Terrestrial biospheric models (TBMs) are used to extrapolate local observations and process-level understanding of land-atmosphere carbon exchange to larger regions, and serve as predictive tools for examining carbon-climate interactions. Understanding the performance of TBMs is thus crucial to the carbon cycle and climate science communities. In this study, we present and assess an approach to evaluating the spatiotemporal patterns, rather than aggregated magnitudes, of net ecosystem exchange (NEE) simulated by TBMs using atmospheric CO2 measurements. The approach is based on statistical model selection implemented within a high-resolution atmospheric inverse model. Using synthetic data experiments, we find that current atmospheric observations are sensitive to the underlying spatiotemporal flux variability at sub-biome scales for a large portion of North America, and that atmospheric observations can therefore be used to evaluate simulated spatiotemporal flux patterns as well as to differentiate between multiple competing TBMs. Experiments using real atmospheric observations and four prototypical TBMs further confirm the applicability of the method, and demonstrate that the performance of TBMs in simulating the spatiotemporal patterns of NEE varies substantially across seasons, with best performance during the growing season and more limited skill during transition seasons. This result is consistent with previous work showing that the ability of TBMs to model flux magnitudes is also seasonally-dependent. Overall, the proposed approach provides a new avenue for evaluating TBM performance based on sub-biome-scale flux patterns, presenting an opportunity for assessing and informing model development using atmospheric observations.