Hydrogeophysics has become a major field of research in the past two decades, and time-lapse electrical resistivity tomography (ERT) is one of the most popular techniques to monitor passive and active processes in shallow subsurface reservoirs. Time-lapse inversion schemes have been developed to refine inversion results, but they mostly still rely on a spatial regularization procedure based on the standard smoothness constraint. We have applied a covariance-based regularization operator to the time-lapse ERT inverse problem. We first evaluated the method for surface and crosshole ERT with two synthetic cases and compared the results with the smoothness-constrained inversion (SCI). These tests showed that the covariance-constrained inversion (CCI) better images the target in terms of shape and amplitude. Although more important in low-sensitivity zones, we have observed improvements everywhere in the tomograms. Those synthetic examples also show that an error made in the range or in the type of the variogram model had a limited impact on the resulting image, which still remained better than SCI. We then applied the method to cross-borehole ERT field data from a heat-tracing experiment, in which the comparison with direct measurements showed a strong improvement of the breakthrough curves retrieved from ERT. This method could be extended to the time dimension, which would allow the use of CCI in 4D inversion schemes.