HiMambaMap: A Hilbert-Scanning Mamba Diffusion Network for Online Map Generation from Remote Sensing Imagery

Remote sensing image-to-online map translation (RSMT) represents a critical research frontier in geographic information processing. With the rapid development of deep learning, generative models have attracted increasing attention for automatic map generation. However, existing methods incur substantial computational overhead when modeling global contextual information. Recently, Mamba has been widely adopted in remote sensing for capturing global dependencies with linear complexity. However, naively applying its sequential scanning mechanism to RSMT tends to disrupt local spatial relationships essential to map generation. In this study, we propose HiMambaMap, a Hilbert-scanning Mamba diffusion network for RSMT. Specifically, it is built upon a Brownian bridge diffusion model, where the proposed denoiser combines a Convolutional Neural Network (CNN) encoder for local spatial feature extraction and a Mamba decoder for efficient global dependency modeling. To further improve the spatial consistency of the generated maps, we introduce a Multi-Scale Feature Fusion (MSFF) module into the denoiser for enhanced multi-scale feature fusion and incorporate a four-directional Hilbert scanning strategy into the Mamba decoder to better preserve spatial locality. Extensive experiments on two widely used datasets demonstrate that HiMambaMap achieves competitive generation quality and pixel-level color consistency, while maintaining the lowest parameter count and computational complexity among several state-of-the-art methods.