It is vital in the study of crustal deformation to reduce system errors and enhance the accuracy of GPS coordinate time series. To eliminate system errors associated in the coordinate time series, which are related to the terrestrial reference frame (TRF) realization, we develop a recursive TRF realization strategy in regional GPS data processing. We have processed the whole set of CMONOC (Crustal Movement Observation Network of China) GPS data by the Bernese software, and employ the controlled datum removal (CDR) filter to solve the rank defect problem in the daily coordinate normal equations. On the recursive TRF realization of stations’ coordinates, we iteratively perform time series modeling with integral trajectory models, TRF realization with all continuous stations treated as “pseudo” fiducial stations, and frame alignment to the ITRF2014 with the 6-parameters Helmert transformation. We obtain the final coordinate time series through 3 times of iterations for the CMONOC data. Compared to the results derived from the conventional TRF realization strategy with the average root mean squares (RMS) being 1.98mm, 2.62 mm and 5.39 mm for east, north and up components, respectively, the average RMS earn significant reduction up to 30%, 43% and 16% in the first loop, with their quantities being 1.41 mm, 1.51 mm and 4.57 mm for east, north and up components, respectively, and negligible changes in the following 2 loops. In contrast to previous studies, our strategy is feasible in the processing of regional geodetic networks, and is concentrated on the TRF-related system errors without any pre-assumption and spatial limitation. In essence, our recursive strategy is to tighten the constraints for the CMONOC GPS stations in the TRF realization through the leaning of the barycenter of the processed geodetic network, thus inevitably loosening the constraints for other globally distributed stations, and slightly magnifying their RMS. On the whole, the north component of coordinates time series has a maximum RMS reduction, resulting in identical precision for both horizontal components, thus indicating that our strategy remedies the frame defects stemming from the extremely uneven distribution of the reference network, and retrieves the “real” precision of GPS observations. The insignificant RMS reduction on the vertical component may be attributed to insufficient time series modeling. Our recursive TRF realization strategy can benefit the velocity estimation for campaign stations.