Future work

This page records the follow-on work scoped for eo-data-embedding. Every item below is method- and data-identified, not vague “future work”: each names the published approach and the dataset that would carry it. The common reason none are done here is a single, external constraint rather than a design gap — the project ran into a hard GPU quota/capacity wall (single GPU slot, GPUS_ALL_REGIONS = 1, auto-denied quota bump, Spot preemption, on-demand T4 stockouts across EU zones) and deliberately pivoted to a CPU path to keep delivery honest (see the V&V report compliance/drd/vv-report.md §3 and research/07-engineering-notes.md). What follows is what the roadmap opens up once accelerator and storage are no longer the binding constraint.

The items are ordered by how directly they attack a verified limitation; the first two close open requirement items in the requirement verification status of the V&V report (compliance/drd/vv-report.md §4).

1. Archive-scale embedding index — the “ML at scale” thesis

The published index is built over a 2,000-tile EuroSAT subset; the “find scenes like this across the archive” claim is therefore demonstrated, not stress-tested. With accelerator time for the one-time encode pass and storage for the resulting vector store, the index would scale to a real corpus:

• Data. SSL4EO-S12 (~1M+ unlabelled S1+S2 patches — the designated “scale” corpus, see research/01-datasets.md) and BigEarthNet-MM (~549k paired patches).

• Method. Embed once with frozen Clay v1.5, persist millions of 1024-d vectors to the parquet store, and back the FAISS index with an ANN structure (IVF/HNSW/PQ) sized for the corpus rather than the exact IndexFlatIP used at demo scale.

• What it proves. Retrieval quality and latency at archive scale — the load-bearing evidence behind REQ-F-03 beyond the EuroSAT smoke set.

• Why blocked. The heavy encode pass needs sustained GPU the trial quota never granted; the vector store needs storage well past the demo bundle.

2. Time-series change detection — the open scientific frontier

This is the #1 open item: it closes the phenological-layer limitation behind the partially verified REQ-F-05. Zero-training two-date embedding distance is rejected by experiment (ROC-AUC at or below chance) because of a seasonality/phenology confound — unchanged vegetated tiles move more in embedding space than truly changed urban ones, and the phenological layer (genuinely different ground cover, not a colour shift) cannot be normalized away or resolved with only two dates (V&V report §3; research/06-change-analysis.md §8–9).

• Method. Model a location’s embedding trajectory across many dates (periodic regression / change-point detection); real change is a departure from the seasonal cycle. This is the literature’s answer — SeCo’s motivation, Element84’s periodic-outlier approach, OPTIMUS’s change-point framing.

• Data. Multi-date stacks that OSCD’s two dates cannot supply — SpaceNet 7 (monthly mosaics with building-change labels) or DynamicEarthNet (daily Planet imagery, monthly semantic labels), or raw Sentinel-2 stacks via the Microsoft Planetary Computer.

• Why blocked. Embedding every date multiplies the encode cost, and multi-date stacks multiply storage — exactly the two resources the trial setup lacked.

3. Seasonally-invariant or better-matched encoders

The frozen-FM thesis was tested with one backbone (Clay v1.5). A natural ablation swaps or compares against encoders pretrained with seasonal-contrast objectives so that phenology is suppressed in the embedding rather than learned around downstream.

• Method. Compare frozen Clay against SeCo / CaCo (seasonal contrast) and against Prithvi / DOFA-CLIP, holding the rest of the pipeline fixed.

• What it proves. Whether the label-efficiency and change-detection results are a property of the thesis (any strong frozen FM) or of Clay specifically — and whether a seasonally-invariant backbone recovers the change signal that §2 attacks from the data side.

• Why blocked. Loading and running several large FMs over the evaluation sets is GPU-bound.

4. Trained cross-modal alignment (SAR ↔ optical)

Clay’s frozen embeddings are only weakly cross-modal (SAR→optical P@1 = 0.042, ~5× chance) because Clay has no cross-modal training objective. The project’s honest result is that a single learned 1024×1024 linear map on just 180 pairs already lifts this to P@1 = 0.142 (~17× chance, median rank 32→8): the modalities are linearly relatable without joint training (V&V report §3).

• Method. Replace the small linear map with a properly contrastive cross-modal alignment trained on a large paired corpus (BigEarthNet-MM, SSL4EO-S12), targeting the DOFA-CLIP band of purpose-built cross-modal retrieval.

• Why blocked. Contrastive training over a large paired corpus needs both the accelerator and the corpus storage.

5. Frozen-vs-fine-tuned ablation

By design the encoder is never trained — the frozen pass is the project’s load-bearing decision (see the design overview). A bounded ablation would quantify the ceiling that decision leaves on the table.

• Method. Compare the frozen probe against lightweight fine-tuning (e.g. LoRA adapters on the encoder) on the same splits and seeds.

• What it proves. The size of the frozen-vs-tuned gap — sharpening the “frozen embeddings buy label-efficiency, not supremacy” finding (at the full label pool the from-scratch CNN already wins 0.949 vs 0.920).

• Why blocked. Fine-tuning, even parameter-efficient, reintroduces the GPU requirement the project pivoted away from.

Summary

None of the above is blocked by the software design: the pipeline already threads device end to end (map_location + .to(device)), the store and index are corpus-agnostic, and the encoder is swappable behind the embed interface. The binding constraints were GPU quota and storage, recorded plainly rather than hidden. With those removed, items 1–2 are the highest-value next steps — they close open verification items (REQ-F-03 at scale, REQ-F-05’s phenological layer) — while 3–5 deepen the scientific claims the project already reports honestly.