HyperSIGMA 1 m Hyperspectral NPY Patches

HyperSIGMA 1 m hyperspectral pretraining patches Summary Kept patches: 2,519,999 Preprocessing errors recorded: 198 Patch shape: C, 64, 64 Tile file shape: N, C, 64, 64 Expected channels: 216 Prepared dtype: float32 (verify individual files) Invalid-patch rule: reject a patch when at least 50% of spatial pixels have all spectral values equal to zero. Tiling: stride 64 where possible; edge windows can overlap slightly. Layout train/ / .npy… See the full description on the dataset page: npy.

Source: Hugging Face Hub (zhu-xlab/ofahsi_npy). Metadata imported from the dataset’s Hub tags.