Multi-Temporal SamplePair Generation for Building Change Detection Promotion in Optical Remote Sensing Domain Based on Generative Adversarial Network

Change detection is a critical task in remote sensing Earth observation for identifying changes the Earth’s surface multi-temporal image pairs. However, due to time-consuming nature of collection, labor-intensive pixel-level labeling with rare occurrence building changes, and limitation location, it difficult build large, class-balanced, diverse change dataset, which can result insufficient changed sample pairs training models, thus degrading their performance. Thus, this article, given that data scarcity class-imbalance issue lead novel pair generation method, namely, Image-level Sample Pair Generation (ISPG), proposed improve performance through dataset expansion, generate more valid overcome small amount information existing datasets. To achieve this, Label Translation GAN (LT-GAN) was designed complete images background pseudo-changes without any complex blending steps used previous works. obtain detailed features detection, especially surrounding context buildings, we multi-scale adversarial loss (MAL) feature matching (FML) supervise quality generated bitemporal On other hand, also consider distribution buildings should follow pattern human-built structures. The approach evaluated on two datasets (LEVIR-CD WHU-CD), results proved method state-of-the-art (SOTA) performance, even if using plain models detection. In addition, plug-and-play solution be model.