GateFormer: Gate Attention UNet With Transformer for Change Detection of Remote Sensing Images
Lili Li, Zhi-Hui You, Si-Bao Chen, Lili Huang, Jin Tang, Bin Luo
Abstract
Extraction of global context information plays a major role in change detection (CD) of remote sensing (RS) images. However, the majority of methods now depend on convolutional neural networks (CNNs), which are difficult to obtain complete context information due to the limitation of local convolution operation. This study proposes a novel gate attention UNet with Transformer for CD of RS images. GateFormer consists of an encoder with Transformer-based Siamese network. Firstly, we propose a gate attention mechanism (GAM), which filters the low-level information by guiding high-level features and focuses on activation of relevant knowledge instead of allowing all to pass. In addition, space pooling module (SPM) in generator extracts more spatial features from pixel level to suppress the generation of noises. Finally, in order to increase the CD accuracy of small-scale ground objects, we design a feature downsampling module (FDM) to minimize the loss of detailed information and compress more small-scale features in feature downsampling of Transformer. The efficiency of our suggested approach has been verified by experiments on three RS CD datasets.