MLGFENet: Multi-Scale Local-Global Feature Enhancement Network for High-Resolution Remote Sensing Image Change Detection

Abstract

The integration of a convolutional neural network (CNN) and a Transformer has become a dominant framework for change detection (CD) in remote sensing images, because of its ability to effectively model both local and global features. However, existing methods still face challenges in extracting features from change regions with diverse shapes and sizes, as well as in capturing edge information. To address these issues, this study proposes a CD model for high-resolution remote sensing images that enhances multiscale local and global features by combining CNN and Transformer. Initially, a multiscale local feature extraction module is constructed. By leveraging the enhanced ResNet50 architecture and incorporating the atrous spatial pyramid pooling technique, this module is capable of accurately capturing the multiscale local details of bitemporal images. Subsequently, a hybrid-scale global context feature extraction module is designed. This module enables the modeling of multiscale global contextual information, thereby further enhancing the model’s feature representation capability. Next, a cascading feature decoder is employed to perform upsampling on the extracted features. Through the use of skip connections, local and global features are efficiently merged at multiple scales. Finally, a differential enhancement unit is utilized to generate differential features that are rich in change information. Additionally, a composite loss function is introduced, which takes into account both pixel-based segmentation errors and edge-based segmentation errors, enabling accurate localization of changed regions. Experimental results on three publicly available high-resolution remote sensing image datasets, namely, LEVIR-CD, WHU-CD, and CDD, demonstrate that the proposed method outperforms several state-of-the-art comparative methods in terms of CD efficacy. It effectively addresses common problems in CD, such as undersegmentation, oversegmentation, and rough edges.