Skip to content

DATASETS.md

Latest commit

 

History

History
58 lines (45 loc) · 2.82 KB

DATASETS.md

File metadata and controls

58 lines (45 loc) · 2.82 KB
Raw

Dataset registry and acquisition record

Graph-RMS was evaluated on nine complete public hyperspectral scenes. The dimensions below are verified against the frozen registry and manuscript. Reference class count and labelled pixels are evaluation metadata; the primary Graph-RMS processing path does not receive the class count.

Dataset Shape Bands Classes Labelled pixels Labelled fraction Role
Salinas-A 83 x 86 204 6 5,348 74.92% development
Indian Pines 145 x 145 200 16 10,249 48.75% development
KSC 512 x 614 176 13 5,211 1.66% development
Pavia University 610 x 340 103 9 42,776 20.62% development
WHU-Hi LongKou 550 x 400 270 9 204,542 92.97% development
WHU-Hi HongHu 940 x 475 270 22 386,693 86.61% development
WHU-Hi HanChuan 1217 x 303 274 16 257,530 69.84% development
Botswana 1476 x 256 145 14 3,248 0.86% development
Trento 166 x 600 63 6 30,214 30.34% independent primary holdout

Loader sources used in the frozen code

The loader records the exact acquisition endpoints used during the study:

  • Salinas-A: danaroth/salinas dataset mirror on Hugging Face.
  • Indian Pines: danaroth/indian_pines dataset mirror on Hugging Face.
  • Pavia University: danaroth/pavia dataset mirror on Hugging Face.
  • KSC and Botswana: Tanishq165/HSI_Datasets dataset mirror on Hugging Face.
  • WHU-Hi scenes: danaroth/whu_hi dataset mirror on Hugging Face.
  • Trento: files Italy_hsi.mat and allgrd.mat in the public tyust-dayu/Trento GitHub repository.

These are reproducibility endpoints, not assertions that a mirror controls the original dataset rights. The release does not redistribute the files.

Original-provider and licensing checklist

Before public deposition, an author should verify and record for each dataset:

  1. the original institutional or project landing page;
  2. the preferred citation requested by the provider;
  3. the dataset licence or explicit terms of use;
  4. whether redistribution of the raw cube and ground truth is prohibited;
  5. the download date and, where practical, a checksum.

The supplied evidence did not contain authoritative licence texts for all nine datasets. Therefore no dataset licence is guessed here. This is a manual FAIR metadata task, not a computational blocker for inspecting the included derived outputs.

Preprocessing used by Graph-RMS

  • All pixels in the complete cube are retained.
  • Each spectral band is standardized over the complete scene.
  • PCA-20 is used for local-neighbour ranking, not for full-spectrum affinity weighting or the initial diffusion state.
  • No labelled-pixel subset is used to estimate preprocessing parameters.
  • Evaluation is restricted to pixels whose reference label is nonzero.