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/<site>/*.npy
test/<site>/*.npy
manifest_hf.csv
errors.log
Each .npy corresponds to one source tile and contains multiple patches.
import numpy as np
array = np.load("path/to/tile.npy", mmap_mode="r")
patch = np.asarray(array[0], dtype=np.float32)
Held-out test sites
2018_MDAS_052024_STER_72025_SERC_7
Normalization
The files preserve the prepared numerical values. Verify the value range and
apply the same physical scaling in training and evaluation. Do not assume the
original HyperSIGMA /4000 normalization is appropriate.
License and redistribution
Complete this section before making the repository public. Confirm that every source permits redistribution of derived patches and document attribution.
- Downloads last month
- 187