Abstract:

Building group pattern recognition is a key issue in fields such as map automatic generalization and urban spatial understanding. To address the limitations of existing methods in terms of pattern coverage, threshold subjectivity, model generalization capability, and reliance on labeled samples, this paper proposes a recognition model that integrates deep graph infomax (DGI) and a multilayer perceptron (MLP), aiming to explore a high-accuracy, strongly generalized approach for recognizing multiple building group patterns under limited labeled samples. First, building groups are partitioned and geometric models are constructed based on the road network and the minimum spanning tree of buildings. Next, individual building features and global group features are extracted, and the DGI model is introduced for unsupervised graph representation learning. By maximizing the mutual information between graph-level and node-level representations, the model effectively captures the complex topological dependencies within groups, generating discriminative low-dimensional graph embeddings. Finally, the graph embeddings and global features are fused into a unified feature vector, which is fed into an MLP classifier for end-to-end pattern discrimination, enabling automatic recognition of four typical building group patterns: linear, curved, grid-like, and irregular. The experimental results indicate that the highest recognition accuracy of the proposed method on the test set reaches 99.20%. Even with a significant reduction in the number of training samples (e.g., using only 20% of the labeled data), the model can still maintain a recognition accuracy of 97.85% along with a high recall rate, demonstrating superior robustness and data utilization efficiency compared to the baseline models.

Key words: building group, pattern recognition, deep graph infomax, multilayer perceptron, deep learning