The Multiscale Differential Feature Optimization Networks for Remote Sensing Image Change Detection

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 17; pp. 16847 - 16859
• Main Authors: Wang, Jinbo, Zhang, Lingling
• Format: Journal Article
• Language: English
• Published: IEEE 2024
• Subjects: Accuracy; Change detection (CD); deep supervision (DS); feature enhancement upsample; Feature extraction; multiscale differential feature; Optimization; Remote sensing; remote sensing image; Spatiotemporal phenomena; Task analysis; Three-dimensional displays
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2024.3452948

In the field of remote sensing, change detection in remote sensing images is vital. It is an important tool in urban planning, land use, and resource management. By leveraging the powerful self-learning capabilities of deep learning, it cannot only significantly improve the accuracy and efficiency of change detection but also handle complex land features. However, current methods have not yet achieved the ideal utilization of spatial information relationships and high and low semantic features between dual-temporal images, leading to fuzzy detection boundaries and limited accuracy. To address these issues, this article proposes a multiscale differential feature optimization network. This network first concatenates remote sensing bitemporal images that contain temporal relationships along the depth dimension, and then uses 3-D convolutional technology to extract multilayer feature information. To further improve the detection effect, we designed a multiscale differential feature optimization module to extract spatial semantic feature information from different levels of feature maps. In order to fully extract the fused feature information and enhance the focus on edges and small details, we have introduced a feature enhancement upsample module and a deep supervision strategy. Furthermore, we use a mixed loss function to calculate the loss value, achieving high-precision identification of change areas. The experimental results on the LEVIR-CD dataset and WHU-CD dataset show that compared with various state-of-the-art methods, the proposed method in this article has significant advantages in terms of the F 1-score and intersection over union index. Specifically, the F 1-scores on the LEVIR-CD dataset and WHU-CD dataset reached 90.91% and 92.23%, respectively. The extracted change area boundaries are clearer, and the false negative and false positive rates are lower, demonstrating the strong robustness of the network model. The experimental results demonstrate that the proposed multiscale differential feature optimization network has made significant progress in the task of remote sensing image change detection, providing new perspectives and methodologies for future research in this field.