CA-2022-S2-NAIP: A High-Quality Cross-Sensor Dataset for Remote Sensing Image Super-Resolution

Paper: CDEHAT: Conditional Diffusion-Assisted Enhanced Hybrid Attention Transformer for Remote Sensing Imagery Super-Resolution

The dataset is derived from publicly available satellite and aerial imagery: Sentinel-2 Level-2A images from the Copernicus Data Space Ecosystem, and high-resolution NAIP imagery from the United States Geological Survey (USGS). All images are projected using the UTM coordinate system (in meters), which preserves geometric consistency and avoids distortions associated with latitude-longitude representations—an essential consideration for SR tasks involving spatial resolution.

During preprocessing, Sentinel-2 images undergo atmospheric correction, while NAIP images are downsampled to 2.5-meter resolution. Spatiotemporally matched image groups are formed based on overlapping spatial extents and acquisition time differences (≈ 1 day). Cloud coverage is assessed using Sentinel-2 Scene Classification Layers (SCL); samples with >25% cloud cover are discarded, and cloud-affected pixels are set to null. NAIP images are assumed cloud-free.

Further refinement includes subpixel-level co-registration of NAIP to Sentinel-2 imagery using ArcGIS, followed by non-overlapping sliding window cropping to generate spatially aligned low-resolution (LR) and high-resolution (HR) submap pairs. Geometric quality is ensured by removing samples with control point errors exceeding a defined threshold. Spectral consistency between sensors is enhanced by applying histogram matching to NAIP submaps, aligning their spectral distribution more closely with Sentinel-2. Submap pairs with significant spectral deviationor residual cloud artifacts are filtered out via spectral angle thresholding.

All valid submap pairs are stored in GeoTIFF format with LZW compression. A comprehensive JSON metadata file accompanies the dataset, recording for each submap: source file name, sensor acquisition time, spatial extent, coordinate system, spatial resolution, data type, compression method, and spatial/spectral quality metrics.

This dataset is constructed with a strict emphasis on geographic independence to ensure fair and unbiased evaluation in super-resolution (SR) tasks. To prevent geographic leakage, the training, validation, and test sets are divided based on the spatial extents of subgraph pairs. Subgraphs that share the same spatial footprint or exhibit spatial overlap are assigned to the same set, ensuring that no same geographic region appears across multiple sets. This design eliminates the possibility of models (e.g. Transformer) learning region-specific knowledge from overlapping spatial contexts.

The final dataset, named CA-2022-S2-NAIP, is well-suited for training and evaluating SR deep learning models under cross-sensor conditions. You can now download the Super Resolution.zip file for super resolution tasks.