In this study, we apply three-dimensional inversion to spatially-dense Self-potential (SP) data in time-series. The fluctuation of SP signals induced by the welling when excavating a borehole in a tunnel at the depth of 300m was observed at the Mizunami Underground Research Laboratory (MIU), the central Japan. We analyze the time series of SP signals and welled-out water flux data for the estimation of hydrogeological characterization surrounding the tunnel. The SP data was contaminated by the noise due to the leakage current from the railway that runs several km away from the site. We first remove the noise using independent component analysis (ICA) prior to the inversion. The flux data is also up-sampled prior to the analysis to match the sampling rate with SP. We compare the estimated hydraulic conductivity structure with those by the other method, such as hydraulic tomography. The hydraulic conductivity structures from our method agree well with those by other methods quantitatively even though the raw SP data is noisy and the number of flux data is insufficient. SP data, in general acquired not so frequently, could become an input to turn the inversion over-determined if sampled in a time-series.