Detección de sistemas silvopastoriles en Uruguay mediante clasificación de imágenes satelitales Sentinel-2

The growth of the forestry sector in Uruguay has increased interest in integrating forestry and livestock activities, leading to an expansion of the area under silvopastoral systems (SPS). However, despite their implementation over several years, limited information is still available regarding their performance. The objective of this study was to identify and characterize SPS using multispectral satellite imagery from the Sentinel-2 satellite through a supervised classification based on the Random Forest (RF) algorithm. The aim was to distinguish SPS from other land cover types and to classify the different tree species present within the stands.

For the development of this study, the available SPS cartography was used, which provides relevant information on the stands, such as planted species, system structure, and year of establishment. In addition, Google Earth Engine (GEE) was employed, an open-access platform that offers a large archive of Sentinel-2 imagery, filtering tools, high processing capacity through Google servers, and a JavaScript-based console that facilitates the implementation of classification processes. As a result of the land cover classification, consumer’s accuracy (CA) of 0.87, producer’s accuracy (PA) of 0.86, and an F1-score of 0.87 were obtained for the SPS class. Overall, the classification achieved a global accuracy (GA) of 0.95 and a Kappa index of 0.94. In contrast, species-level classification showed a considerably lower performance, with a GA of 0.58 and a Kappa index of 0.50. Species-specific results were heterogeneous: Eucalyptus benthamii showed the best performance (F1 = 0.79), followed by E. grandis (0.60) and E. smithii (0.55), while E. globulus, E. maidenii, and E. dunnii presented F1 values below 0.55. Despite these limitations, the study demonstrates that it is feasible to classify SPS and their associated species using Sentinel-2 imagery. However, the structural complexity of these systems and the spatial resolution of the imagery hinder the discrimination of certain species.