| Sumari: | Forest floor or mulch is the carbon stock that regulates most of the functional processes of forest ecosystems, directly influencing soil fertility and site productivity. The forest floor carbon content is highly variable in space and time; therefore, obtaining accurate assessment of the carbon contained in this stock represents a significant methodological challenge at any scale. In this study, four spatial modeling methods were compared to map forest floor carbon content in a temperate forest. The methods were ordinary kriging, generalized linear model, generalized additive model and random forest. Carbon content estimates were made for 2013 and 2018. The predictor variables represent the spatial, canopy and topographic structure present in the study area. All models were evaluated by ten-fold cross validation and the mean absolute error, root mean square error and the coefficient of determination (R2) were determined. The performance of the methods was, in decreasing order, random forest, generalized additive model, generalized linear model and ordinary kriging. The ordinary kriging method reflected the degree of spatial dependence of carbon content, but the spatial estimates were unrealistic (R2 ≤ 0.35). Generalized additive model and generalized linear model showed good performance (R2 ≥ 0.70), but higher overestimation; random forest obtained the best fit (R2 ≥ 0.86) to model carbon content in both years evaluated. It is concluded that random forest is a promising method with high potential to improve estimates of forest floor carbon at landscape scale.
|
|---|
