Algorithms are developed to derive the sea surface air temperature and specific humidity based on AMSU measurements from the NOAA polar orbiting environmental satellites (POES). The air temperature retrieval uses AMSU-A data, while the specific humidity retrieval uses both AMSU-A and AMSU-B observations. The retrieval algorithms are developed with the neural network approach. The training datasets are constructed using a full year of co-located AMSU and buoy/ship data covering the ocean surface from 70S to 70N, in which all four seasons and various weather conditions throughout the year are included to have a global and all-weather representation. Comparison with independent global buoy/ship observation shows that the root-mean-square (RMS) errors of the retrievals are 2.10°C for air temperature and 1.22 g/kg for specific humidity. Additional validation using high-quality research ship measurement shows RMS errors of 1.60°C for air temperature and 1.16 g/kg for specific humidity. Examination indicates that the larger global RMS errors from buoy/ship comparisons are partly due to errors in in-situ data. Global long-term sea surface air temperature and specific humidity since 2001 have been derived for POES satellites and the data are available in a common-date-model format. They can be used as boundary conditions for numerical modeling, in the computation of air-sea turbulent heat fluxes, as well as other applications.