Recent developments in predicting and interpreting seismoelectric (SE) signals suggest a great potential for studying near-surface hydrogeological properties, particularly in the vadose zone. Previous studies have revealed that the SE spectral ratios obtained from earthquake-triggered SE data contain valuable hydrogeological information concerning porous media (e.g., permeability, porosity, fluid viscosity, and salinity). This study introduces Multi-Channel SeismoElectric Spectral Ratios (MC-SESRs) by considering an active seismic source acting on the ground surface. The frequency- and saturation-dependent excess charge density is adopted to calculate the cross-coupling coefficients. Applying a supervised learning task based on a flat neural network, the so-called “broad learning (BL)” model, to map and extract the features of MC-SESRs data, we seek to determine the permeability and the water table depth. Our results indicate that (a) MC-SESRs are sensitive to the water table depth and permeability; (b) using more traces of SESRs data for inversion can increase accuracy; and (c) the changing water table can be rapidly determined by the MC-SESRs by resorting to the BL inverse model, and it can attain an excellent accuracy while disturbed by data noise and misspecified model parameters (e.g., porosity and permeability) with errors of up to 20%. The proposed MC-SESRs inversion has potential applications for non-invasive monitoring in shallow porous media (e.g., frost thawing and geothermal upwelling).