Zhengzhou Yellow River National Wetland Park UAV Hyperspectral Dataset

Before each flight mission, uniform whiteboard correction and dark current correction were performed to eliminate illumination deviation and inherent sensor dark current noise, yielding a total of 314 raw hyperspectral frames.Refined radiometric preprocessing was implemented on each single raw hyperspectral frame via the official supporting software Cuvis Touch. First, the automatic integration time was locked to eliminate spectral bias induced by fluctuating ambient illumination.

Radiometric calibration was conducted using pre-acquired whiteboard reference frames to convert raw sensor digital number (DN) values into surface radiance. Dark current baseline frames captured with the lens cap closed were used to subtract intrinsic circuit noise band by band. Standard frames with clear ground object textures were selected for distance correction to compensate spectral and spatial errors caused by lens optical attenuation and field-of-view distortion.

After radiometric correction, Lanczos upsampling and PanRatio pan-sharpening fusion algorithms were applied to export separate single-frame hyperspectral reflectance imagery and panchromatic imagery. Band filtering was then performed to retain the experimental spectral range of 402–950 nm, producing standardized single-frame pure hyperspectral reflectance imagery with a pixel dimension of 330 × 330 and a spatial resolution of approximately 0.6 m.Photogrammetric stitching was further conducted for full-scene imagery.

All single-frame panchromatic images were imported into Agisoft Metashape to complete aerial triangulation, dense point cloud reconstruction and DEM generation. The panchromatic frames were batch replaced with corresponding single-frame hyperspectral reflectance imagery for rigorous orthorectification and seamless mosaicking. A single full-scene GeoTIFF orthomosaic was generated and subsequently converted to standard ENVI format for professional spectral analysis.

The final full-scene imagery has a pixel dimension of 9954 × 7541, a spatial resolution of approximately 0.12 m, a file size of around 20 GB, and adopts the geographic coordinate system "GCS WGS 1984".Pixel-wise ground truth labeling was completed based on field ground surveys supplemented by low-altitude visible-light UAV reconnaissance. A total of 14 typical land cover classes including platanus, reed, Yellow River water, tidal flat and building zone were delineated and annotated on the full-scene imagery, while 8 typical land cover types such as willow and reed were labeled for single-frame pure hyperspectral subsets.The complete dataset consists of full-scene mosaicked hyperspectral ENVI .dat files with corresponding pixel-wise label GeoTIFF files, as well as single-frame pure hyperspectral subsets and matching pixel-wise labels stored in .mat format, with a total storage volume of approximately 23 GB.

For single-frame samples, stratified pixel sampling was adopted: 5% of annotated pixels are assigned to the training set and the remaining 95% to the test set. The number of training samples from each category is strictly balanced to simulate real-world small-sample field classification scenarios. This subset can support research including fine-grained wetland land cover classification, vegetation spectral inversion, multi-temporal wetland dynamic monitoring, and validation of small-sample hyperspectral deep learning algorithms.

The full-scene hyperspectral dataset preserves the original mosaicked imagery and full-area pixel-wise ground truth labels completely, retaining continuous spatial distribution and native spectral signatures across the entire wetland, covering typical wetland land covers such as water bodies, tidal flats, aquatic vegetation, woodlands and artificial constructions. It is applicable to macro-scale remote sensing research including full-coverage wetland land cover mapping over large regions, quantitative analysis of regional wetland landscape patterns, inversion of full-area vegetation biochemical parameters, research on spatial differentiation of wetland ecosystems, and generalization performance testing of large-area hyperspectral classification models.No open-source external datasets were utilized in constructing this dataset.

All aerial images, field survey records, pixel-wise ground truth labels, and derivative preprocessed data are independently collected and produced by the research group. This dataset fills the gap of high-precision annotated UAV hyperspectral datasets for the Zhengzhou Yellow River wetland located in the middle and lower reaches of the Yellow River.