Abstract
© 2025 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS)Global urban scene classification is a crucial technology for global land use mapping, holding significant importance in driving urban intelligence forward. When applying datasets constructed from urban scenes on a global scale, there are two serious problems. Due to cultural, economic, and other factors, style differences exist in scenes across different cities, posing challenges for model generalization. Additionally, urban scene samples often follows a long-tailed distribution, complicating the identification of tail categories with small sample volumes and impairing performance under domain generalization settings. To tackle these problems, the Uncertainty-aware Domain Generalization urban scene classification (UADG) framework is constructed. For mitigating city-related style difference among global cities, a city-related whitening is proposed, utilizing whitening operations to separate city unrelated content features and adaptively preserving city-related information hidden in style features, rather than directly removing style information, thus aiding in more robust representations. To tackle the phenomenon of significant accuracy decline in tail classes during domain generalization, estimated uncertainty is utilized to guide the mixture of experts, and reasonable expert assignment is conducted for hard samples to balance the model bias. To evaluate the proposed UADG framework under practical scenario, the Domain Generalized Urban Scene (DGUS) dataset is curated for validation, with a training set comprising 42 classes of samples from 34 provincial capitals in China, and test samples selected from representative cities across six continents. Extensive experiments have demonstrated that our method achieves state-of-the-art performance, notably outperforming the baseline GAMMA by 9.79% and 7.42% with average OA and AA metric on the unseen domains of DGUS, respectively. UADG greatly enhancing the automation of global urban land use mapping.
NETL
NSTL
Elsevier