CHANGE MASK-GUIDED MULTI-TEMPORAL IMAGE PAIR GENERATION BASED ON DIFFUSION MODEL

Building change detection (BCD) in remote sensing images primarily involves using multi-temporal images of the same region to extract information about building changes. The development of data-driven deep learning has significantly advanced BCD. However, due to the slow frequency of building changes, it is challenging to obtain effective BCD data pairs, making it difficult to construct a class-balanced BCD dataset. This limitation severely restricts the performance improvement of BCD models. To address this issue, In this paper, a change mask-guided multi-temporal image pair generation (CMMG) is proposed, which utilizes the de-noising process to enlarge the latent representation space for multi-temporal image pair generation. To achieve this, first, a mask-guided building remote sensing image generation approach (MBG) based diffusion model is designed. This approach enables the controlled generation of building information while randomly generating non-building information. Second, a change-controllable multi-temporal background-diverse change detection image pair generation strategy (CDMG) is proposed. This strategy leverages a trained MBG and multi-temporal building masks whose changes can be controlled to generate diverse change detection image pairs. Extensive experiments demonstrate that our method exhibits excellent generative capabilities, and the generated diverse and realistic change detection data significantly unleash the potential of plain change detection model, especially under conditions of limited training data.