Digital Mapping Of Soil Nutrients In Ridge And Valley Area Of Chuandong

The quality evaluation of cultivated land(QECL)plays a key role in national interest and people's livelihood.Measures for Surveying,Monitoring and Evaluating the Quality of Cultivated Land requires that the information of cultivated land quality should be regularly released every five years to ensure a comprehensive and timely understanding of the conditions of cultivated land and its changes.As an important part of QECL,accurate and efficient prediction of soil nutrients can not only provide necessary soil fertility information for QECL,but also provide a reference for predicting spatial distribution of soil nutrients in the areas with similar conditions.In present study,the study area was located in Changshou District,Chongqing,which is at the center of the ridge and valley area of Chuandong.The samples of Soil p H(p H),organic matter(OM),alkali-hydro nitrogen(Av N),available phosphorus(Av P),available potassium(Av K),available sulfur(S),available copper(Cu),available iron(Fe),available zinc(Zn)and available boron(B)were collected and analysed.Vegetation factor [normalized vegetation index(NDVI)],climate factors [annual precipitation(ANPR)and annual average temperature(ANTP)],terrain factors [elevation(ELE),relative slope position(RSP),valley depth(VD),topographic wetness index(TWI)and vertical distance to channel network(VDCN)] and parent material [stratum(ST)] were used as environmental variables.Spatial interpolation methods [ordinary kriging(OK)and inverse distance wighted(IDW)],machine learning methods [classification and regression tree(CART)and random forest(RF)] and hybrid geostatistical method [random forest with residual kriging(RFRK)] were used to model ten soil nutrients based on four kinds of environmetal variables mentioned above,respectively.Random forest was further used to determine the relative importance(RI)of input environmental variables.Finally,the model with the best prediction accuracy was used to map the spatial distribution of soil nutrients in the study area respectively.The main findings are as follows:(1)In the study area,soil p H(5.90 ± 0.86)is generally acidic.The contents of Av P(20.38 ± 11.68 mg/kg),S(49.39 ± 61.65 mg/kg),Cu(2.18 ± 1.28 mg/kg),Fe(98.62 ± 69.45 mg/kg)and Zn(1.60 ± 0.91 mg/kg)are at a high level.The contents of OM(17.46 ± 5.89 g/kg),Av N(59.57 ± 15.76 mg/kg)and Av K(69.49 ± 21.59 mg/kg)are at a medium level.The content of B(0.31 ± 0.21 mg/kg)is at a low level.The coefficients of variation(CV)of soil nutrients are between 14.58% and 124.83%.S is the only strong variability with a CV of 124.83%,while the others are moderate variability with the CVs between 10% and 100%.(2)Significant differences in p H,OM,Av N,Av P,Av K,S,Cu,Fe,B existed among the different STs(P<0.05).On the average,the soils developed from Triassic and Permian limestones had the highest OM and Av K contents of 27.34 g/kg and 79.31 mg/k,respectively.The soils developed from the Triassic Xujiahe Formation sandstone had the highest S,Cu and B contents of 67.20 mg/kg,3.21 mg/kg and 0.39 mg/kg,respectively.The soils developed from the Middle Jurassic Shaximiao Formation siltstone had the highest Av N,Av P and Fe contents of 61.05 mg/kg,21.65 mg/kg and 109.03 mg/kg,respectively.The highest soil p H of 6.77 was in the soils developed from the Upper Jurassic Suining Formation sandstone.There was no significant difference in Zn content among STs.(3)The pH and OM were closely related to S,Cu,Fe,Zn and B.The pH showed significant negative correlations with S,Fe and Zn(P < 0.01),and the Pearson's correlation coefficients were-0.17,-0.57 and-0.25.The correlations between p H and Cu,B was not significant.OM showed significant positive correlations with S,Cu,Fe,Zn and B(P < 0.01),with the Pearson's correlation coefficients of 0.465,0.67,0.36,0.26 and 0.43,respectively.(4)Soil nutrients had moderate or strong spatial autocorrelation in the study area.Among them,p H and B had moderate spatial autocorrelation,with the nugget effects of 27% and 50%,respectively.OM,Av N,Av P,Av K,S,Cu,Fe and Zn had strong spatial autocorrelation,which the nugget effects all were less than 25%.(5)RF was used to determine the main factors influencing soil nutrients variability.The results showed that ST,ANTP and VD are the top three important factors for predicting p H;ANTP,ELE and VD for OM;ANTP,VD and NDVI for Av N;VD,ANPR and ANTP for Av P;ANTP,VD and NDVI for Av K;OM,p H and NDVI for S;OM,NDVI and ANTP for Cu;p H,OM and NDVI for Fe;p H,OM and TWI for Zn;OM,VD and NDVI for B.For p H,OM,Av N,Av P and Av K,ANTP and VD were of general importance,while p H and OM generally exert impact on S,Cu,Fe,Zn and B.(6)The five models were used to predicte soil nutrients respectively,and their performance were tested.The optimal prediction models for p H,Av P and Av K were IDW,with determination coefficients(R~2)of 0.87 ? 0.61 and 0.81,respectively.The optimal prediction models for OM and Av N were the OK,with R~2 s of 0.80 and 0.69,respectively.The optimal prediction models for S,Fe and Zn were RF,with R~2 s of 0.18,0.63 and 0.38,respectively.The optimal prediction model for Cu was CART,with R~2 of 0.34.The optimal prediction model for B was RFRK,with R~2 of 0.25.The results showed that for soil nutrients with a large number of samples(p H,OM,Av N,Av P and Av K,with 5162 soil samples),the spatial interpolation methods provide higher prediction accuracy;while for soil nutrients with small sample numbers(S,Cu,Fe,Zn and B,with 316 soil samples),the machine learning methods and mixed geostatistics method provide higher prediction accuracy.