22 JUL 2025 (TUE) 11:05–11:25

Exploiting Multimodal Remote Sensing Images and GeoAI for Urban Surface Mapping

Mr LIU Rui 

( Supervisor: Prof Hongsheng Zhang )

Abstract:

Urbanization accelerates the conversion of natural surfaces into impervious cover, making accurate urban land-cover data essential for both macro-scale studies of urban expansion and climate change and micro-scale analyses of how urban form affects residents’ well-being. Remote sensing provides detailed, large-scale, multi-temporal observations, but single-sensor, like optical sensor based approaches struggle with adverse conditions and urban complexity. Integrating multisource data, such as synthetic aperture radar (SAR) with optical imagery leverages complementary strengths. Meanwhile, Geospatial Artificial Intelligence (GeoAI), particularly deep learning, can automatically extract intricate urban features. There are still many promising directions and challenging gaps in leveraging GeoAI and multisource data fusion for extracting urban surface.

This thesis addresses these gaps by proposing a general SAR-optical fusion framework in both real and complex domain, developing multimodal fusion strategies under few-sample constraints, and extending two-dimensional mapping into three dimensions. First, we introduce the SAR-Optical Fusion Transformer, a multi-level architecture that balances local and global feature extraction. By integrating both feature- and decision-level fusion, this model significantly improves classification accuracy across varied urban scales and outperforms single-level methods. Second, we preserve SAR’s complex-valued nature by designing a fully complex-valued Transformer incorporating complex attention mechanisms. Decision-level fusion with real-valued optical data, weighted by modality confidence, further enhances urban cover classification. Third, to alleviate the burden of extensive labeling, we employ self-supervised pretraining on large unlabeled datasets followed by fine-tuning with minimal annotations, achieving robust performance under few-sample conditions. Finally, we recognize that urban growth is inherently three-dimensional and develop an unsupervised, fully automated pipeline to extract building heights from high-resolution stereo imagery. While effective, this approach reveals challenges associated with shadows, building density, and height variability.

Together, these studies advance the integration of GeoAI and multimodal fusion for urban remote sensing, offering novel methodologies that improve urban surface monitoring and contribute to sustainable urban development.