Hyperspectral imaging · Foundation-model transfer · PyTorch

Transfer tested.
Failure preserved.

SpectraShift tests whether a wavelength-aware pretrained remote-sensing model can transfer to a new 285-band sensor configuration under limited labels. The benchmark preserves the useful mineral-ranking gain and the calibration failure that prevents a deployment claim.

Inspect the source Open the real-data run

Spectral bands: 285
Frozen parameters: 111.2M
Trainable parameters: 2,307
Hematite AUPRC: 0.533

Research question

Can a pretrained spectral representation survive a new sensor?

Hyperspectral foundation models promise faster adaptation across satellites, but a result is only useful when wavelength identity, geographic leakage, thresholds and model provenance are controlled. SpectraShift turns those boundaries into an executable evaluation protocol.

The system adapts DOFA ViT-B with a frozen encoder and linear probe, compares it with spectral-angle mapping and PCA-logistic baselines, reports ranking, segmentation, calibration and latency metrics, then converts detections into evidence-linked targets for analyst review.

Evaluation path

Four boundaries from raw cube to ranked evidence.

01Ingest
Load the hyperspectral cube, wavelengths and pseudo-label masks with explicit shape and finite-value checks.
02Verify
Pin dataset, source and checkpoint identity; split by source capture before any learned transform.
03Adapt
Freeze the wavelength-aware encoder and fit the small multi-label head using training-only statistics.
04Evaluate
Compare baselines, transfer thresholds from validation, measure calibration and retain failure evidence.

Validity controls

Sensor adaptation needs more than a higher score.

The benchmark makes leakage, provenance, thresholding and evidence authority explicit so the result can be reproduced and challenged.

Wavelength identity

Preserve all 285 EMIT band centres and pass them to the wavelength-aware encoder in the recorded sensor order.

Source-safe geography

Group tiles by source capture, exclude five crossing groups and prevent one acquisition from appearing across train and held-out partitions.

Training-only statistics

Fit normalization, PCA and supervised components on training data only; validation selects thresholds without touching the test tile.

Pinned model identity

Record the DOFA source revision, checkpoint revision and SHA-256 digest so a later run can resolve the same weights.

Frozen adaptation

Keep 111,199,232 encoder parameters frozen and train only a 2,307-parameter linear head, isolating representation transfer from full fine-tuning.

Evidence boundary

The analyst layer retrieves cited, immutable Markdown and JSON evidence. It cannot alter a prediction, threshold, rank or evidence grade.

Held-out tile

Representation improved. Calibration did not.

All metrics use one source-group-held-out test tile. Higher is better for F1, AUPRC and mIoU; lower is better for ECE and latency.

Model	Macro F1	Macro AUPRC	Mean IoU	ECE lower is better	Latency per tile
Spectral-angle mapping	0.199	0.212	0.119	0.164	79 ms
PCA + logistic	0.055	0.242	0.029	0.487	182 ms
DOFA frozen probe	0.273	0.225	0.182	0.776	603 ms

Mineral-specific resultDOFA improved hematite AUPRC from 0.266 for spectral-angle mapping to 0.533. That ranking gain is useful evidence of representation transfer, but it does not override the poor calibration or all-positive threshold maps.

Implementation

A small trainable surface around a pinned encoder.

ModelDOFA ViT-B · wavelength-conditioned patch embedding

Adaptation111,199,232 frozen parameters · 2,307-parameter linear head

Data285-band EMIT tiles · source-group-safe 4 / 3 / 1 split

BaselinesSpectral-angle mapping · PCA with logistic regression

EvaluationMacro F1 · AUPRC · mIoU · ECE · Brier score · latency

Quality27 tests · real-data CI · pinned revisions and checkpoint digest

Verification evidence

What SpectraShift proves today.

01Real sensor pathThe public workflow executes the complete 285-band pipeline rather than reducing the main claim to a synthetic smoke test.

02Model identitySource revision, checkpoint revision and checkpoint SHA-256 are recorded and checked before evaluation.

03Leakage controlTiles remain grouped by source capture, with five groups excluded because they crossed candidate partitions.

04Baseline contextThe frozen probe is interpreted against two transparent traditional methods, not as an isolated headline metric.

05Failure visibilityCalibration, threshold-transfer and weak kaolinite performance remain first-class outputs rather than being hidden behind the hematite gain.

Known limits

This is surface evidence, not a deposit discovery.

The benchmark uses pseudo-labels and only one held-out test tile. Its 112 × 112 inputs produce a coarse 7 × 7 token grid, and validation-selected thresholds became all-positive for goethite and hematite on the held-out geography.

The current result can support representation research and target-screening experiments. It cannot validate underground mineral deposits, replace field geology or justify operational deployment. The next work is a random-initialization ablation, higher spatial resolution and explicit post-hoc calibration.