Generalization of Vision-Language Models for Satellite-Based Urban Monitoring Across Southeast Asia

As Southeast Asian cities rapidly urbanize, scalable tools like vision-language models (VLMs) offer promising solutions for satellite-based urban monitoring. This study evaluates the generalization of RemoteCLIP across four cities—Da Nang, Hanoi, Ho Chi Minh City, and Bangkok—using zero-shot classification with standardized prompts. Top-1 accuracy drops from 85.2% in Da Nang to 68.6% in Bangkok, and cosine similarity falls from 0.82 to 0.70, indicating reduced performance in more informal, complex urban environments. Results show that urban morphology and prompt semantics significantly affect accuracy. We recommend localized prompt design and regional fine-tuning to improve model generalization. This work presents the first multi-city benchmark of RemoteCLIP for satellite imagery in real-world urban contexts.