Predicting soil carbon stock in remote areas of the Central Amazon region using machine learning techniques

The use of covariates derived from remote sensors in combination with machine learning (ML) algorithms has been shown to be promising for mapping soil types and their attributes in large areas. This study explores the feasibility of using the existing knowledge of soil organic carbon stock (SOCS) derived from a relatively low density and irregular dataset to map a large area of 13.440 km2 located in a remote region under the Amazon Rainforest. The objectives of this study were to evaluate: 1-two different types of sampling approach (Reference Area -RA and Total Area-TA) to predict SOCS at depths of 30 and 100 cm; 2-two categories of covariate se-lection; 3-the transferability and the performance of three ML algorithms (regression tree-RT, random forest (RF) and support vector machine (SVM). The dataset consisted of 120 observations of SOCS30, SOCS100 and 21 covariates. Using the RA sampling approach, 96 data located within the RA were used for training the ML models and 24 data (outside the RA) for validation. In the TA approach, the performance of the total area model was evaluated using a 5-fold cross-validation procedure. The results show that the use of previous covariates selec-tion, combined with the RA approach, allows to develop more accurate models. The models developed to predict SOCS100 presented both higher accuracy and transferability than those developed to predict SOCS30. The SOCS30 map was only generated to Urucu Block and the best performance was achieved using RT algorithm (R2 = 0.32). The RF algorithm generated the most accurate maps of SOCS100 for the Urucu and Juru ' a Blocks (R2 = 0.70 and R2 = 0.51, respectively).