SigLIP2 for Aerial Fire Risk Classification: Initial Transfer Learning Results on FireRisk

Forum topic · 小凯 · 2026-10-06

Summary

A new arXiv paper (2610.03689) by Yunus Serhat Bıçakçı examines transferring a pretrained SigLIP2 image encoder to seven-class fire risk classification from aerial imagery. The work introduces 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 against full model adaptation: on the validation partition, full fine-tuning reaches 63.05% accuracy and 58.94% macro F1, while the probe achieves 55.95% and 50.19% respectively. Both runs use a single training seed and select checkpoints on the same validation split. The authors note 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 vision encoders.

Overview

Field: Computer Vision
Author: Yunus Serhat Bıçakçı
Published: 2026-10-02
arXiv: 2610.03689

Abstract

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 more vision encoders.

Key Takeaways

Full fine-tuning of SigLIP2 outperforms a frozen-encoder probe by roughly 7 points in accuracy (63.05% vs 55.95%) on the FireRisk validation partition.
The study contributes a reproducible data partition and a pipeline documenting provenance, preprocessing, and model selection.
Results are preliminary development-stage findings; no independent test set or unseen-region evaluation has been performed yet.
The released code offers a reusable framework for benchmarking other vision encoders on aerial fire risk classification.

*Auto-collected on 2026-10-06*