MSAR-RadioMap: Multimodal SAR-Aided Radio Environment Mapping Dataset

This dataset is intended for researchers and developers working on radio environment map (REM) reconstruction, wireless signal interpolation, and multimodal data fusion in wireless communication environments.1. Dataset StructureEach data sample is provided as a triplet:RSRP.npy: A 256\u00d7256 NumPy array representing sparse Reference Signal Received Power (RSRP) values (in dBm), with missing values masked.SAR_VV.npy: A 256\u00d7256 co-registered SAR backscatter image in VV polarization (normalized intensity).SAR_VH.npy: A 256\u00d7256 co-registered SAR backscatter image in VH polarization (normalized intensity).Building.npy: A 256\u00d7256 binary or normalized float map indicating building region information.All files are grouped into individual sample folders.Total number of samples: 1600.2. PreprocessingRSRP data: Simulated using Altair Feko under urban scenarios with uniform 10-meter building height and 1750 MHz transmission frequency. Random sampling masks at 1%, 3%, and 5% can be applied manually or generated using the provided code.SAR imagery: Acquired from Sentinel-1 (via Google Earth Engine), covering VV and VH polarizations. All SAR images are temporally averaged to reduce speckle noise and seasonal variation, and spatially aligned with the RSRP and building maps.Building maps: Extracted from OpenStreetMap and rasterized to match the resolution and extent of SAR and RSRP data.All modalities are normalized to the [0, 1] range and co-registered in a common spatial grid.3. Recommended UseWe recommend using this dataset to:Train or evaluate deep learning models for sparse REM completion;Test multimodal fusion strategies combining RSRP + SAR (VV\/VH) + building priors;Benchmark signal recovery performance under extreme sampling sparsity;Conduct ablation studies on structural priors and polarization sensitivity in wireless modeling.You may replicate the experiments in our IEEE manuscript using this dataset as input to CMCTNet or alternative architectures.