LoessFormer: a self-supervised framework for loess landform recognition with unified semantic and boundary modelling

Accurate recognition of loess landforms is essential for geomorphological mapping, terrain evolution analysis, and surface process studies. However, existing methods remain limited in jointly representing geomorphological semantics and spatial structural characteristics within a unified framework. In addition, loess landform boundaries in DEM data usually exhibit gradual transitions, which further increases the difficulty of landform recognition and spatial extent representation. To address these issues, this study proposes LoessFormer, a self-supervised framework for loess landform recognition with unified semantic and boundary modelling. Specifically, two complementary self-supervised tasks, namely Masked Patch Prediction (MPP) and Patch Order Prediction (POP), are designed to learn local continuity and global spatial relationships from DEM data. These tasks are integrated into a Transformer-based architecture to enhance the representation capacity for complex geomorphological structures. During downstream fine-tuning, an object-level unified modelling strategy is adopted to jointly learn landform semantics and spatial extent representations. Experiments conducted in typical regions of the Chinese Loess Plateau demonstrate that LoessFormer achieves superior performance over existing mainstream methods across multiple evaluation metrics, including [email protected], F1-score, and IoU. In particular, the proposed method achieves an [email protected] of 55.68%, demonstrating its effectiveness in recognising complex loess landforms and representing their spatial structures.