摘要
为探讨随机森林(random forest, RF)模型对土壤属性空间预测的精度,本文以安徽省为例,收集140个土壤样本,利用GIS和RS技术,获取相关的地形因子、遥感植被指数及气候数据,利用RF模型分析土壤有机碳(SOC)含量、土壤容重和土壤黏粒含量与地形因子、遥感植被指数及气候数据之间的关系,并进行空间分布预测。研究结果表明:①RF建模预测中,当节点分裂次数(mtry)值为1,决策树数量(ntree)值分别为100、1 000和100时,获得的SOC含量、土壤容重和土壤黏粒含量RF模型最优;②高程、归一化植被指数(NDVI),地貌、多尺度山谷平坦指数(MrVBF)和土壤类型是SOC含量的重要预测因子;地貌、年均降水量(MAP)、MrVBF、高程和土壤类型是土壤容重的重要预测因子;高程、MAP、MrVBF和平面曲率是土壤黏粒含量的重要预测因子;③RF模型可以较好地进行土壤属性空间预测,多源环境变量组合可以分别解释SOC含量、土壤容重和土壤黏粒含量的26%、23%和22%;同时RF模型对于土壤类型和地貌等类型变量的处理具有一定优势。研究表明,在大尺度研究区域内,利用RF模型进行土壤属性空间预测有一定的意义。
It is important to study the spatial variability and distribution of soil properties for understanding ecosystems, formulating agricultural policies, conducting soil management and monitoring environmental changes caused by land use. The purpose of this paper is to explore the accuracy of the spatial prediction of soil properties at the provincial scale by the Random Forest (RF) model. Anhui Province in East China was selected as the study area, soil data obtained during the 2nd National Soil Survey and during 2010 2011 were used, the environmental variables were collected with GIS spatial analysis technique, and the correlation between environmental factors and soil properties was analyzed by RF model. The results showed that in the RF modeling process, SOC prediction model was the most robust and the prediction accuracy was the highest when the mtry value was 1 and the ntree value was 1 000;when the mtry value was 1 and the ntree value was 1 000 and 100 respectively, soil bulk density (BD) and clay content prediction models were the best. The elevation, NDVI, landform, muti-resolution index of valley bottom flatness (MrVBF) and soil type were the most important predictors of SOC content;Landform, mean annual precipitation (MAP), MrVBF, elevation and soil type were the most important prediction factors of soil BD;Elevation, MAP, MrVBF and plan curvature were the most important predictors of soil clay content;RF model can be used for spatial prediction of soil properties and has certain advantages in treating the qualitative variables such as soil type and landform;Multi-source environmental variable combinations explained 26% of SOC content, 23% of soil Bd and 22% of clay content, respectively. The use of machine learning for predicting soil properties and digital soil mapping is more efficient than traditional methods, it is of significance to use RF model in spatially predicting soil properties in the large-scale area.
作者
卢宏亮
赵明松
刘斌寅
张平
陆龙妹
LU Hongliang;ZHAO Mingsong;LIU Binyin;ZHANG Ping;LU Longmei(School of Geodesy and Geomatics, Anhui University of Science and Technology, Huainan, Anhui 232001, China;State KeyLaboratory of Soil and Sustainable Agriculture, Institute of Soil Science, Chinese Academy of Sciences, Nanjing 210008, China)
出处
《土壤》
CAS
CSCD
北大核心
2019年第3期602-608,共7页
Soils
基金
国家自然科学基金项目(41501226)
安徽省高校自然科学研究项目(KJ2015A034)
土壤与农业可持续发展国家重点实验室开发基金项目(Y412201431)资助
关键词
土壤属性预测
随机森林模型
环境变量
安徽省
Soil properties prediction
Random Forest model
Environmental variables
Anhui Province