ghost-hsi 0.1.2

GHOST: Generalizable Hyperspectral Observation & Segmentation Toolkit

GHOST

Generalizable Hyperspectral Observation & Segmentation Toolkit

97.5% OA | 0.86 mIoU on Indian Pines — trained on a laptop GPU in 77 minutes.

pip install ghost-hsi

---

What is GHOST?

GHOST is a hyperspectral image segmentation framework. Point it at any .mat hyperspectral dataset and get a segmentation map — no code, no pipeline configuration, no PCA.

It runs on consumer hardware (RTX 3050, 6 GB VRAM) and handles any band count, class count, or spatial resolution automatically.

ghost train_rssp \
  --data data.mat --gt labels.mat \
  --routing forest --loss dice \
  --out-dir runs/my_experiment

ghost predict \
  --data data.mat --gt labels.mat \
  --model runs/my_experiment/rssp_models.pkl \
  --routing forest --out-dir runs/my_experiment

---

Why GHOST?

No PCA Required

Every major hyperspectral deep learning method requires PCA as a preprocessing step — choosing how many components to retain, accepting discarded spectral information. For non-standard domains (planetary science, medical imaging, novel sensors), this is a barrier and a source of information loss.

GHOST uses Continuum Removal: a physics-informed normalisation that strips brightness variation and isolates absorption feature shape. All spectral bands preserved. No dimensionality choices.

Fully Data-Agnostic

Band count, class count, spatial dimensions — all read from the file at runtime. Nothing is hardcoded. The same binary that segments Indian Pines (200 bands, 16 classes) also handles lung cancer pathology slides (61 bands, 3 classes) and Mars CRISM data — with zero code changes.

Runs on Consumer Hardware

Full training on a 6 GB laptop GPU. No A100s, no cloud compute, no multi-GPU setups. Designed for researchers who don't have institutional compute access.

---

Results

All results: --train_ratio 0.2 --val_ratio 0.1 --seed 42, forest routing, NVIDIA RTX 3050 (6 GB).

Indian Pines

Config	OA	mIoU	Dice	Kappa	Time
64 base / 32 num filters	97.52%	0.8593	0.9038	0.9717	6h 2m
32 base / 8 num filters	97.55%	0.8027	0.8391	0.9721	77m

Per-class IoU (64/32 config)

Class	IoU	Class	IoU
1	0.9143	9	0.1905
2	0.9351	10	0.9547
3	0.9425	11	0.9689
4	0.8811	12	0.9607
5	0.9415	13	0.9592
6	0.9961	14	0.9694
7	0.4286	15	0.9890
8	0.9256	16	0.7910

Classes 7 and 9 have <30 training samples. Low IoU on these is expected and consistent with published literature.

Salinas Valley

Results pending re-run with latest pipeline. Space reserved.

Config	OA	mIoU	Dice	Kappa	Time
—	—	—	—	—	—

Pavia University

Results pending re-run with latest pipeline. Space reserved.

Config	OA	mIoU	Dice	Kappa	Time
—	—	—	—	—	—

LUSC (Lung Squamous Cell Carcinoma)

Single image only (1 of 62). Not comparable to published benchmarks. Trained on a 512x512 crop with same-region pixel split. Published methods use patient-level cross-validation across 10 patients. These numbers demonstrate data-agnostic capability, not SOTA claims.

Metric	Value
OA	99.43%
mIoU	88.96%
Dice	93.49%
Kappa	0.9878

See LUSC_ghost_report.md for full details on why these results are not directly comparable.

Mars CRISM / Asteroid Ryugu

Tested on planetary remote sensing data. Results in early exploration phase — not benchmarked against published methods.

---

Quick Start

Install

pip install ghost-hsi

Train

ghost train_rssp \
  --data data/indian_pines/Indian_pines_corrected.mat \
  --gt   data/indian_pines/Indian_pines_gt.mat \
  --loss dice --routing forest \
  --base_filters 32 --num_filters 8 \
  --forests 5 --leaf_forests 3 \
  --epochs 400 --patience 50 --min_epochs 40 \
  --out-dir runs/indian_pines

Predict

ghost predict \
  --data  data/indian_pines/Indian_pines_corrected.mat \
  --gt    data/indian_pines/Indian_pines_gt.mat \
  --model runs/indian_pines/rssp_models.pkl \
  --routing forest --out-dir runs/indian_pines

Visualize

ghost visualize \
  --data    data/indian_pines/Indian_pines_corrected.mat \
  --gt      data/indian_pines/Indian_pines_gt.mat \
  --model   runs/indian_pines/rssp_models.pkl \
  --dataset indian_pines --routing forest \
  --out-dir runs/indian_pines

---

Hardware Requirements

Setting	VRAM	Flags	Indian Pines Time
Lightweight	4 GB	--base_filters 16 --num_filters 4 --d_model 32	~30m
Standard	6 GB	--base_filters 32 --num_filters 8	~77m
Full	8 GB+	--base_filters 64 --num_filters 32	~6h

No multi-GPU support. Single consumer GPU is all you need.

---

Data Format

GHOST accepts .mat files (MATLAB/HDF5 format):

• Data file: 3D array with shape (H, W, Bands) — the hyperspectral cube
• Ground truth file: 2D array with shape (H, W) — integer class labels, 0 = background

Keys inside the .mat file are auto-detected by array dimensionality. No configuration needed.

Standard datasets (Indian Pines, Pavia University, Salinas Valley) are available from the GIC group at UPV/EHU.

---

Architecture Overview

.mat file (H, W, Bands)
    |
    v
Continuum Removal ---- physics-based normalisation, no PCA
    |
    v
Spectral 3D Conv ----- learns cross-band features, kernel (7,3,3)
    |
    v
SE Attention ---------- per-channel importance weighting
    |
    v
2D U-Net -------------- multi-scale spatial context
    |
    v
RSSP Tree ------------- recursive binary class splitting
    |                   each node: independent forest ensemble
    v
Prediction Map (H, W)

See architecture.md for technical details.

---

Documentation

Document	Description
Architecture	Pipeline components, RSSP tree, training details
API Reference	All CLI commands and flags
TODO	Roadmap, known limitations, planned features

---

License

Proprietary. All rights reserved. See LICENSE.

For source code access, research collaborations, or licensing inquiries, contact the author directly.

---

Citation

If you use GHOST in your research, please cite:

@software{ghost2026,
  title  = {GHOST: Generalizable Hyperspectral Observation \& Segmentation Toolkit},
  author = {Ishu},
  year   = {2026},
  url    = {https://pypi.org/project/ghost-hsi/}
}