Weighted mean temperature (T_m) is a key parameter in the retrieval of atmospheric precipitable water (PW) from ground-based Global Positioning System (GPS). In order to improve the accuracy and reliability of the retrieval of PW in Hainan Island, temporal variation characteristics of T_m calculated based on Haikou radiosonde data during 2008-2010 and the relation with meteorological factors at Haikou station are analyzed. On this basis, based on radiosonde and surface observation data during 2008-2012, single-factor and multi-factor T_m regression equations and T_m regression models with day of year factor are established at Haikou, and the models are validated by using radiosonde and surface observation data during 2013-2014. Based on the local T_m regression models, the ground-based GPS PW retrieval of Haikou is performed from May to October 2012, and the retrieval accuracy is verified. The results show that: by comparison of the true T_m, the RMSE of single-factor and two-factor local T_m models are 2.000 and 1.978 K, superior to Bevis and constant model. The local model of T_m has good consistency with T_m calculated by radiosonde data. The GPS PW from single-factor T_m model exhibits much stronger correlations with radiosonde PW than GPS PW based on Bevis model, and the RMSE of GPS PW by single-factor T_m model is lower than that based on Bevis model. Compared with the multi-factor linear T_m model, GPS PW based on the T_m model with day of year factor has significantly improved accuracy. The local models could meet the accuracy requirements of the PW from ground-based GPS data of Haikou.