Prediction Of Spatial Distribution Of Main Soil Texture Types In Chongqing Influenced By Topographic Factors

Soil texture is one of the important physical properties of soil.Soils with different textures have different soil properties,such as porosity,permeability,heat capacity,storage,temperature change and tillage,which affect the transformation of soil moisture,nutrients,heat and air.It is an important input variable for hydrological model,land surface process model and coupled land surface process atmospheric model.Chongqing is located in the southwestern inland of China.Its topography and geomorphology are complex.About 98% of the land in Chongqing is mountainous and hilly.The soil texture types have strong spatial heterogeneity.There are mainly two kinds of methods for determining soil texture types.The laboratory method has the highest accuracy.It determines soil texture types by testing soil particle size and combining with soil texture classification criteria.As this method has higher time cost and economic cost.It is difficult to obtain large-scale data of soil texture types through this method.In practical research,manual measurement is usually adopted,which is considered to be a suitable alternative to laboratory method to determine soil texture types.According to the degree of soil physical and mechanical properties(cohesion and plasticity),soil texture types can be determined,which can be generally divided into clay,loam and sandy soil types.On a larger scale,although geostatistical models can predict soil texture types,their errors can not be ignored and can not meet the needs of relevant research.There is still a lack of suitable spatial distribution prediction methods for soil texture in small watershed scale.The spatial distribution of soil texture is significantly affected by topographic factors under the condition that the stratum of soil formation is relatively consistent with human activities.Therefore,based on limited soil texture samples,the relationship between topographic factors and soil texture types is established,and the optimal prediction models suitable for different regions are analyzed and determined.The spatial distribution data of soil texture types can be obtained quickly and cheaply,which can provide basis and data support for water cycle simulation and agricultural production planning in the current study area,and can also be carried out in other small watersheds.The prediction of spatial distribution of soil texture types provides a reference for improving the accuracy of relevant research results.According to the soil types formed by the main soil stratums in Chongqing,taking into account the main topographic and geomorphological characteristics(mountains and hills),four typical small watersheds of soil types(yellow soil,limestone soil,neutral purplish soil and calcic purplish soil)are selected as the study area.In order to to explore the optimal prediction model of each research area,the relevant topographic index is extracted,and five kinds of machine learning algorithms,such as Support Vector Machine(Polynomial Kernel),Support Vector Machine(Gaussian kernel),Artificial Neural Network,Classification Tree and Random Forest,are applied to study the influence of topographic factors on the spatial distribution of soil texture types.On this basis,the slope cultivated land type variables(slope and terrace)were introduced to reflect the change of micro-topography caused by human activities,and their effects on the spatial change of soil texture types were studied.Normalized Difference Vegetation Index(NDVI)was extracted from multi-temporal remote sensing data(Landsat 5 and Landsat 8).The optimal classification strategy and data set combination approach for predicting spatial distribution of soil texture types were explored by using one-against-one,one-against-all and all-together strategies,as well as topographic factors,stratums and NDVI data combinations.Experimental results indicates:(1)There are obvious differences in texture composition of soils with different types of stratums.A total of 495 soil samples were collected in the yellow soil Watershed with Triassic Xujiahe Formation sandstone as its stratum.The number of samples of clay,loam and sandy soil were 11,445 and 39,with the ratios of 2.2%,89.9% and 7.9%,respectively.537 soil samples were collected in the Limestone Soil Watershed with Triassic Daye Formation limestone as its stratum.The number of samples of clay,loam and sandy soil were 52,52,respectively.409 and 76,with the ratios of 9.7%,76.2% and 14.1% respectively;727 soil samples were collected in the small watershed of neutral purplish soil with sandy mudstone of the Jurassic Shaximiao Formation as the stratum,and the number of samples of clay,loam and sandy soil was 131,341 and 255,with the ratios of 18%,46.9% and 35.1%,respectively;and in the small watershed of calcic purplish soil with sandy mudstone of the Jurassic Suining Formation as the stratum,a total of 3636 soil samples were collected.The number of clay,loam and sand samples were 1872,1764 and 0,with the proportion of 51.5%,48.5% and 0%,respectively.This further illustrates that the stratum of soils significantly affects the composition of soil texture types.(2)Different soil texture types have their corresponding strong correlation topographic index.The main distribution areas of clay are lower elevation,higher topographic humidity index(TWI)and valley floor smoothness multi-resolution index(MRVBF);the main distribution areas of loam are higher elevation and ridge top smoothness multi-resolution index(MRRTF)and smaller slope(Slope);the main distribution areas of sandy soil are slope,vector robustness measure(VRM)and channel length.(Flow_PathL)was higher and TWI was lower.It is found that MRVBF and MRRTF have strong indicative function in identifying clay and loam.(3)Applying the machine learning models,the topographic factors are used to predict the soil texture types,and a spatial distribution prediction map of soil texture types in the study area is created,but the accuracy is different.The best models for small watershed of yellow soil and limestone soil were Support Vector Machine(Polynomial Kernel),with C=90 and P=3,the overall accuracy and Kappa coefficient are 0.943 and 0.79,respectively;the best models for small watershed of neutral purplish soil are random forest(mtry=3,ntree=500),the overall accuracy and Kappa coefficient are 0.656 and 0.601,respectively;the best model for small watershed of calcic purplish soil is also random forest with identical parameters.The overall accuracy and Kappa coefficient are 0.43 and 0.20,respectively.Although the simulation accuracy decreased,they all met the significant requirements.Overall,Support Vector Machine(Polynomial Kernel)model and Random Forest model get highter prediction accuracy.(4)There are differences in topographic indices affecting the distribution of soil texture types.The relative importance of TCI_Low and Flow_PathL to the prediction of soil texture type distribution is more than 70% in the small watershed of yellow soil and limestone soil.According to the order of importance,Elevation > TCI_Low > Flow_PathL > VRM > Slope_Heig > TWI > Slope > Slope_Leng > SPI > MRVBF > MRRTF;in small watershed of neutral purplish soil,elevation is the most important influencing factor,the relative importance of VRM,slope height and TWI ranged from 30% to 52%,ranked as Elevation > TWI > Slope_Heig > VRM > MRRTF > TCI_Low > SPI > Flow_PathL > Slope > Flow_Accum > Slope_Leng > MRVBF;in small watershed of calcic purplish soil,the relative importance of elevation was the greatest,followed by TWI,which reached 89.9%.The relative importance of Slope and Slope_Height and MRRTF ranged from 65.1% to 71.2%,ranked as Elevation > TWI > Slope > Slope_Heig > MRRTF > MRVBF > VRM > Flow_Accum > SPI > Flow_PathL > Slope_Leng.(5)Topographic factors and human activities have great influence on soil texture change.In the small watershed of calcic purplish soil,the types of sloping farmland(sloping land and terraced land)were used to investigate the intensity of human activities.The number of samples was 2502 and 1134,respectively.Variables of sloping farmland type were introduced into the stochastic forest model to calculate.The results showed that the overall accuracy and Kappa coefficient of the small watershed were 0.66 and 0.31,respectively,which significantly improved the prediction accuracy of the model,compared with the original optimal model.The overall accuracy and Kappa coefficient(0.60 and 0.20)were increased by 9.8% and 55% respectively.The effects of topographic index on soil texture change in sloping land and terraced land were studied.The overall accuracy and Kappa coefficients of the corresponding prediction models were 0.64,0.27,0.69 and 0.24,respectively.The results show that the types of sloping farmland(sloping land,terraced land)significantly affect the change of soil texture types in the study area;for sloping land alone,the relative importance of TWI and slope height is more than 90%;for terraced land alone,elevation is the most important variable affecting the change of soil texture.(6)Remote sensing data can assist topographic factors to further identify soil texture types.In small watershed of yellow soil and limestone soil,multi-temporal remote sensing data(Landsat 5 and Landsat 8)are used to select remote sensing image data of four seasons: day of year(DOY 116),summer(DOY 244),autumn(DOY 284)and winter(DOY 351).Radiation calibration and atmospheric calibration are pre-processed by ENVI.NDVI is calculated by calculation and classification tree with parent node = 3 and the child node = 1,is apolied to predict the spatial distribution of soil texture types.The results show that the overall accuracy and Kappa coefficient of NDVI model are 0.939 and 0.781 respectively;the overall accuracy and Kappa coefficient of NDVI model and soil stratum model are 0.935 and 0.764 respectively;the model using one-against-one classification strategy and combination of topographic index and soil stratum as input data is better,and its overall accuracy and Kappa coefficient are 0.975 and 0.918,respectively.The recognition improvements of clay,loam and sand are 144%,0% and 14% respectively.This indicates that the accuracy of identifying clay from loam and sand can be improved by introducing NDVI based on the prediction of soil texture type using topographic index and soil stratum,and it is determined that September 1 is the best remote sensing phase for predicting the spatial distribution of soil texture types.