Landslide recognition with sample augmentation based on a joint DCGAN and Pix2Pix

Landslides are among the most frequent and destructive geological hazards in mountainous regions; however, the ability to automatically identify landslides is often constrained by limited amounts of labelled samples, modest gains derived from traditional augmentation methods, and interference from complex backgrounds. To improve the performance achieved under small-sample conditions, a virtual sample generation framework (D2P) that combines a DCGAN and Pix2Pix is proposed in this paper. The DCGAN is first used to generate diverse landslide label masks, and Pix2Pix then synthesises the corresponding remote sensing images, yielding structurally complete and semantically consistent label–image pairs. On this basis, a transfer learning strategy (D2P-TL) is implemented and evaluated on six mainstream semantic segmentation models. Across various experiments involving different enhancement ratios, original sample sizes, and augmentation strategies, D2P-TL consistently outperforms the no-enhancement (DT), traditional augmentation (CA-TL), and single-model augmentation approaches (DCGAN-TL and Pix2Pix-TL), improving the mIoU by approximately 5%–14% and maintaining high accuracy even when only 10%–30% of the original samples are used. In an independent test area in Wenshan Prefecture, Yunnan Province, D2P-TL also exhibits stable cross-regional performance, indicating its potential for extensions to other geomorphic hazards, such as debris flows and collapses.