川西南山区县域数字土壤制图研究

Digital soil mapping of county level in southwest Sichuan mountainous areas

【目的】研究环境变量的筛选和土壤类型空间推理方法的选择,为有效提高县域数字土壤制图精度提供参考。【方法】以四川省盐源县为研究区,将气候要素、地形数据和遥感影像作为推理制图的辅助因子,利用野外采样点数据和环境协变量因子,采用决策树分类方法对环境特征进行重要性排序,特征筛选和组合优选,通过对比决策树分类、支持向量机和随机森林3种土壤分类方法的制图精度,探索基于土壤-环境关系理论提高具有立体气候特征的山区县域数字土壤制图精度的途径。【结果】(1)气候和地形特征在研究区土壤分类中起重要作用,采用单一气候因素作为土壤筛选分类的环境变量可获得土壤分类筛选精度为83.22%,依次增加地形和生物特征可分别使筛选精度提升至85.78%和89.43%;(2)与其他模型相比,随机森林模型制图效果更好,对比采样点数据的土壤分类总体精度为77.10%,Kappa系数为0.72;(3)气候因子以及地形因素对水热条件的影响是决定研究区土壤类型空间分布异质性的主要原因,年均气温、年积温、年降水量、相对湿度、高程、地形湿润度等环境特征因子与研究区主要土壤类型的空间分布关系密切。【结论】在具有立体气候特征的山区,基于随机森林方法使用气候、地形和遥感数据进行数字土壤分类制图有较好的效果。

[Objective]The selection of environmental variables and the selection of spatial reasoning methods for soil types were studied to provide references for improving the accuracy of digital soil mapping in counties.[Method]Taking Yanyuan county in Sichuan province as the research area,we selected climate factors,terrain data,and remote sensing images as auxiliary factors for inference mapping.Field sampling point data and environmental covariate factors were used to obtain the basic data of soil environment knowledge.The decision tree classifica-tion method was subsequently used for importance ranking,feature selection,and combination optimization of environmental features.Soil classification mapping accuracy of several soil classification methods,including decision tree classification,support vector machine,random forest,and SoLIM model,was compared.Based on the theory of soil environment relationship,the article explored ways to improve the accura-cy of digital soil mapping in mountainous counties with three-dimensional climate characteristics.[Result](i)Climate and topographic char-acteristics played an important role in soil classification in the study area.The screening accuracy of soil classification was 83.22%by using a single climate factor as the environmental variable of soil classification,The screening accuracy was increased to 85.78%and 89.43%by adding topographic and biological characteristics in turn.(i)Compared with other models,the random forest model achieved better mapping results.Using sampling point data as validation data,the overall accuracy of the soil classification was 77.10%,with Kappa coefficient of 0.72.(ii)The influence of climatic factors and topographic factors on regional climatic hydrothermal conditions mainly determined the spa-tial heterogeneity of soil types in the study area.Environmental factors,such as average annual temperature,annual accumulated temperature annual precipitation,relative humidity,elevation and topographic wetness were closely related to the spatial distribution of major soil types.[Conclusion]In the mountain counties with three-dimensional climate characteristics,the use of climate,topographic and remote sensing data for digital soil classification mapping based on random forest method has good results.