Abstract
Building detection and change detection are crucial for urban planning and rescue operations, especially for assessing building damage after natural disasters. Most existing models rely on a single pre-disaster image, which hampers performance due to the presence of destroyed buildings in post-disaster images.
To address this, we propose a Siamese model called SiamixFormer, which utilizes both pre- and post-disaster images as input. Our model features two encoders with a hierarchical transformer architecture, where the output at each stage is fed into a temporal transformer for feature fusion. Specifically, queries are generated from pre-disaster images, while (keys, values) are derived from post-disaster images. Temporal features are also considered in the fusion process.
An advantage of using temporal transformers for feature fusion is their ability to maintain larger receptive fields generated by transformer encoders compared to CNNs. The output from the temporal transformer is then passed to a simple MLP decoder at each stage.
The SiamixFormer model was evaluated on the xBD and WHU datasets for building detection, and on the LEVIR-CD and CDD datasets for change detection, demonstrating superior performance over state-of-the-art methods.
Blogger's Review: The SiamixFormer model significantly enhances the accuracy of building and change detection, particularly in complex post-disaster scenarios, showcasing the powerful potential of transformer architectures in remote sensing image processing. This research provides a more reliable tool for future urban monitoring and post-disaster assessments.