Using An Artificial Neural Network To Estimate High-resolution Organic Carbon Storage In Forest Soil And Its Extension Capability

Soil organic carbon storage(SOCS)estimation is a crucial branch of the atmospheric-vegetation-soil carbon cycle study under the background of global climate change.Against this background,accurate SOCS estimation has become a massive task faced by the forestry and ecological circles.The objective of this study is to develop the artificial neural network(ANN)model with good extension capability to accurately estimate SOCS.The method of this study included two sections:1)The first section took Luoding City,Guangdong Province,as the modeling area.A Back Propagation-Artificial Neural Network(BP-ANN)was built to estimate the SOCS of the modeling area at five soil layers(L1:0-20cm,L2:20-40cm,L3:40-60cm,L4:60-80cm,L5:80-100cm),which adopted the 1:1000000 scale soil organic matter(SOM)map as the required input parameter,the nine topographic and hydrologic parameters derived by 10m resolution digital elevation model(DEM)as the candidate input parameters.The nine candidate input parameters including four topographic variables:slope,aspect,topographic position index(TPI),and potential solar radiation(PSR),and five hydrological variables:soil terrain factor(STF),sediment delivery ratio(SDR),depth to water(DTW),flow length(FL),and flow direction(FD).The optimal candidate input parameters combination of each soil layer would be obtained by evaluating the root mean square error(RMSE),coefficient of determination(R²)and mean absolute error(MAE).Moreover,three common methods,including soil type method(STM),Ordinary Kriging(OK)and Radial basis function(RBF)interpolation methods,were introduced as comparisons.The four methods used 90%(about 203)of the modelling area's soil samples for model building,and the rest of the 10%(about23)soil samples were used as validation samples.The comparisons of the four methods would base on the estimation accuracy,result,and mapping capability.2)The second section,which was divided into two stages,took Xinxing County,Guangdong Province,as the extension area.For the first stage,the optimal ANN models of each soil layer,which built with the modeling area's data,were directly extended to the extension area through Matlab software.For the second stage,a linear model,taking 1:1000000 scale soil organic matter map as modeling criteria,was introduced to capture the regional difference and further improve the accuracy of the first stage.The linear models built by 20%(24)of the extension area's soil samples,and the rest of 80%(96)soil samples were adopted as validation samples to verify the extension accuracies of the two stages.The results are as follows:1)The optimal candidate parameters combination of ANN models of each soil layer were as follows:the model of L1 layer contained seven parameters:Slope,Aspect,SDR,STF,PSR,FL,and FD;the model of L2 layer contained seven parameters:Slope,STF,SDR,PSR,TPI,Aspect,and FD;the model of L3 layer contained six parameters:PSR,Aspect,FD,SDR,TPI and DTW;the model of L4 layer contained six parameters:Slope,PSR,Aspect,DTW,SDR,and FL;the model of L5 layer contained seven parameters:STF,DTW,FL,SDR,PSR,TPI,and FD.The RMSE of L1-L5layers were 1.53-2.40 kg·m^-2,the R² were 0.81-0.84,and the MAE were 0.29-0.41.2)The validation results for the first section show that the ANN method has much better estimation accuracy than STM,OK,and RBF,with the RMSE of the five soil layers decreasing by 0.62-0.90 kg·m^-2,R² increasing from 0.54 to0.65,and the MAE decreasing from 0.32 to 0.42.The SOCS estimated by the four methods all showed a decreasing trend with the increase of soil depth.The SOCD estimated by the ANN method were 3.7 kg·m^-2 at L1,2.7 kg·m^-2 at L2,2.3kg·m^-2 at L3,2.0 kg·m^-2 at L4,and 1.8 kg·m^-2 at L5(the rangeability of the four methods results?0.5 kg·m^-2);and the total SOCS in the 0-100cm soil layer of the modeling area was 28.7Tg.For the spatial distribution map,the ANN method could portray the differences at the grid dimension,which reveal the local variation of the modeling area.In general,the ANN method has the best performance in terms of the SOCS estimation accuracy,result,and mapping when compared with the other three methods.3)For the validation results of the second section,the R² of the linear model was increased by 0.23-0.29 compared with that of the first stage.The mean R²of the five soil layers was increased from 0.36(of the first stage)to 0.62,which displayed a good modify capability.The SOCS estimation results of the first stage were similar to those of the modeling area,which indicated that the direct extension might lead to analogical results when extended to an unmodeled area.By introducing the linear model,the extension results,presenting in the SOCS distribution maps,were rid of the distribution law of the modeling area.In conclusion,the ANN model is an effective method for obtaining accurate SOCS estimation results on a regional scale.By adopting a linear model approach,the ANN method can effectively extend the estimation region with only a small number of field samples with a preferable level of accuracy.