Abstract

Background: The use of mathematical models to estimate soil properties can be an alternative to reduce costs with data collection and analysis, allowing better quantification and characterization of soil attributes, highlighting in that branch artificial neural networks (ANN). Objective: This study aimed to propose, construct and validate a model to estimate soil organic carbon stock in a fragment of a semideciduous submontane forest through Artificial Neural Networks (ANN). Results: The best ANNwere selected based on the Pearson correlation coefficient (r), square root of the mean error (RMSE) and graphical analysis of residuals dispersion. The best correlation, both in training and validation,were observed for ANN1, which comprises the physicochemical variables of the soil, and for ANN 2, with chemical variables. The results of the use of neural network system to predict soil organic carbon stock are favorable. The input variables for the neural network used in this study are relatively easy to obtain, providing the use of the proposed methodology for the estimation of the stock of soil organic carbon. Conclusion: The physical and chemical properties of the soil used in this study allowed the prediction of organic carbon stock in semi-deciduous submontane forest. ANN 2 was considered the most appropriate network to estimate the stock of organic carbon in the soil. However, further work should be carried out with different configurations of neural networks, i.e. using different architectures to obtain the better correlation between the data of SOC stock and the physical and chemical soil properties, thus obtaining minor error in the response variable estimates.

Authors and Affiliations