Improving tree cover estimates for fine-scale landscape ecology

Abstract

Mapping the presence of trees is an important tool for assessing tree-covered habitats, their changes, and calculating variables, like forest area and fragmentation. Despite the popularity of automated pattern recognition to make tree cover maps, their accuracy and precision are rarely tested or compared to more modest methods, like human-based pattern recognition to identify tree cover. Here, we test the performance of two computer-generated tree mapping products, the Global Change Forest database and the Carnegie Landsat Analysis System, against ground surveys and a human-made tree cover map created using Google Earth to hand digitize the presence and absence of trees in a diversified agricultural region in Costa Rica (934 km2). The human-made tree cover map properly classified 100% ground survey sites and explained 81% of the variance in percent of canopy cover values from the field. The Global Change Forest database misclassified 18 of 23 ground survey sites in deforested locations and explained 6% of the variance in percent of canopy cover values from ground surveys. The Carnegie Landsat Analysis System misclassified 9 of 23 ground survey sites in deforested locations and explained 38% of the variance in percent of canopy cover values from the field. Our results suggest that the Global Change Forest database overestimated tree cover by of 20% and the Carnegie Landsat Analysis System by 1%. We caution landscape ecologists working at fine spatial scales against using computer-generated tree cover, especially in the partially forested lands that increasingly cover the planet.