VISSR (visible and infrared spin-scan radiometer) atmospheric sounder (VAS) radiances and conventional surface temperature and dewpoint data are used in several combinations within a regression approach to determine the optimum resolution and accuracy of precipitable water (PW) fields retrieved from satellite observations. Point retrievals at radiosonde satations are used to determine the numerical accuracy of each retrieval technique, and image sequences of the retrieved PW fields are used to determine the temporal stability and spatial coherence of mesoscale PW features. VAS channels 5, 6, 7, and 8 (at 13, 4.5, 12, and 11 μm) and the surface dewpoint contribute the most information to regression- based retrievals of PW. The most accurate PW retrievals are obtained when radiances are averaged to a resolution of 15 to 60 km. A physical ''split window'' approach provides better PW estimates than regression when only the 11- and 12-μm VAS channels are available (as when the satellite operates in a multispectral image mode) or when radiosonde-based training is limited to only one time period.