soil electrical conductivity on the land surface.The results show that:(1)There is a strong correlation between the measured hyperspectral reflectance data and the electric conductivity of soil samples,and the hyperspectral data provides convenience for soil salinity prediction research.(2)The DNN model has high stability,stronger ability to deal with noise and collinearity problems,and relatively high simulation accuracy,while the simulation results of DRF and GBM models are less different.The results show that the DNN model is more suitable for soil salinity prediction in Jingdian irrigation area,which provides a reference for soil salinization research in this area in terms of model applicability.

关 键 词：高光谱反射率 DNN DRF GBM 盐渍化 景电灌区 

分 类 号：S153[农业科学—土壤学]