Cross-temporal and spatial information fusion for multi-task building change detection using multi-temporal optical imagery

Abstract

Accurate detection of changes in buildings is crucial for the understanding of urban development. The growing accessibility of remote sensing imagery has enabled urban scale change detection (CD) in both 2D and 3D. However, existing methods have not yet fully exploited the fusion of feature information in multi-temporal images, resulting in insufficient accuracy in 2D changed regions or in elevation changes. To this end, a Cross-temporal and Spatial Context Learning Network (CSCLNet) aimed at multi-task building CD from dual-temporal optical images is proposed, capturing both 2D and 3D changes simultaneously. It leverages a CNN network to extract multi-layer semantic features. Subsequently, two modules, Cross-temporal Transformer Semantic Enhancement (CTSE) and Multi-layer Feature Fusion (MFF), are developed to refine the feature representations. CTSE enhances temporal information by cross attention of dual-temporal features to enable interactions and MFF fuses multi-layer features and enhances attention to global and local spatial context. Finally, two prediction heads are introduced to separately handle 2D and 3D change prediction, identifying changed building objects and their elevation changes. Experiments conducted with two public datasets, 3DCD and SMARS, show that the CSCLNet achieves state-of-the-art for both 2D and 3D CD tasks. In particular, the change-specific RMSE of elevation changes has been reduced to 4.52 m in real world scenes. The code is available at: https://github.com/Geo3DSmart/CSCLNet.

Keywords

• Change detection
• Multi-task
• Bi-temporal optical images
• Deep learning