nickstheisen/hsdatasets

The hsdatasets package provides pytorch-DataSet wrappers for the most common hyperspectral data sets with pixel-precise ground-truth annotations.

hsdatasets

The hsdatasets-package provides pytorch-DataSet wrappers for the most common hyperspectral data sets with pixel-precise ground-truth annotations. This simplifies the usage of those data sets in deep learning applications.

Currently only wrapper classes for remote sensing data sets are provided but in the future other data sets such as HyKo2 will be provided as well.

Remote Sensing Data

Data Sampling

After loading the data the image is zero-padded and NxN patches are sampled at each pixel position. If necessary the dimensionality can be reduced to M dimensions using PCA. The user can define M and N during class instantiation.

Warning: If N>1 pixels in multiple data patches overlap which may lead to data leakage when not taken care of.

Currently Supported Data Sets

Data Set	Spatial Resolution [px]	Spectral Resolution [bands]	Classes	Sensor
Indian Pines	145 x 145	200	16	AVIRIS
Salinas Scene	512 x 217	204	16	AVIRIS
Salinas-A	86 x 83	204	6	AVIRIS
Kennedy Space Center	512 x 614	176	13	AVIRIS
Pavia Centre	1096 x 1096	102	9	ROSIS
Pavia University	619 x 610	103	9	ROSIS
Botswana	1476 x 256	145	14	Hyperion
AeroRIT	1973 x 3975	51	5	Headwall Hyper E