Abstract: Deep learning methods have been extensively explored in seismic inversion, and semi-supervised frameworks provide an effective solution to the bottleneck caused by the limited availability of well-log data. Conventional one-dimensional semi-supervised inversion approaches process seismic data trace by trace, neglecting the spatial correlations between adjacent traces, and often require an initial model to compensate for missing low-frequency information. In contrast, two-dimensional semi-supervised methods rely on large amounts of labeled seismic data for training, which in practice is challenged by the high cost of data acquisition and the difficulty of ensuring label quality. To make full use of limited well-log data while reducing dependence on large-scale labeled datasets, this study proposes a semi-supervised seismic impedance inversion framework based on one-to-two dimensional dilated convolution kernels. The method first pre-trains a one-dimensional convolutional neural network with well-log and seismic data to learn temporal feature representations of impedance; then, through a physics-constrained convolution kernel dilation, the network is expanded into a two-dimensional form to enable cross-dimensional feature transfer. Supervised fine-tuning is subsequently performed on sparse two-dimensional labeled data from the target area, integrating well-log information with unlabeled seismic data. To ensure the physical plausibility and stability of the transfer process, a spatially aware weight distribution strategy and a progressive training scheme are designed. Experimental results demonstrate that the proposed method outperforms both conventional 1D semi-supervised networks and standard 2D semi-supervised network approaches in terms of prediction accuracy and spatial continuity.