Remote sensing change detection via centralized low-high rank representation

Recently, the advancement of remote sensing equipment and image processing technology has resulted in high-resolution remote sensing (HRRS) images containing more detailed geospatial characteristic. This progress has increased the demands on both the recognition capabilities and the detection speed required for remote sensing change detection (RSCD). Due of a wide variety of surface objects on the earth, RSCD tasks are characterized by “intra-class variance” and “inter-class similarity”. Current mainstream lightweight deep learning models often use fewer network layers to maintain a low model parameters. Therefore, lightweight networks often struggle to capture the rich contextual information present in HRRS images, making comprehensive change detection challenging, specifically as shallow networks struggle to maintain contextual information of single-temporal images while fully detecting changes in bitemporal images. In the paper, we introduce a straightforward and lightweight network called Centralized High-Low Rank and Fusion Network (CHLF-Net), which primarily consists of two modules. First, we design a Centralized High-Low Rank Representation (CHLR) module to extract intra-temporal and inter-temporal information in two ways. One is High-Rank Representation Aggregation (HRRA), which uses an MLP structure to compute difference information between bitemporal images; the other is Low-Rank Representation Comparison (LRRC), which uses learnable low-rank matrices to capture contextual information of single-temporal images and compares the obtained bitemporal low-rank matrices to acquire highly discriminative change features. Second, we design a Lightweight Deep-Shallow Attention (LDSA) module, which lightweightly fuses deep and shallow features, enabling a better integration of rich contextual information and detailed multi-scale features. Employing a shallow backbone network without intricate structures, CHLF-Net, with just 1.27M model parameters, surpasses other advanced methods on three RSCD datasets, showcasing its superiority. Our code is available at https://github.com/yikuizhai/CHLF-Net.