EarthMapper: Visual Autoregressive Models for Controllable Bidirectional Satellite-Map Translation

Cross-modal translation between satellite imagery and cartographic maps constitutes a foundational challenge in geospatial artificial intelligence, with critical applications ranging from autonomous mapping systems to disaster response infrastructures. While existing methods attempt to address unidirectional satellite-to-map translation, the bidirectional satellite-map translation (BSMT) problem remains formally undefined and technically under-explored, particularly due to: 1) nondifferentiable geometric discrepancies between pixel-level observations and symbolic representations, causing cross-modal feature misalignment; and 2) conflicting learning objectives requiring simultaneous high-level geographic abstraction and pixel-wise photorealism. To address these limitations, we introduce EarthMapper, a novel autoregressive (AR) framework for controllable BSMT. EarthMapper employs geographic coordinate embeddings to anchor generation, ensuring region-specific adaptability, and leverages multiscale feature alignment within a geo-conditioned joint scale autoregression (GJSA) process to unify bidirectional translation in a single training cycle. A semantic infusion (SI) mechanism is introduced to enhance feature-level consistency, while a key point adaptive guidance (KPAG) mechanism is proposed to dynamically balance diversity and precision during inference. We further contribute CNSatMap, a large-scale dataset comprising 302132 precisely aligned satellite-map pairs across 38 Chinese cities, enabling robust benchmarking. Extensive experiments on CNSatMap and the New York dataset demonstrate EarthMapper's superior performance, achieving significant improvements in visual realism, semantic consistency, and structural fidelity over state-of-the-art methods. In addition, EarthMapper excels in zero-shot tasks like in-painting, out-painting, and coordinate-conditional generation, underscoring its versatility. The source code for EarthMapper and the CNSatMap dataset will be publicly available at https://github.com/HIT-SIRS/EarthMapper