Soil Organic Carbon Mapping By Geostatistics and Artificial Neural Network Methods (Kohgiluyeh& Boyer-Ahmad Province)

Predicting and mapping soil organic carbon (SOC) contents and stocks are important for C sequestration, greenhouse gas emissions and national carbon balance inventories. The SOC plays a vital role in sustaining agricultural productions in arid ecosystems. It shows very quick and direct changes with atmosphere through the photosynthesis and the SOC decomposition. The depletion of C storage not only exacerbates the risk of soil erosion but also reduces agricultural production. An accurate knowledge of regional SOC contents and stocks and their spatial distribution are essential to optimize the soil management and land-use policy for SOC sequestration. Today, digital soil mapping methods such as geostatistics and artificial neural network (ANN) have focused more on SOC contents and stocks mapping. Geostatistics is a robust tool widely applied to model and quantify soil variation and analyze the spatial variability of SOC in large scale. The ANN as a nonlinear technique has been received much less attention for modeling SOC contents and stocks. Therefore, in this study, we aimed to develop and compare the performance of ordinary Kriging, co-kriging, inverse distance weighting (IDW) and artificial neural network models in predicting and mapping the SOC contents and stocks in East and Southeast of the Kohgiluyeh and Boyer-Ahmad province, southern Iran. The composite soil samples were collected randomly from the 0-15 cm soil depths at 204 sampling sites at different land uses in east and southeast of the Kohgiluyeh and Boyer-Ahmad province. The collected soil samples were air-dried, ground, and sieved to pass through a 2 mm mesh. Soil properties such as organic carbon contents and stocks, pH, electrical conductivity (EC), bulk density (BD) and soil texture were determined according to the standard analysis protocols. The normality tests were done according to the Kolmogrov–Smirnov method, and the variability of SOC contents and stocks were analyzed by the classical statistics (mean, maximum, minimum, standard deviation, skewness, and coefficient of variations). The digital elevation model (DEM), slope gradient, precipitation and temperature and Normalized Difference Vegetation Index (NDVI) were used as co-variables (auxiliary data). The NDVI was obtained by the remotely sensed data of LANDSAT 8. The geostatistical parameters were calculated for each soil property as a result of corresponding semivariogram analysis. The spatial prediction maps of soil properties were generated by ordinary kriging (OK), cokriging (Co-K) inverse distance weighting (IDW) with powers of 1, 2, 3, 4 and 5 as well as the Artificial Neural Network (Multilayer Perception model, MLP) methods. The mentioned interpolation methods were used to prepare the SOC spatial distribution maps by using the 80 % of data as the training datasets. The prediction results were then evaluated by the validation data set (20 % of all data). The differences between the observation and prediction values were evaluated by Mean Error (ME), Root Mean Square Error (RMSE), Correlation Coefficient (R2) and Concordance Correlation Coefficient (CCC). The spatial distribution maps of the SOC contents and stocks in the study area were finally developed by ArcGIS 10 software. The SOC content for all samples largely varied from 0.20 to 3.96 % .The high coefficient of variation of 53.38 % demonstrates the strong spatial variation of SOC content in the study area. The SOC stocks had also a relatively high variability compared with other soil properties. Such strong variation could be attributed to the diverse soil types, land covers and other environmental conditions across the study area. The average SOC content for forest land use was significantly higher than the other land uses. The intensive tillage in cropland soils appears to have induced the acceleration of organic carbon oxidations leading to the lowest SOC contents and stocks. By increasing the mean precipitation within our study area (in eastern and northeastern regions), the SOC contents and stocks increased significantly. The inverse trend was, however, observed for temperature implying the fact that the higher the temperature, the lower the SOC. Gaussian model was found to be the best model for parameters such as SOC contents and stocks due to the lowest RSS and R2.Overall, the results denoted the higher ability of ANN compared to geostatistical techniques (cokriging, kriging and IDW methods) in estimating both soil organic carbon contents and stocks. According to the results, ANN (MLP) method with one hidden layers with 50 neurons performed better in estimating soil organic carbon contents and stocks atunsampled points, whereas the largest errors were obtained for IDW method. The good performance of ANN method can be attributed to the division of the study area and the capability of ANN to capture the nonlinear relationships between SOC and environmental factors i.e. slope, DEM, precipitation, temperature and NDVI. The results suggest that the proposed structural method for ANN can play a vital role in improving the prediction accuracy of SOC spatial variability in large scale.