PID-CD: PID Control is What Change Detection Needs

Remote sensing change detection (RSCD) is a remote sensing technique used to identify geomorphological changes in bitemporal images of the same region. However, the identification of change regions is often hindered by false changes in similar objects, resulting in inaccurate boundary recognition of the change regions. This article adopts the proportional-integral-derivative (PID) control concept and designs two modules: PID feature alignment (PIDFA) and PID feature fusion (PIDFF), which makes up the PID-CD model. The PIDFA enhances and corrects the change region features using manifold structures and adaptive strategies, while the PIDFF introduces a third branch to guide multilevel feature fusion, progressively refining the boundaries of the change regions. Finally, extensive experiments on benchmark datasets, including LEVIR-CD, S2Looking, and CL-CD, display that PID-CD achieves state-of-the-art performance in RSCD. Specifically, PID-CD achieves the F1 score of 91.94% on LEVIR-CD, 66.61% on S2Looking, and 74.11% on CL-CD, showing significant improvements in detail and boundary of change area.