Concetti Chiave
RSBuilding proposes a unified framework for building extraction and change detection in remote sensing imagery, enhancing generalization and task complementarity.
Sintesi
The content introduces RSBuilding, a model designed for building extraction and change detection tasks in remote sensing imagery. It emphasizes the importance of interpreting buildings for urban planning, macroeconomic analysis, and population dynamics. The model aims to unify these tasks within a comprehensive understanding framework by leveraging shared knowledge and enhancing cross-scene generalization capabilities. The proposed model is trained on a dataset of 245,000 images and validated on various datasets, demonstrating robust zero-shot generalization capabilities.
Structure:
- Introduction to Buildings in Remote Sensing
- Buildings as crucial components of man-made structures.
- Importance in geographic information databases.
- Challenges in Building Interpretation
- Current methodologies treating extraction and change detection separately.
- Complexity of remote sensing image scenes.
- Proposed Solution: RSBuilding Model
- Foundation model approach for enhanced generalization.
- Multi-level feature sampler for scale information augmentation.
- Training Strategy and Results
- Federated training strategy for smooth convergence.
- Validation on multiple datasets showcasing robust generalization capabilities.
Statistiche
RSBuilding achieves an IoU score of 92.15% on the WHU dataset.
RSBuilding outperforms other methods with an IoU of 86.19% on the LEVIR-CD dataset.