西南农业学报2025，Vol.38Issue(10)：2064-2075,12.DOI:10.16213/j.cnki.scjas.2025.10.003
多模型集成驱动下的川中丘陵区土壤属性制图方法研究
Soil attribute mapping method assessment in hilly area of central Sichuan province driven by multi-model integration
摘要
Abstract
[Objective]The complex terrain and highly heterogeneous soil-forming environment in hilly areas pose challenges to achieving high-precision spatial prediction of soil properties by traditional methods.The paper focuses on Suining city,Sichuan province as a representa-tive area to evaluate the applicability of multiple models and identify the optimal model scheme.The study reveals the spatial differentiation patterns of soil properties and their main controlling environmental factors,and examines the evolution trend of soil properties over the past 40 years,providing data support and technical pathways for precise soil management.[Method]Based on 4133 surface sample points in Suining city,the study selected four indicators:pH,organic matter(OM),cation exchange capacity(CEC)and total nitrogen(TN).A total of 25 en-vironmental variables across five categories:climate,topography,parent material,remote sensing indices and human activities-were included in the analysis.Random Forest(RF)was employed to screen key environmental factors.The performance of five models:Ordinary Kriging(OK),extreme gradient boosting(XGBoost),Random forest(RF),Stacking and random forest residual Kriging(RF+OK)in soil property mapping was compared.Model performance was comprehensively evaluated using indicators such as the coefficient of determination(R2),root mean square error(RMSE),concordance correlation coefficient(CCC)and Bias.Further analysis of variable importance was conducted to identify the main controlling environmental factors of soil attribute differentiation.[Result](i)Among all models,the RF+OK model ex-hibited the best performance on the independent test set(average R2=0.78±0.04),significantly outperforming other models(such as Stacking R2=0.65).This model effectively combined the nonlinear analytical ability of RF for environmental factors with the residual correc-tion of OK for spatial autocorrelation.(ⅱ)The spatial differentiation of soil properties showed attribute specificity:endogenous properties(pH,CEC)were mainly controlled by geological background(strata+lithology,importance＞22.8％);Exogenous attributes(OM,TN)were dominated by human activities and hydrological processes(land use+river network distance,importance＞42.2％).(ⅲ)Spatial pattern:the spatial patterns predicted by different models showed similar trends but exhibited critical local differences.(iv)Spatio-temporal evolution:over the past 40 years,soil properties had shown a trend of'three increases and one decrease':OM(average 15.67 g/kg),CEC(average 23.21 cmol/kg)and TN(average 1.12 g/kg)had significantly increased,but the area with extreme pH(＞8.5 or ≤4.5)ex-panded,forming the core of soil degradation risk.[Conclusion]The RF+OK model provided an effective path for mapping soil properties in hilly areas.The evolution of soil properties in Suining city was driven by the three-dimensional composite of'geological background-human activities-spatial dependence'.There was an urgent need to establish a smart soil management platform based on the concept of'geological zoning-targeted measures'to address the risks of increasing soil alkalization and unbalanced fertility.关键词
随机森林残差克里格/Stacking模型/环境变量/时空演变/川中丘陵区Key words
Random forest residual kriging/Stacking model/Environmental variables/Spatial and temporal evolution/Central Sichuan hilly area分类
农业科技引用本文复制引用
邓春秀,李源洪,何霞,陈小辉,董秀春,唐康其,曾瑜苹,郭伟,黄平,王思..多模型集成驱动下的川中丘陵区土壤属性制图方法研究[J].西南农业学报,2025,38(10):2064-2075,12.基金项目
四川省科技计划项目(2023YFN0070) （2023YFN0070）
