Research map / Vision-language and multimodal / Multimodal LLMs

FireScope: Wildfire Risk Raster Prediction With a Chain-of-Thought Oracle

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces FireScope-Bench (US wildfire risk rasters from Sentinel-2 and climatology plus European EFFIS fire events) and FireScope, a GRPO-trained Qwen2.5-VL-7B Oracle with chain-of-thought whose scalar risk output conditions a FiLM-modulated raster decoder, improving cross-continental generalization.

Problem

Wildfire risk estimation requires integrating imagery, climate and geography to produce continuous high-resolution risk rasters, but vision models trained on imagery alone generalize poorly across regions and continents.

Approach

A two-stage reasoning-to-generation pipeline: a VLM Oracle is RL-trained (GRPO) to reason with CoT and output an ordinal risk score, and this scalar (via FiLM) conditions a lightweight vision encoder-decoder that generates the risk raster, aiming to transfer better OOD.

Key results

• CoT Oracle conditioned U-Net reaches OOD wildfire-event ROC AUC 0.750 vs 0.679 for image-only U-Net and Brier 0.191 vs 0.217.
• CoT Qwen Oracle achieves OOD ROC AUC 0.748 and Brier 0.196 versus 0.701 and 0.225 for non-CoT Qwen fine-tuned.
• Expert study: one expert reached QWK 0.33 from Oracle CoT summaries versus 0.50 with golden CoTs; consistency 0.91 and fidelity 0.33 for FireScope.

Surprising findings

• Climate-conditioned models perform best in-distribution but degrade sharply OOD, suggesting climate overfitting at coarse resolution.
• Oracle conditioning on a single scalar improves per-pixel OOD discrimination, which the authors describe as surprising.

Limitations

• Communication between Oracle and decoder is bottlenecked by a single scalar, limiting spatial granularity.
• Expert study used only two experts with varying results (QWK 0.33 and 0.11).
• Fine-tuned Oracle loses about 7.5 points on MMMU relative to base Qwen2.5-VL-7B.

Experimental setup

Datasets: Wildfire Risk to Communities risk rasters (USA) Sentinel-2 L2A imagery NASA POWER climatology EFFIS burnt areas (Europe) MTBS (USA)

Benchmarks: FireScope-Bench

Metrics: Brier score ROC AUC ECE IoU MSE MAE SSIM QWK fidelity consistency

Baselines: AlphaEarth SegFormer U-Net Climate MLP Fire Weather Index GPT-5 Qwen2.5-VL-7B (no CoT) Qwen+decoder

Builds on: GRPO FiLM Qwen2.5-VL SegFormer U-Net

Closest papers by meaning

• DetectiumFire: A Comprehensive Multi-modal Dataset Bridging Vision and Language for Fire Understanding (NeurIPS 2025 · similarity 0.85)
• Sim2Real-Fire: A Multi-modal Simulation Dataset for Forecast and Backtracking of Real-world Forest Fire (NeurIPS 2024 · similarity 0.85)
• CHOICE: Benchmarking the Remote Sensing Capabilities of Large Vision-Language Models (NeurIPS 2025 · similarity 0.85)
• GST-UNet: A Neural Framework for Spatiotemporal Causal Inference with Time-Varying Confounding (NeurIPS 2025 · similarity 0.85)
• TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation (CVPR 2026 · similarity 0.85)
• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.84)