A transferable framework for roof vegetation mapping across tropical cities based on open-source multimodal data and morphology-aware UDASS

Green roofs play a vital role in urban sustainability, offering substantial environmental, ecological, and social benefits. However, automated roof vegetation extraction mostly relies on costly high-resolution data, with limited investigations in Global South cities. Existing methods insufficiently fuse multimodal priors and overlook domain adaptation, thereby constraining extraction accuracy and cross-regional generalization. This work presents a transferable open-source data-driven framework for cross-domain multimodal extraction and analysis of roof vegetation. Specifically, we introduce a multimodal transformer encoder to fuse Google imagery, Sentinel-2-derived products and OpenStreetMap building footprints at the feature level to improve extraction accuracy. We then propose an unsupervised domain-adaptive semantic segmentation (UDASS) model with morphology-aware pseudo-label correction to improve cross-domain generalization cost-effectively. Validated across six Southeast Asian cities, our framework achieves an F1-score of 81.71% for supervised source domain extraction and up to 70.85% for UDASS. Further validation on four additional cities spanning East Africa, South America, West Asia, and South Asia verifies its strong cross-regional generalization ability. Finally, approximately 297 ha of roof vegetation are detected across the six Southeast Asian cities, mainly on low-to-mid-rise buildings with distinct regional disparities and diverse vegetation intensity. This study delivers the first large-scale dataset on roof vegetation and green roofs for tropical Southeast Asia, establishes a cost-effective and scalable multimodal paradigm for cross-regional monitoring, and supports evidence-based green roof planning and analysis.