Savannas are significant global carbon sinks that are increasingly threatened by land-use change. In Uruguay, subtropical wooded savannas cover approximately 100,000 ha and are the focus of conservation and climate change mitigation efforts. In this context, assessing woody species structure and carbon stock is important for the sustainable management of this ecosystem. This study integrates a global canopy height model (CHM), satellite-based multispectral data, and a global soil dataset to model diameter at breast height (DBH), mean height (H), maximum height (MAXH), and above-ground carbon stock (AGC) of woody species from subtropical wooded savannas using machine learning algorithms. Data from 64 plots of the National Forest Inventory of Uruguay were employed to train and validate four algorithms: random forest (RF), support vector machine (SVM), gradient boosting machine (GBM), and k-nearest neighbors (KNN). Height-related variables were primarily influenced by CHM features, whereas DBH and AGC required a complementary combination of data from all sources. The nested spatial 10-fold cross-validation showed that RF outperformed for DBH (R2 = 0.44, RMSE = 2.80 cm), H (R2 = 0.59, RMSE = 0.53 m), and AGC (R2 = 0.51, RMSE = 8.71 Mg ha-1), whereas GBM performed best for MAXH (R2 = 0.59, RMSE = 0.97 m). Nationwide maps revealed a right-skewed distribution of all variables, reflecting the predominance of small trees, possibly due to historical disturbances. The present results provide a valuable baseline for defining conservation strategies, monitoring ecosystem changes, and promoting the sustainable management of subtropical wooded savannas as carbon sinks.
扫码关注我们
求助内容:
应助结果提醒方式:
