This study proposes a hydraulic tomography neural network (HT-NN) based on a convolutional encoder-decoder neural network (DenseNet) combined with a head data sampling strategy to estimate hydrogeological parameter fields. Numerical experiments demonstrate that HT-NN effectively captures the spatial characteristics of both hydraulic conductivity ( \({\text{K}}\) ) and specific storage ( \({S_s}\) ) fields and achieves high accuracy in delineating subsurface heterogeneity. By selecting late- and early-time head data to construct the input matrix, HT-NN substantially improves parameter estimation while significantly reducing computational time. Compared to the successive linear estimator (SLE), HT-NN achieves more accurate parameter estimation and reduces computation time from 28.5 h to 0.76 s. The simulated heads derived from HT-NN’s estimated parameter fields closely match the reference heads across all experiments. Additionally, adopting a smaller input matrix with a simplified encoder-decoder structure greatly enhances computational efficiency while maintaining estimation accuracy. These findings demonstrate the potential of HT-NN as an efficient and reliable alternative for estimating hydrogeological parameters in heterogeneous aquifer systems.
