Influence of Protected Areas in the Carbon Cycle: A Machine Learning Approach for Future Forecasting

Abstract Protected Areas (PAs) are destined to the conservation of natural resources, whose quantity and quality are menaced by Climate Change and anthropogenic pressure. The objective of this research was to quantify and forecast the Carbon Absorption of PAs, attending these factors. We used the Net Primary Productivity (NPP) and implemented machine learning algorithms, as Random Forests (RF), Gradient Boosting Trees (GBT), and Multilayer Perceptrons (MLPNN), to forecast it in four differentiated PAs of Galicia (NW Spain): the Central Massif, the Sil Canyons, Fragas do Eume, and the Tambre River. Two testing stages were carried, one in the areas where the model was trained, and another in the whole territory. Finally, we set several scenarios based on projections SSP2-4.5 and SSP 5-8.5 in and land use changes. GBT was the most accurate algorithm, with a Root Mean Squared Error (RMSE) of 0.05 kgC/m 2 (5.7% of the average NPP) and correlation of 0.9. RF obtained an error of 0.07 and correlation of 80%, and MLPNN 0.06 and 86%, respectively. The GBT obtained a RMSE of 0.04 kgC/m 2 and R 2 of 0.95 in the first test. These results were worsened in the second test, with an RMSE of 0.09 kgC/m 2 and 71% of correlation. For the SSP2-4.5 scenarios, a decrease around 7% can be expected, barely influenced by the land use. On the other hand, the SSP5-8.5 is expected to record a decrease of 5% in the NPP, with no significant differences between the land uses, but with a significant trend. All the pilot sites stick to this trend, except for the Tambre River, which was forecasted to record a slight increase of the NPP (~ 3%). The prediction and analysis of future scenarios can help the management of the territory, focusing on the mitigation of the effects of Climate Change in PAs. Graphical Abstract