Abstract:

Due to factors such as differences in acquisition time, angle, and sensor characteristics, dual temporal remote sensing images often manifest various pseudo-changes. Moreover, certain changes may have an uninteresting nature and typically correlate with adjacent objects. However, the utilization of a fully convolutional neural network (FCN) may lead to the loss of long-range information. To address this issue, this study proposes a network that integrates convolutional neural networks (CNN) and Transformer (C-T Net), which has an overall network architecture consisting of a deep feature extraction section and a detection head section. The network backbone combines CNN and Swin Transformer. Additionally, two novel fusion modules, C-to-T and T-to-C, are designed to amalgamate local features and global features. The detection head section utilizes Transformer encoding and decoding to derive refined feature maps for discerning change regions. Comparative experiments with multiple change detection models validate the efficacy of C-T Net. On the LEVIR-CD and WHU-CD datasets, the proposed method achieves the highest F₁_1 (90.63%, 86.24%) and $ {p}_{\mathrm{I}\mathrm{o}\mathrm{U}} \_1 $(82.87%, 75.81%). Results across both datasets affirm that our proposed algorithm outperforms existing methodologies from both visual and data-centric perspectives.

Key words: multi-temporal, change detection, convolutional neural networks (CNN), transformer, feature fusion