A Hierarchical Local-Global-Aware Transformer With Scratch Learning Capabilities for Change Detection

Most transformer-based methods rely on pretraining weights on large datasets such as Imagenet or pretraining from specific change detection (CD) datasets and then fine-tuning on the target dataset. When the target dataset significantly diverges from the dataset used for pretraining, the model’s ability to generalize to remote sensing imagery may be compromised due to the domain gap. In this letter, we propose HierFormer, which has the advantage of processing semantic features hierarchically, using simple operations for shallow features, spatial position transformation for middle-level features, and channel information interaction for high-level features. In addition, we propose a local-global-aware (LGA) attention block, which reduces the computational overhead of self-attention by sparse attention and increases the locality inductive bias (LIB) of the transformer by focusing attention on the local region and sparse part of the global region, which enables the model to be trained from scratch on small to medium-sized CD datasets. Finally, a new feature fusion decoder (FFD) is proposed to fuse the bitemporal features, which reweights the features of each channel through attention mechanism. Compared with other transformer-based or transformer-CNN-based hybrid networks, our method significantly improves F1, reaching 91.56% and 97.56% on the LEVIR-CD and CDD-CD change detection datasets. Our code is available at
https://github.com/WesternTrail/HierFormer
.