Hyperfm250k Eval

---
name: hyperfm250k-eval
description: Evaluates hyperspectral foundation models and task-specific deep learning architectures on pixel-level regression tasks for retrieving cloud optical and microphysical properties (COT, CER, CWP, CTH) from NASA PACE-OCI imagery. Use when the user wants to benchmark on HyperFM250K, or asks about evaluating this task. Reports MSE.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.21127
  bibtex_key: tushar2026hyperfm
  confidence: high
---

# hyperfm250k-eval

> HyperFM: An Efficient Hyperspectral Foundation Model with Spectral Grouping — Tushar et al. (2026) (arXiv:2604.21127, 2026)

## What this evaluates

Evaluates hyperspectral foundation models and task-specific deep learning architectures on pixel-level regression tasks for retrieving cloud optical and microphysical properties (COT, CER, CWP, CTH) from NASA PACE-OCI imagery.

## Datasets

- **HyperFM250K** — total ?; splits: train (2000), val (250), test (2000); repo https://github.com/umbc-sanjaylab/HyperFM

## Metrics

- `MSE` **(primary)** — range: other
  - Mean Squared Error between predicted and ground truth pixel values for each cloud property regression task.

## Input / output format

**Input**: Hyperspectral image patches of shape 291×96×96 (291 spectral bands, 96×96 spatial pixels). Baseline models using only two wavelengths receive 2×96×96 patches. Patches are extracted non-overlapping, with border patches using overlapping sampling and averaged predictions.

**Output**: Pixel-level regression values for COT, CER, CWP, and CTH. COT and CWP targets are log-transformed during training.

## Scoring recipe

```python
def compute_mse(predictions, targets):
    # predictions and targets: arrays of shape (N, 4) for COT, CER, CWP, CTH
    # If targets were log-transformed during training, apply inverse log to predictions
    mse = np.mean((predictions - targets) ** 2)
    return mse
```

## Common pitfalls

- COT and CWP targets are log-transformed before training, so predictions must be inverse-transformed or the metric must account for this to match reported values.
- Inference is performed patch-wise with overlapping borders at image edges, requiring averaging of overlapping predictions to reconstruct full-image retrievals.
- Baseline models require spatial resizing to 128×128 or spectral interpolation to match input dimensions, which can affect performance if not applied consistently.

## Evidence (verbatim from paper)

> As in related work, we use Mean Squared Error (MSE) as the evaluation metric for all regression tasks. ... The finetuning dataset contains 2000 training images, 250 validation images, and 2000 test images, and following[48], COT and CWP are log transformed before training.

## Citation

```bibtex
@misc{tushar2026hyperfm,
  title={HyperFM: An Efficient Hyperspectral Foundation Model with Spectral Grouping},
  author={Tushar et al. (2026)},
  year={2026},
  note={arXiv:2604.21127}
}
```

- arXiv: 2604.21127