Surface water is experiencing significant changes due to anthropogenic activities and climate change. Recent studies have explored long-term surface water dynamics with the development of multi-source remote sensing data and mapping technology. However, cost-effective and time-saving reliable references for mapping surface water from remote sensing images at multiple times are still lacking. Further, negligence of abnormal changes in time-series surface water maps seriously affects surface water mapping accuracy. In this study, we propose a high-accuracy method for long-term mapping of surface water by automatic update of training samples and temporal consistency modification in surface water sequences. Taking the surface water in Huizhou from 1986 to 2020 as a case study, the method is applied and tested. Temporal training samples were updated through Robust Mahalanobis distances and statistical filtering. Combined with the random forest algorithm and the Google Earth Engine platform, annual surface water maps were generated. A method for temporal consistency detection in moving window and abnormal changes modification rule with intra-annual information are proposed to improve annual surface water mapping accuracy. The accuracy of surface water dynamics and abnormal water modifications were quantified as 89.8% and 95.7%, respectively. With our proposed method, the resultant maximum annual water areas of Huizhou were detected, with a statistically significant upward trend and net increase of 103.1 km² over 35 years. Reservoir construction and aquaculture expansion constitute the main sources of increased waters in this case study. It is demonstrated that our proposed method offers important advantages in terms of the accuracy of annual surface water maps and especially change monitoring.