A transformer-based landform mapping method combining topographic knowledge and digital elevation data

• We propose a topographic knowledge-integrated transformer for landform mapping. • Terrain feature-guided attention enhance feature encoding of landform structures. • A topographic prior loss is proposed to constrain the learning process. • Efficient decoder design supports multi-scale fusion with lower computation. • Outperforms seven baseline models in both qualitative and quantitative assessments. Deep learning (DL) methods have emerged as potential tools for automated landform mapping. However, current approaches frequently overlook critical topographic features and prior topographic knowledge, which are essential for accurate landform differentiation. Consequently, these approaches struggle to resolve indistinct class boundaries and misclassification in transitional landform zones. The application of transformer-based approaches to landform mapping remains rare, although these methods are well-suited for capturing global geomorphic patterns. In this study, we develop TKGFormer, a topographic knowledge-guided and transformer-based DL model specifically designed for landform classification. This model explicitly integrates terrain residuals and multi-scale terrain feature-guided attention mechanisms to emphasise key topographic features and improve classification accuracy. We introduce a topographic feature prior loss function, which constrains the learning process by enforcing consistency between predicted landform classes and fundamental terrain characteristics. Moreover, we design an efficient decoder to enhance multiscale feature fusion. Experimental results demonstrate that TKGFormer outperforms five CNN-based models and two transformer-based model in landform mapping by using bare Earth 30 m resolution Copernicus digital elevation data. In terms of mIoU, TKGFormer exhibits an improvement that ranges from 4.4 % to 20.7 % over the other models. The integration of topographic knowledge into DL provides a promising direction for improving the interpretability and reliability of automated landform classification.