<p>Maintaining building shape across different levels of map generalization is essential for preserving spatial accuracy and readability. Differentiated map generalization algorithms can achieve higher shape-maintaining performance, and these algorithms require the strong support of accurate building shape recognition. Traditional building shape recognition methods based on template matching often suffer from high complexity, low accuracy, and poor generalization capability. Meanwhile, existing convolutional neural network (CNN)-based approaches lack optimized network structures tailored to building features, resulting in limited classification performance. To address these challenges, this study proposes an improved ZFNet-based classification method specifically designed for single-channel building images, which typically exhibit regular geometries and simple feature representations. To reduce complexity and avoid overfitting, the model simplifies the original ZFNet by removing its fifth convolutional layer. Batch normalization layers are added after each convolutional layer, kernel sizes are refined, and the activation function is adjusted for improved performance. Experimental results demonstrate that the improved ZFNet achieves a validation set accuracy of 98.53%, outperforming 14 comparative CNN models, and ablation experiments exhibit the effectiveness of the proposed structural improvements to ZFNet. Furthermore, it exhibits superior F1-scores across most shape categories, except for the T shape and the L shape, while maintaining a more lightweight architecture with lower computational cost. Compared with two variants of Graph Convolutional Neural Networks (GCN), the proposed method also shows better model performance, providing a robust foundation for addressing key challenges in building synthesis, such as feature preservation.</p>