Spatially distributed near surface air temperature data with high temporal resolution is a very important input parameter for several land surface models. Such data are often lacking because there are few traditional meteorological stations. It is of great significance in both theoretical research and practical applications to retrieve air temperature data from remote sensing observations. According to the surface energy balance and the radiance balance theories, the paper summarizes the retrieval methods of near surface air temperature from remote sensing images based on analysis of physical mechanism. The main methods are as follows: ① Single factor statistical method. It builds a relation model between bright temperature of thermal infrared channels or land surface temperature and land surface air temperature. ② Multiple factors statistical method. In relation models, at least two influence parameters are included to retrieve land surface air temperature by multiple regression analysis. ③ Artificial neural network method. Using multiple influence parameters as the input and the air temperature as the output, artificial neural network models are trained and then built to retrieve air temperature. ④ Temperature-vegetation index method. The approach is based on the hypothesis that vegetation canopy temperature approximates near surface air temperature and regression parameters used to compute air temperature are determined within a moving window around a location. ⑤ Surface energy balance methods. On the basis of physical approach, the air temperature is derived after calculation or measurement of the needed parameters for the energy balance equation. And then problems within air temperature retrieval are discussed in terms of spatial-temporal resolution, influence factors as well as models portability and utility considered. Finally the probable research fields in the future are proposed.