Transformer-driven monocular high-resolution DTM generation on mars via multimodal integration of CaSSIS imagery and MOLA altimetry

High-resolution digital terrain models (DTMs) are essential for geomorphological analysis of planetary surfaces, yet Martian topography remains constrained by the coarse resolution of global datasets such as MOLA and by the limited coverage of stereo-derived products. In this work, we present SurfPlaNet, a deep learning framework based on a Dense Residual Connected Transformer Architecture (DRCT) that reconstructs intermediate-resolution DTMs from single-view (monocular) orbital observations. Unlike traditional photogrammetric pipelines, SurfPlaNet does not require stereo geometry, but instead fuses panchromatic CaSSIS imagery with HRSC–MOLA coarse elevation data to infer detailed topography. The architecture integrates Swin Transformer blocks with an affine calibration head and is trained end-to-end using terrain-aware loss functions combining masked pixel losses, gradient consistency, and total variation. Experimental evaluation against CaSSIS stereo-derived DTMs shows that SurfPlaNet achieves average elevation errors on the order of 61 meters. While this accuracy remains lower than stereo-based methods, the model is capable of recovering geomorphological features such as crater rims, ridges, and localized slope variations that are absent in HRSC–MOLA inputs. Crucially, by leveraging MOLA as a global elevation prior, SurfPlaNet produces metrically calibrated predictions that can be generalized across the Martian surface, including areas not covered by CaSSIS stereo. This demonstrates the potential of monocular transformer-based approaches to complement stereo pipelines, enabling broader coverage of Mars with consistent, scalable topographic reconstructions. • We present SurfPlaNet, a transformer-based model that reconstructs high-resolution Martian DTMs from single-view CaSSIS images fused with low-resolution HRSC-MOLA data. • The architecture combines Swin Transformer blocks with an affine calibration module and is trained using terrain-aware losses, including L masked-L1 , gradient consistency, and total variation loss. • The model reliably reconstructs fine-scale geomorphological structures across heterogeneous landscapes, surpassing the performance of conventional baseline model and facilitating scalable monocular topographic mapping.