摘要
精准的数字化土壤地图及空间属性数据库对于适地适树、适地适种分析,生态环境保护规划和决策,土壤生态系统多样性研究等都具有重要的理论及实践意义。为探索滇中及滇东南地区红壤空间分布,基于机器学习中的决策树和随机森林两种树形模型,利用1145个样点训练获取了滇东高原红壤与地形地貌、气候、生物等多种环境因子之间的非线性关系,并以250 m为最小栅格单元实现了滇东红壤的空间分布制图;通过31个已知剖面样本进行了模型检验,并与云南省1995年红壤分布图对比。结果显示:(1)决策树算法在验证集上的模型精度为82%,但过拟合严重;(2)使用随机森林算法拟合的验证集模型精度为81.38%,红壤空间分布结果精度为67.74%,与1995年版土壤类型图相比精度提高了9.68%;(3)树形模型中的随机森林算法具有更好的泛化和抗过拟合能力,拟合结果可展现更详细的空间细节和空间变化信息,与云南自然地理环境要素空间格局的吻合度更好,更适合大尺度的数字土壤制图;(4)影响滇东高原红壤分布的主要因素是海拔、温度、地表切割度和地表隆起度。本研究首次尝试了应用树形模型和空间大数据开展复杂山地区域的大尺度土壤类型制图,机器学习方法可以用于我国西南地区大尺度的精准化土壤数字制图。
Accurate digital soil maps and spatial attribute databases have important theoretical and practical significance for the analysis of suitable trees and plants,ecological restoration management planning and decision-making and soil ecosystem diversity research.In order to explore the spatial distribution of red soil in central and southeastern Yunnan,this paper uses 1145 sample points to obtain the nonlinear relationship between the red soil of the eastern Yunnan Plateau and various environmental factors such as topography,climate,and biology based on the decision tree and random forest tree model in machine learning,and the grid array unit with 250 m realizes the mapping of the spatial distribution of the red soil in eastern Yunnan Plateau.Compared with the red soil distribution map of Yunnan Province in 1995,the results tested by the data of 31 known profile samples.The result show that:(1)The accuracy of the decision tree on the verification set is 82%,but has serious overfitting;(2)The accuracy of the verification set fitted by the random forest algorithm is 81.38%,and the accuracy of the red soil spatial distribution result is 67.74%,increasing 9.68%accuracy compared with the current version;(3)The random forest algorithm in the tree model which has better generalization and ability to resist over-fitting,shows more detailed spatial details and spatial change information.The results fit by random forest have better goodness of fit with the spatial pattern of Yunnan’s natural geographic environment and more suitable for large-scale digital soil mapping;(4)The main factors affect the red soil distribution in eastern Yunnan Plateau are elevation,temperature,surface cut and surface uplift.This study is the first attempt to apply tree model in machine learning to map soil types in a large area in the complex terrain of Southwest China.Machine learning methods can be used for large-scale precision soil digital mapping in Southwest China.
作者
杨阳
叶江霞
王艳霞
蔡志勇
周汝良
YANG Yang;YE Jiang-xia;WANG Yan-xia;CAI Zhi-yong;ZHOU Ru-liang(School of Geography and Ecotourism,Southwest Forestry University,Kunming Yunnan 650224,P.R.China;School of Forestry,Southwest Forestry University,Kunming Yunnan 650224,P.R.China;AVIC General Aviation Research Institute Co.,Ltd./China Special Aircraft Research Institute,Zhuhai Guangdong 519000,P.R.China)
出处
《西部林业科学》
CAS
北大核心
2021年第6期31-39,共9页
Journal of West China Forestry Science
基金
国家自然科学基金(31760212)
云南省科技厅重大科技专项(202002AA100007)。
关键词
红壤
机器学习
决策树
随机森林
数字土壤
空间大数据
数字制图
red soil
machine learning
decision tree
random forest
digital mapping
spatial big data
digital cartography
