Modeling network of research sites for monitoring carbon flows by Random Forest method

Environmental observing networks provide information for understanding and predicting the spatial and temporal dynamics of Earth biophysical processes. The optimization of resources for large-scale environmental monitoring activities is required. The paper describes and then tests spatial structure of Tyumen region research sites network. The network is based on principles of landscape approach, taking into account cost minimization. At the baseline of research, two testing sets of 40 and 105 points were determined. Proposed locations were evaluated using Random Forest (RF) method. The study accomplished in two stages for each test set. At the first stage, the model was trained; its capacity and indicators of additional diagnostics were studied. At the second stage, the trained model was used to predict the points formed of regular grid covering entire territory of this region (544 points). In conclusion, the obtained results were compared with similar point sets of the same volume but generated randomly. Primary Productivity Gross (GPP) was chosen as predictable variable because it is one of the major complex environmental indicators associated with carbon production in this area. The ability of an area to absorb or produce carbon is one of the main parameters that determine climate processes. As independent variables characterizing geosystemic processes, a set of indicators associated with climate, terrain parameters, and variability of soil resources has been selected. The problem was solved using Forest-Based Classification and Regression tool from Spatial Statistics—Modeling Spatial Relationships toolkit of ArcGIS Pro software package. As the result of the study, a high forecast accuracy and reliability for both approaches to research sites locations was obtained. The study was based on open source data.