Fire disturbance is poorly simulated in dynamic global vegetation models (DGVMs). The occurrence of fire and its effect on vegetation is often prescribed. Process-based models of fire activity are a better approach, although their complexity and parametrisation is an issue. In the current work, Earth observation (EO) data is used to better understand a coupled DGVM and fire model through probabilistic calibration. The mehodology outlined is general, and results in the model improving its predictive capabilities as the EO data constrains model parameters, provided the model is able to reproduce the observations. The data used is fundamentally burnt area derived from MODIS data, and only a handful of parameters controlling ignition patterns and rate of spread are considered. Poor agreement between calibrated model and observations is found in areas where the DGVM predicts unrealistic vegetation, which results in the fire model not being able to spread fires to match the observations. In areas where the DGVM simulates vegetation well, we find good agreement between simulations and observations.