Labeled dataset of Sentinel-1 SAR imagery Despeckled with multitemporal fusions

This paper presents a labeled multitemporal dataset designed to support supervised despeckling approaches for Sentinel-1 synthetic aperture radar (SAR) imagery. The dataset consists of 512×512-pixel SAR regions acquired in Interferometric Wide Swath (IW) mode (GRD-HD product) for both VV and VH polarizations, paired with corresponding speckle-reduced reference images generated by multitemporal averaging. For each location, ground-truth images were constructed from temporal stacks of 5, 10, 15, 20, and 25 acquisitions, enabling controlled analysis of speckle attenuation as a function of the number of fused scenes. The data were collected globally using an automated workflow implemented in Python and Google Earth Engine, with random spatial sampling and quality filtering based on polarization-specific backscatter thresholds. In addition to the SAR imagery, the dataset includes ESA WorldCover v200 land cover maps and a metadata file containing acquisition parameters and geographic information. Baseline quantitative metrics, including Equivalent Number of Looks (ENL), Mean Squared Error (MSE), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR), are provided for representative samples to facilitate benchmarking. The proposed dataset provides a structured, reproducible resource for training and evaluating machine learning and deep learning models for speckle noise reduction in SAR data.