Quantifying uncertainty from large-scale model predictions of forest carbon dynamics

Linking environmental computer simulation models and geographic information systems (GIS) is now a common practice to scale up simulations of complex ecosystem processes for decision support. Unfortunately, several important issues of upscaling using GIS are rarely considered; in particular scale dependency of models, availability of input data, support of input and validation data, and uncertainty in prediction including error propagation from the GIS. We linked the biogeochemical Forest-DNDC model to a GIS database to predict growth of Eucalyptus globulus plantations at two different scales (similar to 0.045 ha plot(-1) scale and similar to 100 ha grid(-1) scale) across Victoria, in south-eastern Australia. Results showed that Forest-DNDC was not scale dependent across the range of scales investigated. Reduced availability of input data at the larger scale may introduce severe prediction errors, but did not require adjustment of the model in this study. Differences in the support of input and validation data led to an underestimation of predictive precision but an overestimation of prediction accuracy. Increasing data support, produced a high level of prediction accuracy ((-)e%), but a medium level of predictive precision (r(2)=0.474, ME=0.318) after statistical validation. GIS error contribution could be detected but was not readily or reliably quantified. In a regional case study for 2653 ha of E. globulus plantations, the linked model GIS system estimated a total standing biomass of 95 260 t C for mid-2003 and a net CO2 balance of -45 671 t CO2-C yr(-1) for the entire year of 2002. This study showed that regional predictions of forest growth and carbon sequestration can be produced with greater confidence after a comprehensive assessment of upscaling issues.