عنوان مقاله [English]
Background and objectives: Considering global warming as a prominent challenge in the world and based on increase and protection of soil carbon stock in forest ecosystems, optimal monitoring carbon sequestration in Forests is very efficient to verify its sustainability and climate change. Therefore, a comprehensive protocol is needed for developing models with high certainty in relation to soil organic carbon (SOC) stock. Furthermore, using the models including explanatory variables attributed by easy measurement capabilities and the lowest cost is conducted to properly protective control in association with variation of carbon sequestration in different soil layers.
Materials and methods: This research was conducted in Nour Forest Park as a largest plain forest in north of Iran. The number of 25 plots with 400 m2 was placed in the three stands (Alder- Ironwood, Maple- ironwood, Elm- Hornbeam) based on randomized block design sampling to measure plants biodiversity and soil carbon factor. Regression analysis and simulating by ANN was used to monitor soil carbon pool optimally.
Results: The results showed that multiple linear regression including correlated indices based on validation parameters such as Akaike information coefficient and variance inflation factor (VIF < 10) were valid to estimate the SOC stock; however, there was found no considerable accuracy. Pertaining to non-linear regression analysis, S-curve model including Abundance of herbal species was the best predictor to SOC in the organic layer and log-transformed power model (CF = 1.00) including species dominance (D) of trees, Abundance of and evenness (J') of grasses was the highest accurate model to predict SOC in the mineral layer. After back propagation algorithm in the neural network, SOC stock was predicted well through input matrix of Abundance and J' with topology of one hidden layer and 15 neurons including function of Tan-sigmoid. Furthermore, SOC stock in the mineral layer was predicted with the highest accuracy by adding layer of D to the prior matrix in the network having architecture of two hidden layers and 35 neurons in each layer.
Conclusion: Irrespective of ecological and biological interpretations, an accurate type of mathematic relationship (Function) between plant biodiversity indices and the response variable is dubious with low certainty and statistical issues such as variance inflation factor. Hence, ANN application can be a best alternative to traditional models for the response variables. Subsequently, like the parameters in the traditional models, the topology of each model in ANN is determining the architecture and efficiency of monitoring the responses in different soil layers.