Comparison of RBF and MLP neural network performance and regression analysis to estimate carbon sequestration

Abstract

Soil organic carbon has favorable effects on the chemical, physical and thermal properties of soil, as well as its biological activities. Organic matter carbon is one of the important elements in soil, which plays a crucial role in soil quality of the forest ecosystems. In this research, to exactly estimate carbon sequestration (CS) according to the organic carbon and bulk density, we used RBF, MLP and multiple regression models. To do so, we took 60 soil samples from the depth of 0–15 cm of soil, across an altitudinal gradient of the forest, located at the Tarbiat Modares University Training Forest, and physicochemical soil properties (i.e., nitrogen, calcium, potassium, clay, silt, sand, organic carbon, pH, EC, bulk density and soil water content) as input variables for prediction of CS were measured. The results showed that CS of the study region was affected by soil physical and chemical characteristics. Furthermore, in all states, the RBF model statistically proved to have better prediction of CS compared to the MLP neural network and regression analysis, where the highest correlation between input variables and CS predicted with the least error was evident for RBF model followed by MLP and regression analysis, respectively. Moreover, the rate of carbon sequestration was not significantly affected by the amount of silt, whereas soil water content and soil electrical conductivity slightly affected the CS rate.