Latent Space Feature-driven Semi-supervised Change Detection Network for Illumination Change Scenarios

With the increasing abundance of remote sensing images resources, the challenges faced by change detection are becoming more and more severe. Among them, the illumination change problem significantly increases the detection difficulty. This paper proposes a latent space feature-driven semi-supervised change detection network (LSF-CDNet) for illumination change scenarios, aiming to solve the problems of limited feature extraction ability and high training difficulty of existing methods when dealing with illumination change scenarios. This method constructs a generalized feature extraction paradigm with the stable diffusion model, uses unlabeled data to fine-tune the latent space encoder to improve the feature extraction ability, and adopts a parameter freezing strategy to reduce training difficulty and deployment time. At the same time, a Dual Pathway Attentional Convergence (DPAC) module is designed to fuse the channel attention and spatial attention mechanisms to enhance the model’s perception and localization ability of changed regions. The experimental results on the WHU and CLCD datasets show that the F1-score of LSF-CDNet in change detection reaches 92.62% and 69.35% respectively, and various indicators reach the current state-of-the-art (SOTA) level. Moreover, the deployment efficiency is considerable, and the single-epoch training time in the WHU dataset experiment is only 4.88 minutes.