Land subsidence provides important information about the spatial and temporal changes occurring in the subsurface (e.g. groundwater levels, geology, etc.). However, sufficient subsidence data are difficult to obtain using only one sensor or survey, often resulting in a tradeoff between spatial resolution and temporal coverage. This study aims to estimate the high spatio-temporal resolution land subsidence by using a kernel-based vector data fusion approach between annual leveling and monthly subsidence monitoring well data, while invoking an invariant relation of subsidence information. Subsidence patterns and processes can be identified when spatio-temporal fusion of sensor data are implemented. In this subsidence investigation in Yunlin and Chunghua counties, Taiwan, the root mean square error (RMSE) is 0.52 cm in the fusion stage, and the mapping RMSE is 0.53 cm in the interpolation. The fused subsidence data readily show that the subsidence hotspot varies with time and space. The subsidence hotspots are in the western region during the winter (related to aquaculture activities) but move to the inland areas of Yunlin County during the following spring (related to agricultural activities). The proposed approach can help explain the spatio-temporal variability of the subsidence pattern.