WISE: A Lightweight, Weakly-Supervised Model for Onboard Fire Smoke Detection and Localization
Sha Lu, Yu Sun, Liang Zhao, Jixue Liu, Lin Liu, Jiuyong Li, A. K. Qin, Alejandro Mousist, et al.
cs.CV
Quick verdict
Promising weakly supervised onboard smoke detector (0.12M params, in-orbit demo), but pixel eval rests on 11 tiles, test-set-tuned thresholds; no code release.
Summary
WISE is a deployment-oriented framework for wildfire smoke detection from satellite imagery that requires only tile-level labels. A ResNet-50 teacher is fine-tuned for three-class tile classification (clear/other/smoke), and fused multi-layer Grad-CAM maps (Layers 2-4, depth-weighted) serve as spatial pseudo-labels for a lightweight dual-branch student that outputs a tile-level prediction and a 256x256 smoke probability map in one forward pass. Trained on 1,836 Landsat-5/8 tiles (6 bands, 64/16/20 split), four width variants span 0.008-0.116M parameters and 0.23-3.04 GFLOPs. Over 10 runs, WISE-L reaches tile-level F1 0.964 and pixel-level F1 0.750 (IoU 0.627) on an 11-tile densely annotated subset; ablations show Grad-CAM supervision drives localization (F1 falls to 0.456 without it) and improves classification except for the smallest variant. Against tile-level onboard baselines (CloudScout, VIB-SD, MobileNetV2, PhiFire), WISE-M trails the best by about 1.6 F1 points but adds localization. Supervised segmentation references (VTrUNet, F1 0.826) still exceed the best WISE-L run (0.806). In-orbit execution on the ISS-mounted IMAGIN-e payload achieved 0.10-0.26 s per tile on CPU with image acquisition bypassed. Key limitations: pixel evaluation on only 11 tiles, thresholds selected on the test set, best-run reporting in the supervised comparison, no statistical testing, and no code or data release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 4
● low 0
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Tile-level labels alone, distilled via fused multi-layer Grad-CAM maps from a ResNet-50 teacher, suffice to train a 0.12M-parameter student that jointly outputs tile-level detection (F1 0.964) and pixel-level smoke probability maps (F1 0.750) in a single forward pass.
- Adding spatial distillation improves tile-level classification for all but the smallest variant (WISE-M F1 0.938 to 0.962, FNR 9.4% to 4.1%), acting as an implicit regularizer when capacity allows.
- The best WISE-L run approaches fully supervised smoke segmentation (F1 0.806 vs VTrUNet 0.826) despite never observing a pixel-level mask during training.
- WISE executes in orbit on the ISS-mounted IMAGIN-e payload at 0.10-0.26 s per 256x256 six-band tile on CPU, within that platform's resource limits.
Strengths
Single-pass dual output at extreme compactness
WISE-L produces both tile-level classification and a 256x256 smoke probability map in one forward pass at 0.116M parameters and 3.042 GFLOPs (Table 1); a full layer-by-layer configuration is given in Appendix Table A.1, supporting scrutiny of the architecture.
Genuine in-orbit execution evidence
Section 4.5 reports real timing aboard the ISS-mounted IMAGIN-e payload: 0.10/0.14/0.26 s per tile for WISE-S/M/L on CPU without GPU, with batch size, overlap, tile counts, and peak memory (582-903 MB) disclosed in Table 7 - rare deployment evidence for an academic model.
Ablation cleanly isolates the core contribution
Above-average reporting hygiene
Weaknesses
Tile-level performance does not lead the onboard baselines
Pixel-level evaluation set is only 11 tiles
Operating threshold selected on the test set
Inconsistent run selection in the supervised comparison
No code or data availability statement
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 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
Does the WISE model need pixel-level smoke annotations for training?
No. WISE trains only on tile-level (image-level) labels indicating whether a satellite tile contains smoke. A fine-tuned ResNet-50 teacher produces Grad-CAM attention maps that serve as soft spatial pseudo-labels for the lightweight student network, so no segmentation masks are used in training. Pixel-level evaluation in the paper did use a separate set of 11 manually annotated tiles to measure localization quality.
How accurate is WISE at detecting and localizing wildfire smoke from satellites?
In the paper's Landsat-5/Landsat-8 experiments averaged over 10 training runs, the largest variant (WISE-L, 0.116M parameters, ~3 GFLOPs) achieved a mean tile-level F1 of 0.964 for detecting smoke, and a mean pixel-level F1 of 0.750 (IoU 0.627) on a small set of 11 densely annotated tiles. The medium variant trades about one point of accuracy for roughly half the compute.
Has WISE actually been run in space?
Yes. The authors report in-orbit execution on the IMAGIN-e edge-AI payload mounted on the International Space Station, with CPU-only (no GPU) inference of 0.10 s, 0.14 s, and 0.26 s per 256x256 six-band tile for the S, M, and L variants. Image acquisition was bypassed during the demo, so the timings cover preprocessing, inference, and output generation rather than full end-to-end scene processing, and the payload is an ISS demonstration platform rather than a CubeSat.
Is the WISE code or dataset publicly available?
The paper text contains no code or data availability statement. The underlying Landsat tile dataset and the 11-tile pixel-annotated subset come from earlier published works (Zhao et al. 2022 and Liu et al. 2024), and the paper provides a detailed layer-by-layer architecture table and full training hyperparameters, but independent reproduction currently requires reimplementing the pipeline from those descriptions.
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.
- Figures (pipeline diagram, Grad-CAM visualizations, qualitative smoke maps in Fig 4) cannot be assessed from the extracted text, so qualitative claims rest on the authors' descriptions.
- Equations and tables are lightly garbled by text extraction; cell-level readings of Tables 2, 4, and 7 were reconstructed from captions and prose and could differ in layout details.
- The metadata author list omits Stefan Peters, who appears as last author in the full text; affiliation list was taken from the text.
- No supplementary material, review history, or code repository was available to verify implementation details.
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.