Source-linked AI summary
SigLIP2 for aerial fire risk classification
Yunus Serhat Bıçakçı
TL;DR
FireRisk hazard classification from aerial imagery involves map-derived categories that can be difficult to distinguish from visible surface cues. This study creates a reproducible split and compares frozen and fully adapted SigLIP2 encoders; full adaptation reaches 58.94% validation macro F1 versus 50.19% for the probe. The results are development measurements, not an independent-test or unseen-region evaluation.
Problem
Aerial images may contain cues for wildfire hazard categories, but similar vegetation can span levels and the map-derived labels incorporate information beyond RGB, making within-category recognition difficult to establish.
Source: §1 · p. 1Method
The study creates a reproducible FireRisk mirror split and compares a frozen SigLIP2 image-encoder probe with full adaptation to hazard labels.
Source: §1 · p. 1Results
58.94% validation macro F1 for full SigLIP2 adaptation versus 50.19% for the frozen encoder probe on the fixed mirror split.
Source: §8 · p. 6Takeaways & Limitations
The results support further evaluation of SigLIP2 within a transparent framework for repeated experiments and comparisons with other visual encoders.
Source: §1 · p. 1
Abstract
from arXiv · showhide
We examine the transfer of a pretrained SigLIP2 image encoder to seven class fire risk classification from aerial imagery. We introduce a reproducible partition of the public FireRisk training mirror and an implementation that records data provenance, preprocessing and model selection. Two initial runs compare a frozen encoder probe with full model adaptation. On the validation partition, full adaptation reaches 63.05% accuracy and 58.94% macro F1, compared with 55.95% and 50.19% for the probe. Both runs use one training seed and select their checkpoint on the same validation partition. These development results support further evaluation of SigLIP2 but do not establish performance on an independent test set or unseen regions. The accompanying code provides a common framework for repeated experiments and comparisons with additional visual encoders.
1 Introduction
The paper studies whether a pretrained SigLIP2 image encoder transfers to aerial-image classification of wildfire hazard categories. It establishes a reproducible starting point for comparing frozen and adapted encoders while distinguishing hazard classification from fire forecasting.
1 Introduction
- FireRisk labels derived from a wildfire hazard map distinguish categories of relative potential for wildfire that would be difficult to suppress, not ignition or fire-consequence forecasts.Aerial imagery may help distinguish categories through vegetation, buildings, and surface cover, but the labels incorporate information beyond the displayed RGB crop.Source: §1 · p. 1
- Similar vegetation can occur across hazard levels, and aggregate accuracy alone cannot establish whether a classifier learns distinctions within those categories.The five ordered hazard levels are more demanding than separating water, vegetation, and built surfaces.Source: §1 · p. 1
- The study asks how SigLIP2’s image encoder transfers to aerial imagery by comparing a frozen representation with one updated using hazard labels.The frozen probe tests information already present in the representation; adaptation lets it change with the labels but changes optimization demands.Source: §1 · p. 1
- The paper supplies a reproducible split, audited data access, two initial adaptation results, and error analysis as a transparent basis for further comparisons.The public mirror lacks the original evaluation partition, so the study claims neither a new best benchmark result nor a validated system for unseen locations.Source: §1 · p. 1
2 Related work
Prior work spans FireRisk benchmarks, remote-sensing-specific pretraining, aerial segmentation, and reusable vision-language encoders. This paper evaluates supervised SigLIP2 image-encoder adaptation for tile-level hazard classification, distinct from those other tasks and settings.
2 Related work
- FireRisk provides supervised and self-supervised representation benchmarks for aerial fire-hazard classification, while this study uses a separately declared partition.The paper examines transfer from a more recent pretrained image encoder.Source: §2 · p. 2
- SatMAE uses temporal and spectral encodings, while RemoteCLIP adapts vision-language learning to remote sensing; neither model is evaluated in the reported runs.Their published results do not establish performance under this study’s split.Source: §2 · p. 2
- Aerial building segmentation with attention gates and a Transformer targets building masks, unlike this study’s single RGB-derived hazard category per tile.Segmentation performance therefore does not imply hazard-classification performance.Source: §2 · p. 2
- Prior SigLIP work includes image-text training and geolocation retrieval, whereas this study adapts only the SigLIP2 image encoder to aerial hazard labels.It does not evaluate text prompts, language generation, or geographic retrieval.Source: §2 · p. 2
- The code includes a DINOv2 preset for a later comparison, but this paper reports no DINOv2 result.The proposed comparison could distinguish pretraining content from adaptation strategy.Source: §2 · p. 2
3 Data and protocol
The study uses the public FireRisk training mirror, which lacks the published evaluation partition, to create a documented, stratified train-validation-reserved-test split. Scores are development measurements because the reserved test set is unevaluated.
3 Data and protocol
- The image audit confirms 70,331 320 × 320 RGB PNGs; Figure 1 shows one uniformly sampled training image per class, selected with seed 2026 without visual screening.These examples illustrate inputs and labels, not whether hazard can be inferred reliably by visual inspection.Source: §3 · p. 2, Figure 1 · p. 3
- The new split allocates 49,231 images to training, 10,552 to validation, and 10,548 to reserved test, stratified by class with seed 2026.It preserves the observed class imbalance across seven categories.Source: §3 · p. 2
- Figure 2 summarizes mirrored class counts and their allocation; nonburnable and very low together comprise 56.47% of the mirror, while very high comprises 4.65%.The validation partition contains 30.93% very low examples, compared with 4.64% very high and 2.46% water.Source: §3 · p. 2, Figure 2 · p. 3
- Macro F1 weights the seven categories equally, while class recall reveals missed categories; the checkpoint is selected using validation macro F1.The validation set retains observed class frequencies rather than being resampled.Source: §3 · pp. 3–4
- Pixel hashes found 70,331 unique image groups and no exact duplicates, but cannot rule out overlapping or nearby tiles or establish geographic or temporal independence.The mirror lacks coordinates and acquisition dates.Source: §3 · p. 3
- The reserved test partition is unevaluated, so validation-selected scores are development measurements and cannot be compared directly with the original FireRisk benchmark.The mirror exposes only training examples and does not provide the published evaluation partition.Source: §3 · pp. 2, 4
4 Model and training
The system applies a seven-class head to SigLIP2 pooled image representations, with either a frozen encoder or full adaptation. The two training recipes differ in optimization and preprocessing, so their comparison does not isolate encoder adaptation alone.
4.1 Classifier and objective
- Full adaptation updates the SigLIP2 encoder, while the frozen probe trains only its 6,919-parameter head; the full classifier has 92,891,143 parameters.Both retain pretrained attention pooling and use trainable LayerNorm with a seven-output linear layer.Source: §4.1 · p. 4
- Both recipes minimize smoothed categorical cross-entropy over seven classes with ε = 0.05, without an image-text pairing objective during adaptation.The frozen recipe updates the head and LayerNorm parameters while keeping the encoder fixed.Source: §4.1 · p. 4
4.2 Preprocessing and optimization
- Training uses randomized crops, flips, and right-angle rotations; evaluation uses deterministic bicubic resizing, the pretrained center-crop policy, and channel normalization.Training crops cover 80%–100% of the image area.Source: §4.2 · p. 4
- Both runs use AdamW, batch size 64, gradient clipping, and cosine learning-rate schedules after warmup, but use different learning rates and maximum epoch counts.They share training seed 42; full adaptation also uses gradient accumulation and checkpointing.Source: §4.2 · p. 4
- Both runs stop after 14 epochs and select epoch 7, but differing learning rates, dropout, and schedule lengths prevent attributing their score difference to encoder adaptation alone.Early stopping requires seven epochs without a validation macro F1 improvement greater than 0.0001.Source: §4.2 · p. 4
5 Preliminary results
Full adaptation outperforms the frozen encoder probe on the validation partition. Class-level results show that F1 improves across categories, while some hazard categories remain difficult to identify.
5 Preliminary results
- Full adaptation improves validation accuracy by 7.10 percentage points and macro F1 by 8.75 percentage points relative to the probe.The reported validation results are from one training seed.Source: §5 · p. 4, Table 1 · p. 5
5.1 Errors within hazard categories
- Full adaptation has higher F1 for every category, but high recall falls from 46.56% to 45.08%.Table 2 reports support, F1, and recall for each class on the same validation images.Source: §5.1 · pp. 4–5, Table 2 · p. 5
- Full adaptation correctly identifies 164 of 490 very high examples, for 33.47% recall.The passage notes that this recall limits usefulness when missing high-hazard areas has substantial cost.Source: §5.1 · p. 5
- Under full adaptation, 29.70% of low and 26.88% of high examples shift down a hazard category; 51.02% of very high cases are classified high or moderate.Figure 3's matrices normalize predictions within each true class, with rows summing to 100% before rounding.Source: §5.1 · p. 5, Figure 3 · p. 5
- Exact accuracy on the five ordered hazard levels rises from 44.50% to 54.01%; 723 of 749 added correct predictions come from these classes.Predictions of water or nonburnable remain errors in the focused calculation.Source: §5.1 · pp. 5–6
- No repeated training seeds are available, so variation from training randomness is unknown; concurrent runs on separate GPUs do not provide a controlled speed comparison.Reported timings exclude environment installation, downloads, and the initial pixel audit.Source: §5.1 · p. 6
6 Discussion
The comparison does not isolate the effect of encoder updates, and residual errors and unresolved geographic independence constrain what the results support.
6 Discussion
- The two recipes do not separate benefits from encoder updates from differences in head optimization and schedule.A controlled follow-up should match search budgets and schedules, with independent seeds to measure training variation.Source: §6 · p. 6
- Very high recall remains 33.47%, while errors frequently shift examples toward lower hazard categories.Class weighting, alternative sampling, and ordinal objectives are validation candidates, not demonstrated improvements; the current loss treats errors categorically and does not encode hazard-level distance.Source: §6 · p. 6
- Without spatial holdouts, the study cannot establish geographic independence or quantify any spatial bias in this mirror.The model predicts map-derived labels from RGB without weather, ignition, or exposure inputs, so its outputs are not supported as future fire probabilities or operational warnings.Source: §6 · p. 6
7 Reproducibility
The implementation records data and experiment provenance while preserving the original shards. Further experiments should use a committed release and declared validation search budgets before reserved-test evaluation.
7 Reproducibility
- Each run saves its configuration, environment, preprocessing, data manifest, history, and selected weights.The documentation lists immutable revisions and the split hash, though initial artifacts record a modified development tree without a source commit.Source: §7 · p. 6
- Original Parquet shards occupy about 10.78 GiB, and shard access avoids the similar additional allocation required by the optional image archive.The validation comparison can be regenerated from saved predictions, and analysis scripts record their inputs and numerical provenance.Source: §7 · p. 6
- Final experiments should use a committed release and declared validation search budgets before evaluating fixed recipes on the reserved test partition.Possible overlap with pretraining data remains unknown.Source: §7 · p. 6
8 Conclusion
In a preliminary comparison, full SigLIP2 adaptation outperforms a frozen encoder probe on validation macro F1. The results are a starting point; repeated experiments and independent evaluation remain necessary, with code and checkpoints available.
8 Conclusion
- 58.94% validation macro F1 for full SigLIP2 adaptation compares with 50.19% for a frozen encoder probe on one preliminary fixed-split comparison.Repeated controlled experiments and independent evaluation remain necessary before claiming reliable transfer to new regions.Source: §8 · pp. 6–7
- Code is available under the GNU General Public License version 3, with initial adaptation and frozen-probe checkpoints, inference instructions, and validation records.Upstream imagery and pretrained weights retain their own terms.Source: §8 · p. 7