We present accurate photometric redshifts in the 2-deg 2 COSMOS field. The redshifts are computed with 30 broad, intermediate, and narrow bands covering the UV (GALEX), Visible-NIR (Subaru, CFHT, UKIRT and NOAO) and mid-IR (Spitzer/IRAC). A χ 2 template-fitting method (Le Phare) was used and calibrated with large spectroscopic samples from VLT-VIMOS and Keck-DEIMOS. We develop and implement a new method which accounts for the contributions from emission lines ([O II], Hβ, Hα and Lyα) to the spectral energy distributions (SEDs). The treatment of emission lines improves the photo-z accuracy by a factor of 2.5. Comparison of the derived photo-z with 4148 spectroscopic redshifts (i.e. ∆z = z s − z p ) indicates a dispersion of σ ∆z/(1+zs) = 0.007 at i + AB < 22.5, a factor of 2 − 6 times more accurate than earlier photo-z in the COSMOS, CFHTLS and COMBO-17 survey fields. At fainter magnitudes i + AB < 24 and z < 1.25, the accuracy is σ ∆z/(1+zs) = 0.012. The deep NIR and IRAC coverage enables the photo-z to be extended to z ∼ 2 albeit with a lower accuracy (σ ∆z/(1+zs) = 0.06 at i + AB ∼ 24). The redshift distribution of large magnitude-selected samples is derived and the median redshift is found to range from z m = 0.66 at 22 < i + AB < 22.5 to z m = 1.06 at 24.5 < i + AB < 25. At i + AB < 26.0, the multi-wavelength COSMOS catalog includes approximately 607,617 objects. The COSMOS-30 photo-z enable the full exploitation of this survey for studies of galaxy and large scale structure evolution at high redshift.

Aims. We present and release photometric redshifts for a uniquely large and deep sample of 522286 objects with i AB ≤ 25 in the Canada-France Hawaii Telescope Legacy Survey (CFHTLS) "Deep Survey" fields D1, D2, D3, and D4, which cover a total effective area of 3.2 deg 2 . Methods. We use 3241 spectroscopic redshifts with 0 ≤ z ≤ 5 from the VIMOS VLT Deep Survey (VVDS) as a calibration and training set to derive these photometric redshifts. Using the "Le Phare" photometric redshift code, we developed a robust calibration method based on an iterative zero-point refinement combined with a template optimisation procedure and the application of a Bayesian approach. This method removes systematic trends in the photometric redshifts and significantly reduces the fraction of catastrophic errors (by a factor of 2), a significant improvement over traditional methods. We use our unique spectroscopic sample to present a detailed assessment of the robustness of the photometric redshift sample. Results. For a sample selected at i AB ≤ 24, we reach a redshift accuracy of σ ∆z/(1+z) = 0.029 with η = 3.8% of catastrophic errors (η is defined strictly as those objects with |∆z|/(1 + z) > 0.15). The reliability of our photometric redshifts decreases for faint objects: we find σ ∆z/(1+z) = 0.025, 0.034 and η = 1.9%, 5.5% for samples selected at i AB = 17.5-22.5 and 22.5-24 respectively. We find that the photometric redshifts of starburst galaxies are less reliable: although these galaxies represent only 22% of the spectroscopic sample, they are responsible for 50% of the catastrophic errors. An analysis as a function of redshift demonstrates that our photometric redshifts work best in the redshift range 0.2 ≤ z ≤ 1.5. We find an excellent agreement between the photometric and the VVDS spectroscopic redshift distributions at i AB ≤ 24. Finally, we compare the redshift distributions of i selected galaxies on the four CFHTLS deep fields, showing that cosmic variance is still present on fields of 0.7-0.9 deg 2 . These photometric redshifts are made publicly available at http://terapix.iap.fr (complete ascii catalogues) and http://cencos.oamp.fr/cencos/CFHTLS/ (searchable database interface).Key words. galaxies: distances and redshifts -galaxies: photometry -methods: data analysis Article published by EDP Sciences and available at http://www.edpsciences.org/aa or http://dx