The Application Of Time-Series Sentinel-2 Extracted Agricultural Activity Factors In SOC Digital Mapping In Cultivated Land

Soil Organic Carbon（SOC）is an important indicator of Soil health and the main source of Carbon sequestration and emission reduction in farmland.Currently,Random Forest（RF）model based on collaborative variables is widely used in spatial mapping of soil organic carbon,but the selection of environmental variables is still inadequate.Human activities strongly affect the distribution of soil organic carbon content in topsoil of cultivated land,but at present,the spatial prediction of cultivated land SOC is still lacks the description of human factors,especially for soil mapping in large-scale complex geomorphic areas,and there was little regional comparative research.For this purpose,random forest（RF）model was applied to map the spatial distribution of topsoil organic carbon contents for farmlands in Ningde City in the northeast of Fujian and 11 counties under the jurisdiction of Sanming and Longyan City.The random forest model was used to extract crop rotation（CR）and vegetation characteristic transformation variables（HANTS）that can reflect the rotation pattern information based on sentinal-2 time series data,which were used as qualitative and quantitative factors of agricultural activities respectively.The natural environmental factors composed of climate and terrain are combined as environmental synergistic variables to participate in the construction of soil organic carbon RF model.For different regions,RF models driven by four combinations of environmental variables（climate+terrain,climate+terrain+rotation mode,climate+terrain+HANTS variable,climate+terrain+rotation mode+HANTS variable）are trained based on a large number of SOC measured sample point data.By comparing the performance of RF models with different variable combinations,this paper explores the possibility of improving the spatial prediction accuracy of cultivated land SOC by extracting agricultural activity factors from temporal sentinel-2 images,and selects the optimal RF model.Based on this,the spatial distribution pattern of SOC in the two areas was obtained,and the SOC mapping accuracy,SOC spatial distribution characteristics and main environmental factors in the two areas are analyzed and compared.The main conclusions are as follows:（1）By comparing the average prediction accuracy of RF model corresponding to four random sampling sets,it can be seen that for Western Fujian and Southeast Fujian,RF-D（CF+TF+Faa）model containing two agricultural activity factors obtains the highest prediction accuracy,while RF-A（CF+TF）model driven by natural environmental factors only obtains the lowest prediction accuracy.Compared with RF-A model,R²and R in Western Fujian increased by 89.47%and 36.36%respectively,and RMSE and Mae decreased by 10.66%and 12.05%respectively;In Northeast Fujian,R²and R increased by 275%and91.43%respectively,and RMSE and Mae decreased by 24.92%and23.88%respectively.The above results show that adding agricultural activity factors related to crop rotation mode can effectively improve the accuracy of SOC spatial mapping.（2）From the importance analysis of environmental variables in the two regions,it can be seen that all categories of variables were retained to participate in the modeling,but the retention factors and contribution were different.In the analysis of the importance of variables of the four RF models in Western Fujian,the top four variables were annual precipitation（Rainfall）,elevation（DEM）,annual maximum and minimum temperature（Maxt,Mint）,indicating that microclimate and regional topography affect the distribution of SOC content,especially climate factors play an important role in soil SOC mapping.At the same time,the rotation mode and some HANTS variables were also screened and retained,indicating that the distribution of SOC content in this region was also affected by agricultural activity factors.Among the four RF models in Northeast Fujian,the top four factors in the RF-A model are rainball,maxt,mint and aspect.When the agricultural activity factor CR or HANTS variable was added to the RF-A model,the factor ranking changes greatly;The top four in RF-B model are CR,DEM,slope and plane curvature respectively.The top four in RF-C model were DEM,second harmonic imaginary part（Imaginary2）,slope and mint in turn.DEM and slope were included in the top three of the two model factors,and the importance of agricultural activity factors Cr and imaginary2were ranked first and second.In the best model RF-D（CF+TF+Faa）,the top three factors were CR,imaginary2 and DEM,indicating that agricultural activity factors have a strong impact on the distribution of SOC content in Northeast Fujian.At the same time,climate factors only retain mint,and the ranking was relatively low,indicating that the distribution of SOC content in this area was also affected by regional terrain and microclimate,but the impact of climate factors was lower than that of terrain factors.To sum up,the distribution of SOC content in the two areas was affected by factors such as topography,climate and agricultural activities.（3）Based on the HANTS variables obtained from NDVI time series images with different cloud cover ratios,the RF model was reconstructed in cooperation with terrain,climate and rotation mode category factors.The prediction accuracy of models with cloud amount≤50%and≤100%in Western Fujian was lower than that of RF-D model（cloud amount≤10%）,R²decreases by 19.44%and 13.89%respectively,and RMSE increases by 4.36%and 3.25%respectively;The results of northeast Fujian are consistent with those of West Fujian.For the models with cloud amount≤50%and≤100%,R²decreases by 20%and8.89%,R decreases by 10.45%and 4.43%,RMSE increases by 7.92%and 3.20%,MAE increases by 7.24%and 1.55%respectively.It shows that due to the influence of cloud content,even if the time series remote sensing image was more abundant,it can not improve the prediction accuracy of the model,but reduce the calculation efficiency.（4）From the SOC content distribution predicted by the four RF models in the two areas,the overall distribution trend is consistent.In Western Fujian,the prediction range of the best environmental variable combination model RF-D（CF+TF+Faa）was 8.25～30.69 g kg^-1,with an average of 18.22±2.99 g kg^-1.The prediction results show that the content of soil organic carbon in the south was higher and that in the middle were lower.In Northeast Fujian,the range of SOC predicted by RF-D（CF+TF+Faa）was 7.70～29.50 g kg^-1,with an average of 16.47±2.94g kg^-1;The overall prediction results show that the SOC content in the western and northern regions was relatively high,and the SOC content in the southern and eastern coastal regions was relatively low.