Multi-class classification of parcel-level morphological urban change using a Siamese network: Evidence from Portland, Oregon

Urban regeneration is crucial for cities adapting to growing populations and shifting socio-economic conditions. The regeneration process is often monitored through field surveys, permit records, or medium-resolution remote sensing, but these sources can be costly, inconsistent, or insufficient for identifying parcel-level morphological change. This study develops a Siamese network-based framework for classifying morphological urban change during regeneration from paired high-resolution aerial images. Using 2000 residential parcels from Portland, Oregon, we classify four operational change types: No Change, New Development, Redevelopment, and Demolition. We compare Siamese network variants using ResNet, UNet, and YOLO backbones, including a UNet variant with a local similarity attention module. Five-fold cross-validation shows that the UNet-based models achieve the most consistent performance, with an overall accuracy above 85%. A transfer-learning test using data from Charlotte, North Carolina, suggests that direct cross-city transfer is limited, but local fine-tuning substantially improves performance. This approach provides a cost-effective, replicable tool for parcel-level morphological change analysis, facilitating rapid, reliable assessment of urban regeneration even in data-limited environments.