As part of the Coastal Ocean Dynamics Experiment, 58 infrared satellite images were analyzed for the magnitude and source of their errors in predicting sea surface temperature. Comparisons with near-surface thermistors on moorings and differences between images from consecutive days showed that a two-channel atmospheric correction using the 3.7-μm and the 11-μm channels was ineffective in reducing errors. Large systematic errors, which have the form of a constant over an image, were found in both channels. A simple viewing angle correction reduced fluctuations in the area-mean 11-μm channel temperature from 0.74¿ to 0.45¿C; a satisfactory correction could not be found for the large fluctuations in the 3.7-μm channel, which appears to be degraded by reflected or scattered solar radiation. By taking account of the known dependence of atmospheric effects on viewing angle geometry, it was possible to isolate the primary causes of error, to estimate typical magnitudes, and to correct for them. The 11-μm channel temperature error from the northern California atmosphere was found to be nearly constant in time over a 100-day period, but to vary slowly with distance from the coast.