基于地理加权回归的地形平缓区土壤有机质空间建模

Spatial modeling of soil organic matter over low relief areas based on geographically weighted regression

气候变化效应评估、土壤固碳潜力和肥力管理等,迫切需要详尽的土壤有机质(soil organic matter,SOM)空间分布信息。该文以江苏省第二次土壤普查的1519个典型土壤剖面的表层(0~20 cm)SOM含量为例,选择1217个样本为建模集,302个为验证集,选取年均温度、年均降雨、物理性黏粒和土壤pH值等因子进行SOM的地理加权回归(geographically weighted regression,GWR)建模。从建模集中分别随机抽取100%(1217个)、80%(973个)、60%(730个)、40%(486个),20%(243个)的样点,对比不同样点数量下GWR和传统全局回归模型的精度差异,并选择最优模型进行SOM空间预测制图。结果表明:1)江苏省SOM含量在不同空间尺度上存在极显著的空间自相关性。不同样点数量的建模集的全局自相关性和局部空间自相关聚类图结果相似。全局Moran’s I值介于0.25~0.61(P<0.001)。SOM含量空间分布以空间聚集特征为主,“高-高”聚集区主要分布在苏中和苏南地区,“低-低”聚集区主要分布在苏北地区。2)GWR建模结果均优于传统的全局回归建模,其残差在不同的空间尺度上均不存在空间自相关性。不同建模集的GWR的R2 adj较全局建模均提高0.15~0.20,其AIC和RSS均比全局模型有大幅降低,为56.08~360.19和17.40~76.67。不同建模样本数量的GWR模型对SOM的解释能力差异较小。3)建模样点数量(除建模样本n=243)对GWR预测制图结果的精度影响不大,RMSE介于5.56~5.75 g/kg之间,MAE介于3.87~4.05 g/kg之间,R2介于0.52~0.48之间,均优于全部建模样点的普通克里格插值验证结果。该研究可为样点数较少的省级尺度地区SOM空间建模与制图提供借鉴。

Accurate estimates of the spatial variability of soil organic matter(SOM)are necessary to properly evaluate climatic chagne,soil carbon sequestration potential and soil fertility.In plains and gently undulating terrains,soil spatial variability is not closely related to relief,and thus digital soil mapping(DSM)methods based on soil-landscape relationships often fail in these areas.Therefore,different predictors or methods are needed for DSM in plains.In provincial regional scale,climatic factors influence spatial distribution of soil properties.For this research,Jiangsu Province was selected as example and mean annual temperature(MAT),mean annual precipitation(MAP),physical clay content,and soil pH were selected for SOM spatial modeling using geographically weighted regression(GWR).The SOM content in the surface layer(0-20cm)of 1519 typical soil profiles of the Second National Soil Survey in Jiangsu Province were collected.1217 samples were selected as the modeling set and 302 were the validation set.Fristly,100%(1217),80%(973),60%(730),40%(486),and 20%(243)samples were randomly selected from the modeling set,and global and local spatial autocorrelation of SOM content were analyzed at different spatial scales using spatial statistics tools in ArcGIS.Secondly,comparison of the accuracy between GWR model and the global regression model under the different sampling size was conducted.Akaike information criterion(AIC),residual sum of squares(RSS)and adjustment determination coefficient(R2 adj)were used modeling comparison.Thirdly,the optimal model was selected for mapping SOM spatial prediction.Independent validation was used for model evaluation,using four indices:mean error(ME),mean absolute error(MAE)and root mean of squared error(RMSE),and determination coefficient(R2).Results show that:1)There was a significant spatial autocorrelation of SOM content in Jiangsu Province at different spatial scales.The clustering pattern of global and local spatial autocorrelation of modeling set with different sampling size were similar.The global Moran’s I ranged from 0.25 to 0.61(P<0.001).The spatial distribution of SOM content was mainly characterized by spatial clustering pattern.The“high-high”clustering areas were mainly distributed in the central and south of Jiangsu,and the“low-low”clustering areas were mainly distributed in the north of Jiangsu.2)The modeling results of GWR were better than the global regression modeling,and the residuals had no spatial autocorrelation at different spatial scales.The R2 adj of GWR in different modeling sets was increased by 0.15 to 0.20 compared with the global model.The AIC and RSS were significantly lower than the global model,which were decreased by 56.08 to 360.19 and 17.40 to 76.67 respectively.There were slight difference between GWR models with different sampling size.3)The number of modeling samples(except for the number of modeling samples was 243)had little effect on the accuracy of prediction and mapping results of GWR,the RMSE was between 5.56 and 5.75 g/kg,MAE was between 3.87 and 4.05 g/kg and R2 was between 0.48 and 0.52.The results were all better than the validation result of Ordinary Kriging using all modeling sampling points.This study can provide reference for SOM modeling and mapping in large and low relief areas with sparse samples.