全球智库信息集成服务系统

Acquiring forest resources information for tropical developing countries is challenging due to financial and logistical constraints. Yet, this information is critical for enhancing management capability and engaging in international initiatives such as Reducing Emissions from Deforestation and forest Degradation (REDD +). The use of multi-source inventories (i.e., remote-sensing, field, and other data) in integrated models has shown increasing promise for accurately estimating forest attributes at lower costs. In this study, we compared the use of Landsat 8 OLI versus RapidEye satellite imagery in four modeling approaches (generalized linear model (GLM), generalized additive model (GAM), k-Nearest Neighbors (k-NN), Random Forests), with and without auxiliary information (e.g., soils characteristics, distance to roads, etc.) to estimate percent canopy cover by pixel for an ~ 1,000,000 ha area in Zambia. We derived plot-level canopy cover as the dependent variable, using field-measured data collected according to current National Forest Inventory (NFI) protocol. Using cross-validation statistics, Landsat 8 OLI exhibited better results than RapidEye across modeling approaches likely due to the additional short-wave infrared bands which consistently improved model performance (average root mean squared prediction error = 10.1% versus 11.0%). The GAM approach was more precise, though more challenging to fit. For both remote sensing data sources and all modeling approaches, other auxiliary information improved the model; soil variables were commonly selected for inclusion using a Genetic Algorithm. Using a binomial GAM with Landsat 8 OLI and soil variables, and by applying the current FAO forest/non-forest definition (i.e., canopy cover > 10% for a 0.5 ha area), we estimated the total forest area as 758,100 ha (95% bootstrapped confidence interval of ± 3,953 ha). Overall, our research indicates that sufficiently accurate forest area estimates for Zambia can be obtained using canopy cover GAM models that incorporate NFI data and freely-available remote sensing imagery and soil information.