Abstract: Deep learning methods have been extensively applied in seismic inversion, and semi-supervised frameworks effectively overcome the bottleneck of limited log data. Conventional 1D semi-supervised inversion approaches process seismic data trace by trace, neglecting spatial correlations between adjacent traces and often requiring an initial model to compensate for missing low-frequency information. In contrast, 2D methods rely on large amounts of labeled seismic data for training, which is restricted by the high cost of data acquisition and the difficulty of ensuring labeling quality. To make full use of limited log data and reduce dependence on large-scale labeled datasets, this study proposes a semi-supervised seismic impedance inversion framework based on 1D-to-2D dilated convolution kernels. The method first pre-trains a 1D convolutional neural network using log and seismic data to learn temporal feature representations of impedance. Subsequently, through physics-constrained kernel dilation, the network is expanded into a 2D form to enable cross-dimensional feature transfer. Supervised fine-tuning is then performed on sparse 2D labeled data from the target area to integrate well-log information with unlabeled seismic data. To ensure the physical plausibility and stability of the transfer, this study designs a spatially aware weight distribution and a progressive training strategy. Experimental results on synthetic and field data demonstrate that the proposed method maintains consistency at well sites and achieves high inversion accuracy and good spatial continuity. It exhibits strong stability and engineering application potential with limited log data.