LAUSR

A reasonable predicting model for spatial variation of soil carbon would be a useful tool in monitoring and restoration of salt marshes. In this study, radial basis function neural networks model (RBFNN) and back propagation neural networks model (BPNN) were built to predict total carbon (TC), total organic carbon (TOC) and dissolved organic carbon (DOC) contents in salt marsh of the Yellow River Delta. Both models contained thirteen input parameters, i.e., nine topographic factors selected from ASTER GDEM Version2 and Geographical Information System (GIS), one vegetation index – MODIS 16-day composite Enhanced Vegetation Index (EVI), and three soil physicochemical properties. For prediction of TC, the MAE, MSE and RMSE values of RBFNN were smaller than those of BPNN by 61.87%, 81.36% and 56.82%; for TOC, the MAE, MSE and RMSE values of RBFNN were smaller than those of BPNN by 37.13%, 58.06% and 35.23%; both models had no significant difference in accuracy for DOC prediction, but the MAE, MSE and RMSE values of RBFNN were smaller. All ME values of RBFNN rather than BPNN were closer to zero. RBFNN integrating environmental factors had a higher accuracy than BPNN in predicting soil carbon content at a relatively small regional scale.