# Dataset Card: LWIR Hyperspectral Image Restoration Inputs

## Overview

This bundle supports ground-based long-wave infrared (LWIR) hyperspectral image restoration: per-band destriping of pushbroom cubes and physically grounded spectral recalibration of the camera band centers. It combines five scenes from the [DARPA IH Dataset](https://www.kitware.com/ihdataset/) with a **per-scene** atmospheric product: a zenith (vertically downward) downwelling spectral radiance computed for the conditions of each capture.

The downwelling files are **not** a shared U.S. Standard Atmosphere snapshot. For each capture, the zenith downwelling was computed with [libRadtran](http://www.libradtran.org/) (`uvspec`, DISORT, thermal emission) from a layered temperature and gas-density profile at the **nearest analysis time and nearest grid location** to that scene, plus high-resolution HITRAN line-by-line optical depths (H₂O, CO₂, O₃, NO₂). The product uses a 6–14 µm grid at 4 nm spacing.

The ENVI cubes are too large for GitHub. The image build fetches this same layout from the frozen Hugging Face dataset [dccc2025/tmp_lwir_image_restoration](https://huggingface.co/datasets/dccc2025/tmp_lwir_image_restoration) and checks SHA256. That fetch is pinned to a commit (not a moving `main` branch) and is not a live scrape of Kitware.

## 1. Source and provenance

The five hyperspectral cubes are taken from the DARPA IH Dataset, publicly released under Distribution Statement A (see the scene headers). The experimental data were provided by the U.S. Army Night Vision and Electronic Sensors Directorate (NVESD) and the Johns Hopkins University Applied Physics Laboratory, and were acquired with a pushbroom LWIR hyperspectral imager built around a cooled HgCdTe focal-plane array. The same instrument and data family underlie published LWIR hyperspectral studies, including heat-assisted detection and ranging (HADAR) and concurrent band selection / traversability estimation in off-road settings.

**Dataset homepage:** https://www.kitware.com/ihdataset/

### Source citation

```bibtex
@InProceedings{Yellin_2024_WACV,
    author    = {Yellin, Florence and McCloskey, Scott and Hill, Cole and Smith, Eric and Clipp, Brian},
    title     = {Concurrent Band Selection and Traversability Estimation from Long-Wave Hyperspectral Imagery in Off-Road Settings},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024}
}
```

## 2. Instrument and acquisition conditions

The sensor is a pushbroom LWIR hyperspectral imager with a cooled HgCdTe detector. For each spatial pixel the spectrometer records 256 spectral bands over the thermal LWIR window around 8.0–13.2 µm. Instrument characteristics as reported for this sensor:

| Property | Value |
|----------|-------|
| Focal plane array | HgCdTe, cooled |
| Spectral bands | 256, ~8.0–13.2 µm |
| Vertical field of view | 11.6° |
| Horizontal field of view | 57° |
| Focal length | 50 mm |
| f-number | f/0.9 |
| Typical noise | ~1 microflick (≈ 1000:1 SNR at 10 µm) |

All five scenes are night-time ground-based captures from April 2021 at Fort A.P. Hill, Virginia:

| Scene stem | Acquisition time (UTC) | Lines | Samples | Bands |
|------------|------------------------|------:|--------:|------:|
| `IHTest_202104_Path15_Step2_LWHSI1_collect0_DistStA` | 2021-04-16T17:53:52 | 260 | 1600 | 256 |
| `IHTest_202104_Path17_Step1_LWHSI1_collect0_DistStA` | 2021-04-16T21:43:22 | 260 | 1600 | 256 |
| `IHTest_202104_Path17_Step3_LWHSI1_collect0_DistStA` | 2021-04-16T21:52:54 | 260 | 1600 | 256 |
| `IHTest_202104_Path26_Step2_LWHSI1_collect0_DistStA` | 2021-04-20T23:13:02 | 260 | 1600 | 256 |
| `IHTest_202104_Path27_Step25_LWHSI1_collect0_DistStA` | 2021-04-21T03:24:47 | 260 | 1600 | 256 |

Because the instrument is a pushbroom system, each cube shows some level of along-track column striping from detector non-uniformity, and the strength of the striping varies from band to band. The header `temperature` field records near-surface **air temperature** at the site, in kelvin.

## 3. Hyperspectral scenes (`original_datasets/`)

Five ENVI-format LWIR hyperspectral cubes (`.bsq` + `.hdr`). Each pair is a band-sequential cube plus acquisition metadata.

- **Format:** ENVI BSQ, 32-bit float (`data type = 4`), little-endian.
- **Wavelengths:** A nominal band-center list is stored in each `.hdr`, in micrometers. Treat this axis as the instrument's **nominal** calibration, not ground truth: it is known to be offset, and atmospheric absorption features in the measured spectra do not sit where the header claims. Do not use the header axis for quantitative spectral work without recalibrating it first.
- **Distribution:** `Distribution Statement A` (see scene headers).

### Cube radiance units

The `.bsq` arrays are **not** already in
\(\mathrm{W\,m^{-2}\,sr^{-1}\,\mu m^{-1}}\). They are stored a factor of
\(10^{2}\) high relative to physical spectral radiance. Convert with

\[
L_{\lambda}\;[\mathrm{W\,m^{-2}\,sr^{-1}\,\mu m^{-1}}]
= 10^{-2}\,L_{\mathrm{bsq}}.
\]

Use this scaled \(L_{\lambda}\) in any comparison to Planck \(B_\lambda(T)\)
or to the bundled zenith downwelling (which is already in physical units).
Do not treat header `temperature` as the surface temperature map; it is
local air temperature.

## 4. Zenith downwelling radiance (`downwelling_radiance/`)

One file per scene, already in physical radiance units, computed with libRadtran as described above. A single column: **zenith only** (vertical downward / zenith angle 0°, surface observer, `umu = −1`). This is not an angle-resolved sky basis.

| Key | Shape | Description |
|-----|-------|-------------|
| `lambda` | `(1, 2001)` | Wavelength grid, **6–14 µm** at **4 nm** |
| `downwelling_res` | `(2001, 1)` | Zenith downwelling spectral radiance, **W m⁻² sr⁻¹ µm⁻¹** |
| `zenith_angle_deg` | `(1,)` | `[0]` |

Load example:

```python
import numpy as np

stem = "IHTest_202104_Path26_Step2_LWHSI1_collect0_DistStA"
data = np.load(f"downwelling_radiance/{stem}.npz")
lam = data["lambda"]                    # (1, 2001), µm
I_down = data["downwelling_res"]        # (2001, 1), W·m⁻²·sr⁻¹·µm⁻¹
zenith = data["zenith_angle_deg"]       # [0]
```

## Units and conventions

| Quantity | Symbol / key | Units |
|----------|--------------|-------|
| Wavelength | `lambda` | µm |
| Zenith downwelling spectral radiance | `downwelling_res` | W m⁻² sr⁻¹ µm⁻¹ |
| Hyperspectral cube | ENVI `.bsq` | stored \(10^{2}\) high; \(L=10^{-2}L_{\mathrm{bsq}}\) in W m⁻² sr⁻¹ µm⁻¹ |
| Header `temperature` | air temperature | K |

## Notes for use

1. **The header wavelength axis is nominal.** It carries a calibration error, so the atmospheric absorption structure in a cube (ozone, water vapor) will not line up with the same features in the downwelling reference if you take the header at face value. Recovering the operating wavelength axis — from the data, against the downwelling reference — is part of the work this bundle supports.
2. **Grid mismatch.** The downwelling `lambda` grid is 6–14 µm at 4 nm and does **not** match the camera bands. Once the camera axis is calibrated, project the reference onto it rather than resampling the cube.
3. **Paired products.** Each scene has its own matching zenith-downwelling file, computed from the nearest analysis time and nearest grid location to that capture. Pair files by scene stem; do not mix products across scenes. This per-scene pairing is more accurate than a shared standard atmosphere.
4. **Zenith only.** `downwelling_res` has shape `(2001, 1)`. There are no additional zenith-angle columns.