Wetlands not only affect the local hydrology and ecosystems, but also regulate the conditions of human-environment. However, the availability of accurate wetland data remains a key challenge in wetland research. This study attempts to address this problem through a novel mapping framework that is based on the Google Earth Engine (GEE), feature optimization, and the random forest (RF) model (GFORF). This framework was built to map high-accuracy wetland data on the headwaters of the Brahmaputra, Ganges, and Indus rivers (HBGIR) in the western Tibetan Plateau (TP). Four time periods were examined: 1990, 2000, 2010, and 2017. Our results showed that the overall accuracy for the acquired wetland data was 82.73%, 83.16%, 82.47%, and 88.14% in 1990, 2000, 2010, and 2017, respectively. Furthermore, the feature optimization results showed that the spectral indices feature was the main contributor to the accuracy of wetland mapping, with the highest value being 26.9%. The seasonal factors, surface reflectance, auxiliary data, and texture contributed 21.8%, 21.6%, 21.5%, and 8.1%, respectively. Combining the seasonal features and auxiliary data of distances to rivers significantly improved the mapping accuracy of the wetlands by approximately 14%, 24%, 11%, and 10% in 1990, 2000, 2010, and 2017, respectively. In addition, our analysis showed that the wetland areas in the HBGIR amounted to 5177.39 km2, accounting for 5.82% of the total area. Over the 30-year observation period, the overall consolidation of the wetlands was characterized by a slight expansionary phase, with an average increase of 0.16% per year from 1990 to 2017. As a result of the improvement in the accuracy of wetland mapping in alpine areas, the change dynamics of wetlands was revealed, which provides justification for implementing ongoing wetland ecological services and protection measures.