FTRNet: triplet fusion temporal relationship network for change detection in bitemporal remote sensing images

Wei Wu,Tong Li,Qi Xuan,QiMing Wan,Zuohui Chen
DOI: https://doi.org/10.1080/10106049.2024.2353253
IF: 3.45
2024-06-03
Geocarto International
Abstract:Change detection (CD) in remote sensing (RS) images aims to identify surface changes based on images acquired at different times. However, existing methods are still unsatisfactory in locating fine details of change in RS images, due to overlooking the inherent temporal information. To address the issue, we introduce a novel Triplet Fusion Temporal Relationship Network (FTRNet). FTRNet incorporates a triplet input backbone that enables the extraction of both spatial and temporal features. We design a change attention module to enhance bitemporal features, making the backbone network retain temporal information and fuse cross-scale features to extract the high-level location information. We evaluate our method on three benchmark datasets, including LEVIR-CD, WHU-CD, GZ, and DSIFN. The experimental results showcase that FTRNet achieves IoU scores of 83.60%, 77.06%, 73.67%, and 77.30% in LEVIR-CD, WHU-CD, GZ, and DSIFN datasets, respectively. These results surpass the second-best baseline by 1.20%, 0.49%, 1.31%, and 1.20%, respectively.
geosciences, multidisciplinary,environmental sciences,remote sensing,imaging science & photographic technology
What problem does this paper attempt to address?
The paper aims to address the problem of Change Detection (CD) in high-resolution remote sensing images. Specifically, the research focuses on how to accurately identify surface changes in bi-temporal remote sensing images, especially the challenges in handling fine change details. Current methods often overlook the inherent temporal information of images, making it difficult to accurately capture the details of changes. To solve this problem, the authors propose a new Triplet Fusion Temporal Relationship Network (FTRNet). FTRNet extracts spatial and temporal features by introducing a three-input backbone network and designs a Change Attention Module to enhance bi-temporal features, retain temporal information, and fuse cross-scale features to extract high-level positional information. The main contributions of this method include: 1. Proposing the FTRNet framework, which is based on three inputs (two temporal remote sensing images and their difference map), to better detect key changes and improve robustness to background noise. 2. Developing the Change Attention Module (CAM) and Change Segmentation Module (CSM) to effectively utilize temporal information to model change areas. 3. Conducting extensive experiments on three standard datasets, demonstrating that FTRNet has better performance compared to other state-of-the-art change detection algorithms. In summary, the goal of this paper is to improve the detection accuracy and robustness in the task of bi-temporal remote sensing image change detection, particularly in handling fine changes.