Cloud phase recognition is important for cloud studies. Ice crystals correspond to physical process and properties that differ from those of liquid water drops. The angular polarization signature is a good mean to discriminate between spherical and nonspherical particles (liquid and ice phase, respectively). POLDER (Polarization and Directionality of Earth Reflectances) has been launched on the Japanese ADEOS platform in August 1996. Because of its multidirectional, multispectral, and multipolarization capabilities this new radiometer gives useful information on clouds and their influence on radiation in the shortwave range. The POLDER bidirectional observation capability provides the polarization signatures within a large range of scattering angles in three spectral bands centered on 0.443, 0.670, and 0.865 μm with a spatial resolution of 6.2 km¿6.2 km. These original features allow to obtain some information both on cloud thermodynamic phase and on cloud microphysics (size/shape). According to POLDER airborne observations, liquid cloud droplets exhibit very specific polarization features of a rainbow for scattering angles near 140¿. Conversely, theoretical studies of scattering by various crystalline particles and also airborne measurements show that the rainbow characteristics disappear as soon as the particles depart from the spherical shape. In the paper the POLDER algorithm for cloud phase classification is presented, as well as the physical principle of this algorithm. Results derived from the POLDER spaceborne version are also presented and compared with lidar ground-based observations and satellite cloud classification. This cloud phase classification method is shown to be reliable. The major limitation appears when thin cirrus clouds overlap the liquid cloud layer. In this case, if the cirrus optical thickness is smaller than 2, the liquid phase may be retrieved. Otherwise, the ice phase is correctly detected as long as cloud detection works. ¿ 2000 American Geophysical Union