A self-supervised dual-branch network for terrain matching localization

Abstract

Terrain matching localization provides an effective passive navigation solution for unmanned aerial vehicles. However, traditional methods often rely on manually designed descriptors, raw elevation comparisons, or iterative optimization, making them sensitive to measurement noise, terrain ambiguity, and rotation disturbances. To overcome these challenges, this paper proposes DCLNet, a self-supervised dual-channel network for terrain matching localization. DCLNet is constructed as a terrain-matching autoencoder, in which the encoder learns discriminative terrain descriptors from DEM patches without requiring position labels. The local channel extracts fine-scale terrain variations using residual convolution and spatial attention, whereas the global channel models extended spatial relationships via multi-head self-attention. A spatial attention fusion module is further introduced to adaptively integrate local and global terrain features. Experiments on the ASTER GDEM V3 dataset show that DCLNet achieves better localization performance than traditional terrain matching and retrieval-based methods while maintaining competitive computational efficiency.