Modeling The Spatial Dynamics Of Key Soil Properties Using Multisource Data In Fuzhou City,China

Soil is an integral component of the ecosystem,serving as a backbone to achieve the global sustainable development goals of food security,mitigation of land degradation,and the provision of key ecosystem services.However,the urban soils are massively disturbed by intensive human activities where urbanization and industrialization processes release an increased amount of organic and inorganic contaminants to the ecosystems that accumulate on the topsoil.Soil contamination and deterioration of soil organic matter are receiving more attention because it is increasingly affecting the atmosphere,water bodies,climate,human health,and whole ecosystems.The rapid economic growth and urbanization have caused soil contamination and the decline of soil organic matter in the coastal regions of China.Fuzhou is one of the largest coastal cities in China,located in the estuaries of Mingjian River,Fujian Province,Southeast China.It has experienced a huge expansion of urbanization and the built-up areas in recent decades,causing a decline in ecological quality.An increasing decline of soil organic matter and soil contamination in the current study area needs a comprehensive assessment of the spatial variability of key soil properties.Recently,remotely sensed earth observation(EO)products,proximal sensing data,and non-linear machine-learning models allowed soil monitoring,mapping,and assessing the spatial distributions of key soil properties in a large geographic area more comprehensively,quickly,and robustly.Even though machine-learning methods and remotely sensed predictors are widely used to model soil properties,there is still a critical need to find appropriate spatial soil modeling approaches and optimal sets of predictors.To that end,this study aimed to model the spatial dynamics of key soil properties using multisource data in Fuzhou City,China.Firstly,it modeled the spatial dynamics of soil organic carbon(SOC)with remotely-sensed predictors using the RF and Classification and Regression Tree(CART)and identified the most influential variables for prediction.Additionally,it assessed the spatial variability of SOC across land use,landform and lithology.The result indicated that the mean SOC concentration was11.70 mg/g,where most of the area was classified as humus and organo-humus in the mountainous regions.RF model had good prediction performance with corresponding high R~2and RMSE of 0.96 and 0.91 mg/g,respectively.The biophysical land surface indices,brightness removed vegetation indices,topographic indices,and soil spectral bands,respectively,were the most influential predictors.SOC was variable across land use,landform,and lithology,where forestland had the highest SOC(13.60 mg/g)content.Secondly,it evaluated the performance of selected machine-learning approaches to predict total soil nitrogen(STN)using remote sensing and multisource environmental covariates.Additionally,it assessed the spatial variability of STN across land use,landform,lithology and characterized it based on multisource data.The results from this study showed that random forest(RF),support vector machine(SVM),artificial neural network(ANN),multi-linear regression(MLR),and locally weighted regression(LWR)can achieve high R~2 values of 0.96,0.92,0.80,0.97,and 0.93 with corresponding RMSECV values of 0.08,0.35,0.37,0.43,and 0.65,respectively.RF was the most effective model with the highest R~2 and lowest RMSECV.RF and SVM models were used to select important predictors,where RF selected mainly vegetation indices while SVM selected VIS-NIR spectra.Additionally,STN contents had relationships with most environmental covariates derived from remote sensing,soil spectra,and topographic variables.Spectral transformations had improved the correlations with STN.Thirdly,the spatial distribution,extent,and sources of selected soil heavy metals and their distribution pattern across land use,landform and lithology were assessed.The result showed that the spatial distribution of heavy metals was highly variable for each element,but the Cangshan district and its surroundings were hotspots for most heavy metals.Cr,Cu,Pb,Fe,As,and Mn had strong spatial dependence implying their distribution may be impacted by natural factors such as topography and parent material.Pb,Cr,Ni,and Zn had strong relationships,and their peak values were in urbanized regions.However,Cd,Mn,and As may be influenced by natural and anthropogenic sources such as industrial effluents and agricultural pollution.Land use,landform and lithology had a significant impact on the variability of Cd,As and Pb at p<0.01.The study area was clean for most heavy metals except for Cd and Zn,which had strong and moderate contaminations,respectively,in parts of the study area.Generally,the modeling approaches used in this study had acceptable results,and similar approaches can be used to model SOC,STN,and heavy metals in a similar eco-environment.Particularly,remotely sensed predictors including land moisture,land surface temperature and built-up indices were important for predicting SOC and STN.Hence,the land surface indices may provide new insights into SOC modeling in complex landscapes of warm sub-tropical urban regions.Moreover,machine-learning methods are more practical approaches for predicting STN and can be used in similar complex coastal environments.Based on the results,this study recommends implementing sustainable agricultural practices to mitigate soil contamination and maintain healthy coastal ecosystems.This study derived biophysical land surface indices of soil moisture,land surface temperature,and human influences from LANDSAT-8 images;further studies could consider high-resolution Sentinel products.Moreover,this study compared SOC and STN with biophysical land surface variables extracted from LANDSAT-8 images captured during the winter months.However,future studies may assess the effects of the seasonal biophysical land-surface variations on the spatial distributions of SOC and STN.Additionally,future research is recommended to evaluate the performances of more machine learning and deep learning methods.This research can be used as a scientific basis for developing an appropriate soil monitoring system in the study area and other related environments in other regions.This study contributes to soil remote sensing in general and spatial modeling of SOC,STN,and heavy metals in complex coastal cities.Moreover,it adds knowledge to remote sensing and GIS in general and soils remote sensing in particular.