The usefulness of eddy covariance for understanding terrestrial carbon exchange has been hampered by uncertainty over the magnitude and causes of a systematic underestimation of CO2 efflux on calm nights. We combined in situ measurements of the temperature, wind and CO2 profile with nocturnal Land Surface Temperature (LST) imagery from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) to infer the patterns of cold air drainage in an Amazonian tropical forest. The meteorological tower was located on a flat plateau that sloped gently to the southwest. The vertical profile from 0 to 64 m above ground level was divisible into two air parcels at night: a warm, low CO2, upper parcel and a cool, high CO2, lower parcel that was stable with respect to the upper parcel. The nocturnal wind direction of the upper parcel was from the east, a pattern that is consistent with the general circulation, whereas the wind direction of the lower parcel was from the northeast, a pattern that implies drainage down the local topographic gradient. The nocturnal patterns of LST were closely related to local topography and land use. In general, a nearby river was warm, gullies were cold, plateau centers were cold, stream drainages were cold, pastures were particularly cold, and upper slopes and plateau edges were warm. The in situ temperature and wind observations, combined with the observed relationship between elevation and nocturnal LST and the occurrence of warm thermal belts extending inward from the edges of plateaus, imply that cold air drainage occurs on clear nights. The slope of the relationship between LST and elevation varied between nights, indicating that the degree of thermal stratification, and possibly the extent of cold air drainage, also varied. The night-to-night variation in stratification across the landscape was correlated with the vertical temperature gradient at the tower but not the above-canopy friction velocity (u*). Criteria associated with vertical temperature gradients may prove better than u* for screening nocturnal eddy covariance observations to eliminate periods that underestimate CO2 flux.