Using spatial models of temporal tree dynamics to evaluate the implementation of EU afforestation policies in rangelands of SW Spain

Iberian silvopastoral systems known as dehesas in Spain and montados in Portugal are undergoing a spatially polarized process by which many of the main areas lack tree recruitment, whereas marginal lands suffer abandonment and shrub encroachment. These ongoing processes should be considered when designing afforestation measures and policies. We analyzed the temporal tree dynamics in 800 randomly selected plots of 100 m radius in dehesas and treeless pasturelands of Extremadura by comparing aerial images taken in 1956 and 2012. Based on this data, spatial models that predict areas prone to undergo or lack natural tree recruitment were developed using three data mining algorithms: MARS (Multivariate Adaptive Regression Splines), Random Forest (RF) and Stochastic Gradient Boosting (TreeNet, TN). A number of 51 candidate environmental, physical and land use and cover spatial variables were used as predictors in models, from which the main 15 were selected. The statistical models developed were deployed to the spatial context of the rangelands in Extremadura and, separately, to the afforested areas performed under the UE First Afforestation of Agricultural Land program between 1992 and 2013. The percentage of area predicted as prone to tree recruitment was calculated in each case. The three data mining algorithms used showed high fitness and low misclassification rates. Although the drivers and patterns of the different models were similar, outstanding differences were observed among models attending the area prone to tree recruitment. A model ensemble was also produced as a map of agreement reflecting the majority vote among models. Despite these differences, when maps of the model results were related to the afforested surfaces, the three algorithms pointed to the similar conclusion, i.e., the afforestations performed in the studied rangelands barely discriminated between areas that already showed or lacked natural tree regeneration. In conclusion, data mining technics are suitable to develop high-performance spatial models of vegetation dynamics. These models are useful to help policy design, decision-making and assessment about the implementation of afforestation measures and could be used to improve the spatial targeting of future programs