Retrievals of cloud optical thickness and particle size are now being produced operationally from measurements of reflected solar radiation made by MODIS aboard NASA’s Terra satellite. These global, high-resolution estimates are useful in climatological and process studies only if we know how much faith to put in the retrievals. Current understanding of the uncertainty is limited in two ways. First, most uncertainty estimates have focused on the average uncertainty over a wide variety of conditions, when in fact the uncertainty depends on cloud optical properties, viewing and illumination geometry, the state of the atmosphere, and other factors. Secondly, uncertainty calculations are typically based on feeding radiances from perturbed radiative transfer calculations into the retrieval scheme, which convolves the radiative impact of uncertainty sources with the retrieval behavior.

We describe a new approach for assessing the uncertainty in retrievals of optical thickness and particle size that decouples the inherent sensitivity of the retrievals from the physical factors affecting reflectance. We separate the uncertainty into its components: the sensitivity of the derived parameters to changes in cloud-top reflectance, and the sensitivity of cloud-top reflectance to changes in the optical properties of the atmosphere and underlying surface. The reflection sensitivities depend on the details of the radiative transfer calculations, and must be computed separately for each potentially uncertain aspect of the cloud, atmosphere, and surface. The retrieval sensitivities, however, are entirely determined by the forward calculations. This allows us to identify those portions of the retrieval space that are naturally sensitive to changes in reflectance, without regard to what exactly causes those changes.

We show how the retrieval sensitivity varies with optical properties and viewing and illumination geometry, illustrate how uncertainties can be calculated from these sensitivities under a wide variety of circumstances, and show how the uncertainty in statistical measures of cloud properties can be computed in an operational setting.