Accurate acquisition of building age information is of significant importance for understanding urban conditions. However, existing building age prediction methods face several challenges, including insufficient consideration of multi-dimensional spatial features, imbalanced sample distribution, and inadequate extraction of deep semantic features. To address these issues, this study designed a progressive multi-task learning-based graph neural network (PMTGNN) framework. First, the study innovatively integrates multi-source geospatial data to construct a multi-dimensional feature indicator system encompassing building geometric scale, building visual scale, resident perception scale, and socioeconomic scale. Based on this, graph-structured node representations are constructed to achieve comprehensive integration of building-related features. Second, the study introduces two auxiliary tasks (building height estimation and building function classification) to enhance the model’s generalization capability under sample imbalance conditions through knowledge transfer. Third, the study designs a progressive learning strategy, which leverages the synergistic effects of expert modules and gating networks to guide the model in gradually learning high-level semantic features from low-level building geometric features. Finally, taking Changzhou as an empirical case study, the study trained the model using existing building age data and filled in the missing age information for 51,281 buildings. Experimental results show that, compared to traditional methods, PMTGNN achieves improvements of at least 6.9% and 5.4% in Accuracy and F1-score, respectively, and both the constructed building age prediction indicators and the designed multi-task learning modules play crucial roles in the prediction results. The PMTGNN framework enables accurate and reliable building age prediction and can provide support for urban planning.
Lu et al. (Mon,) studied this question.