Cross-Architecture Foundation-Model Distillation for Edge Flood Segmentation
Fabian Schmalstieg, Karsten Mueller, Wojciech Samek
cs.CV · cs.LG
Quick verdict
Rigorous Prithvi-300M→0.7M distillation yielding a 1.5MB INT8 engine (5.57ms, Jetson NX); honest, but the matched-budget gain is geometry-confounded.
Summary
The paper distills a 300M-parameter Prithvi-EO-2.0 teacher, fine-tuned on Sen1Floods11's 252 hand-labeled scenes, into a 0.7M-parameter EfficientViT-B0 student for Sentinel-2 water segmentation. The frozen teacher pseudo-labels 23,720 unlabeled GDACS flood-event tiles; a pre-filter plus farthest-point sampling curate a 2,500-scene training pool without new manual annotation. At the matched 252-scene budget, teacher-supervised training edges out direct training on STURM-Flood in all five architecture-objective cells (+0.017 to +0.024 mIoU per flood map; four survive Holm correction), but the authors' own geometry-matched control collapses the OFA gap to +0.006 (p=0.071), showing acquisition geometry confounds the label-source effect. Scaling to 2,500 scenes narrows the student-teacher gap: 0.787 vs 0.822 water IoU on the Sen1Floods11 test split, parity on STURM (0.756 vs 0.755), still below on WorldFloods-v2. A fixed MNDWI threshold is competitive on both clean external benchmarks, which the authors honestly frame as generalization evidence only. After ReLU6 retraining and quantization-aware training, the student deploys as a 1.5 MB mixed INT8/FP16 TensorRT engine at 5.57 ms per 512x512 image on a Jetson Xavier NX (~14 MB device memory). The statistical protocol (Wilcoxon signed-rank, Holm correction, scene-level replication, contamination audits) is unusually careful; the main residual risks are benchmark-guided development and evaluation-scale sensitivity.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- At a matched 252-scene budget, teacher-supervised training is a competitive substitute for direct training and raises STURM-Flood mIoU in all five architecture-objective cells (four of five survive Holm correction), though a geometry-matched control shows label source alone does not explain the difference.
- Scaling the teacher-supervised pool from 252 to 2,500 scenes narrows the student-teacher gap: 0.787 vs 0.822 water IoU on Sen1Floods11 test, parity on STURM-Flood (0.756 vs 0.755), still below on WorldFloods-v2.
- After ReLU6 retraining and quantization-aware training, the student runs as a 1.5 MB mixed INT8/FP16 TensorRT engine on a Jetson Xavier NX at 5.57 ms of GPU compute per 512x512 image with about 14 MB of device memory.
- Because a fixed MNDWI threshold is competitive on both clean external benchmarks, external accuracy evidences generalization of the compact model rather than superiority over a spectral rule.
- Teacher supervision amplifies a fixed 252-scene annotation budget into a 2,500-scene training set without new labels, while transferring the teacher's surface-water annotation semantics (96.1% agreement on flagged pre-event baselines).
Strengths
Statistical protocol far above field norm
Unit of replication is the flood scene, not the pixel or seed; paired Wilcoxon signed-rank with Pratt zeros, Holm correction over the five-cell family, exact sign test as conservative supplement (only 2/5 survive), event-level coarsening for headline claims, one 10,000-sample bootstrap CI, and an explicit micro/macro aggregation split (§IV-A, §S1).
Self-undermining honesty rarely seen at this density
A geometry-matched control that collapses their own headline effect (OFA +0.006, p=0.071; CE -0.0004, §IV-C), a competitive MNDWI baseline bounding external claims (§IV-D), a teacher-semantics audit (96.1% agreement on flagged baselines, §IV-G), and an explicit benchmark-guided-development admission (§VII).
End-to-end deployment chain is measured, not asserted
Contamination auditing is explicit and quantified
Strong reproducibility package
Weaknesses
The central matched-budget claim is confounded by the authors' own control
Curation sophistication outruns its measured benefit
Headline STURM parity is evaluation-protocol-dependent
Benchmark-guided development without an untouched holdout
Deployment framing has soft edges
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper attacks the gap between geospatial foundation models (GFMs), which are accurate but too large for edge hardware, and deployable flood mapping on Sentinel-2. The pipeline: (1) a frozen, unmodified public Prithvi-EO-2.0-300M teacher (Sen1Floods11 checkpoint; reproduced at 0.897 mIoU / 0.822 water IoU on the S11 test split vs 0.900/0.826 published); (2) teacher pseudo-labeling of 23,720 unlabeled GDACS Sentinel-2 tiles; (3) curation — a pre-filter (flood fraction 0.05-0.85, confidence sigma >= 0.05) to 5,987 tiles, then farthest-point sampling over a 15-D descriptor to N=2,500; (4) cross-architecture distillation into EfficientViT-B0 (0.7M params) with OFA branch heads and a soft-target term; (5) ReLU6 replacement plus QAT into an INT8 TensorRT engine on Jetson Xavier NX. Claims are deliberately narrow — substitution (not superiority) of teacher labels at matched budget; gap narrowing (not parity) under scaling; a measured deployment footprint. Contribution 1 states +0.017 to +0.024 mIoU per flood map for pseudo vs S11 labels in all five architecture-objective cells on STURM-Flood, four of five surviving Holm correction (§IV-C).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper do in one sentence?
It compresses NASA/IBM's Prithvi-EO-2.0 (a 300-million-parameter satellite foundation model) flood-mapping ability into a 0.7-million-parameter EfficientViT-B0 network that runs as a 1.5 MB quantized engine on a Jetson Xavier NX edge device, using the big model's automatic labels on 23,720 flood-event Sentinel-2 images instead of new manual annotations.
How accurate is the distilled student model compared to the original 300M teacher?
On the Sen1Floods11 test split the student reaches 0.787 water IoU versus the teacher's 0.822; on the out-of-distribution STURM-Flood benchmark it matches the teacher under the paper's evaluation protocol (0.756 vs 0.755 mIoU); on WorldFloods-v2 it remains below. A simple fixed MNDWI threshold is competitive with both on the two external benchmarks.
Do teacher-generated labels beat human labels in this setup?
Not clearly. At the same 252-scene budget, teacher labels edge out human labels on STURM-Flood, but the authors' own geometry-matched control shows most of that advantage reflects differing image footprint rather than label quality; the real benefit of the teacher route is that its labels can be produced at scale (2,500 scenes) without new annotation cost.
Is the code and data available for this paper?
Yes. The paper points to a public repository (github.com/sycz00/FloodDistill) containing dataset construction and curation code, the pseudo-label pool identifiers and curation index, experiment configurations, ONNX exports and TensorRT build settings; the teacher is the unmodified public Prithvi-EO-2.0 Sen1Floods11 checkpoint, and all three benchmarks are open datasets.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction truncated portions of §III-C, Tables I and II contents, and some figure-only material (Fig. 2 scaling curves not analyzable); key numbers were cross-checked against prose, Tables III/IV/S1/S3, which survived extraction.
- The abstract in the metadata block is cut mid-sentence; the full-text version was used for interpretation.
- Supplementary repository contents (code, pool manifests) were not verifiable from the text alone.
- arXiv id 2609.20441 is dated September 2026; pool acquisition extends to July 2026, consistent with the stated timeline but not independently verifiable here.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.