multivariate linear regression model

Soil organic carbon is very important in increasing the production of agricultural products and reducing soil erosion and greenhouse gas emissions. A better understanding of the amount and storage of soil organic carbon is essential to the use of soil and maintaining its productivity. The extended pastures of Fandoqloo region in Ardabil province potentially could support and sequestrate considerable amounts of carbon but they are subjected to land use change, high population and tourism. Such tensions led to land degradation and low carbon sequestration in some parts. This study aimed to find effective natural variables on soil carbon density and sequestration. Such variables could be applied for remediation and resilience of degraded pastures and croplands in Fandoghloo region. By application of Latin Hypercube techniques and spatial variables like soil, land use and geological maps the location of sample points was determined. For each land use unit (pasture or cropland) 70 composite surface soil samples (in total 140 samples) were collected from zero to 15 cm deep. After pretreatments of soil samples, soil's physical and chemical properties were measured according to standard protocols. A digital elevation model of the study area was formed using digital isolines of topographic maps and primary and secondary derivates were constructed. Landsat 8 bands were extracted from the Geology Department site of the USA and Normalized difference vegetation index and thermal map of the study area were calculated. Multiple linear regression and artificial neural network models were applied and soil data, digital elevation model and its derivates, normalized difference vegetation index and land surface temperature map have entered the models as independent variables for the prediction of soil organic carbon density. To evaluate the results of the models, the coefficients of determination and root mean square error was used. Application of multiple linear regression model indicated that total soil nitrogen, normalized difference vegetation index and mean weight diameter of aggregates justified 53% of soil organic carbon density in the pasture while 46% of soil organic carbon density in croplands land use was explained by total soil nitrogen and plane curvature. The results of the paired t-test showed a significant difference (P< 0.01) between soil organic carbon density of the pasture and cropland land use. An artificial neural network with an arrangement of 1-10-24 and tan-sigmoid transfer function in the hidden layer justified 71% and 82% of the soil organic carbon density in the pasture and croplands, respectively. Root mean square error of multiple linear regression models for predicting soil organic carbon density in the pasture and croplands were 1.07 and 1.04, respectively but of artificial neural network model was decreased to 0.85 and 0.58 for pasture and cropland, respectively. Sensitivity analysis revealed that the artificial neural network model was severely sensed to digital elevation model and soil moisture content at wilting point. The result of this study indicated the significance of natural vegetation cover and the mean weight diameter of soil aggregates for carbon sequestration in the pastures. Artificial neural networks performed to simulate changes in soil organic carbon density showed that these networks are successful in tracking an important part of soil organic carbon density changes at pasture and cropland land uses. In other words, artificial neural networks were able to identify the effects of these variables on soil organic carbon density using tracing and identifying the effects and interactions between an independent variable series with organic carbon density status. The results showed that transmission functions can be used by the data of primary and secondary topographic parameters for estimating soil organic carbon density in the region.