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

AI summary

Overview

Research area: Computer vision and multimodal machine learning applied to geospatial modeling — specifically, predicting continuous wildfire risk rasters from satellite imagery and climate data, with a vision-language model (VLM) providing chain-of-thought (CoT) reasoning.

Technical level: Advanced. The paper assumes familiarity with vision-language models, reinforcement learning from group-relative policy optimization (GRPO), feature-wise linear modulation (FiLM), encoder-decoder segmentation architectures, and geospatial raster metrics.

Scope: The paper introduces FireScope-Bench, a multimodal dataset and benchmark for wildfire risk raster prediction spanning the USA and Europe, and FireScope, a two-stage "reasoning-to-generation" framework that conditions a vision encoder-decoder on scalar risk judgments produced by a CoT-trained VLM, yielding large out-of-distribution gains without sacrificing in-distribution accuracy.

What This Paper Is About

Wildfire risk estimation is not a simple image task: it requires combining visual appearance with climatic and geographic causal drivers to infer a continuous, spatially structured risk field. The paper argues that existing work addresses only observational tasks (fire detection, segmentation, spread estimation) or relies on coarse physics-based indices like the Fire Weather Index, and that no unified benchmark integrates imagery, climate, and topography, nor any framework spans the range from pixel-level vision to multimodal reasoning.

The goal is therefore twofold: build a large-scale benchmark that tests reasoning-intensive wildfire risk estimation and cross-continental generalization, and build a model that couples language-based reasoning with raster generation so that explicit reasoning grounds the predicted risk map.

Key Contributions

1. FireScope-Bench, a large-scale multimodal dataset and benchmark that couples Sentinel-2 imagery and climate data with expert-defined risk rasters across the USA, plus real wildfire events in Europe for cross-continental evaluation. It spans over 5.7M km² across 55K regions and 6.3B pixels, with a USA partition of 50K areas (2021) and a European partition of 4,989 areas (2018–2025), and includes metrics for accuracy, calibration, and interpretability.

2. FireScope, a reasoning-to-generation framework in which a VLM fine-tuned with GRPO produces a scalar wildfire risk estimate ("Oracle") and then conditions a lightweight vision encoder-decoder through FiLM to generate fine-grained risk rasters, combining interpretability of language reasoning with spatial precision of vision decoding.

3. Empirical evidence that explicit reasoning improves cross-continental generalization. When trained in the USA and tested in Europe, FireScope achieves substantial out-of-distribution (OOD) gains while remaining robust in-distribution, and these gains persist even against models trained on forty times more data.

4. A systematic interpretability evaluation, combining an expert study with automated fidelity and consistency metrics, showing that FireScope's reasoning traces are semantically meaningful and that raster generation is actively and coherently guided by them.

The paper explicitly frames items 1 and 2 as its two key contributions, and additionally claims three "firsts" for the framework: the first to demonstrate that language-based reasoning can improve generalization in visual generation, the first high-resolution wildfire risk model applicable across continents, and the first to enable systematic studies of robust cross-continental generalization for multimodal fire risk models.

Main Findings

• CoT conditioning produces the best OOD event discrimination across every vision backbone. In Table 1, conditionings are compared as image-only, +climate data, +Oracle (Qwen fine-tuned without CoT), and +CoT (FireScope). With U-Net, CoT conditioning yields Brier 0.191, ROC AUC 0.750, and ECE 0.068; the image-only U-Net baseline scores Brier 0.217, ROC AUC 0.679, ECE 0.050. With SegFormer, CoT conditioning yields Brier 0.205 and ROC AUC 0.727 versus 0.222 and 0.705 for image-only. With AlphaEarth, CoT yields Brier 0.218 and ROC AUC 0.717 versus 0.252 and 0.692 for image-only.

• OOD gains extend to the pixel level. For OOD wildfire pixels, CoT Oracle conditioning obtains the best ROC AUC and IoU scores. CoT SegFormer reaches ROC AUC 0.658 and IoU@0.5 0.184; CoT U-Net reaches 0.652 and 0.178; CoT AlphaEarth reaches 0.653 and 0.177. The paper notes this is surprising because the Oracle conditions on only a single scalar yet improves generated rasters at pixel scale.

• Reasoning improves the Oracle itself. In Table 2, CoT Qwen achieves OOD Brier 0.196, ROC AUC 0.748, ECE 0.077, compared with 0.225, 0.701, and 0.134 for non-CoT Qwen. Other Oracles perform worse on OOD discrimination: GPT-5 (Brier 0.281, ROC AUC 0.636, ECE 0.229), Climate MLP (0.276, 0.524, 0.150), and FWI (0.321, 0.551, 0.255).

• Climate-only models can match in-distribution but fail to transfer. The Climate MLP Oracle reaches an in-distribution QWK of 0.766, comparable only to CoT Qwen's 0.766, but its performance does not transfer OOD. Non-CoT Qwen reaches 0.751 QWK and GPT-5 only 0.316; in-distribution Brier/MAE are 0.029/0.123 for Climate MLP, 0.031/0.127 for CoT Qwen, 0.036/0.135 for Qwen, and 0.081/0.225 for GPT. FWI's in-distribution scores are listed as not reported ("–").

• In-distribution raster quality is largely preserved. In Table 3, climate-conditioned encoder-decoders tend to slightly outperform the rest; for example AlphaEarth with climate data reaches MSE 0.025, SSIM 0.552, MAE 0.110, versus 0.026, 0.544, 0.113 for image-only AlphaEarth. CoT conditioning is generally on par with other baselines rather than systematically better or worse (e.g., CoT AlphaEarth: MSE 0.028, SSIM 0.547, MAE 0.119).

• The OOD gain outweighs the ID loss quantitatively. Using USA MTBS events (2,284 real wildfire events and 1,522 control areas) as in-distribution and European EFFIS events as OOD, averaging across encoders gives an ID loss of 0.001 ROC AUC against an OOD gain of 0.037 for wildfire events, and an ID loss of 0.012 against an OOD gain of 0.026 for pixels.

• Ablations support the specific design, not merely the presence of a VLM. Training U-Net on the full training volume with translational augmentations — 40 times larger in total — improves in-distribution performance but remains inferior to FireScope's CoT-conditioned configuration OOD. Attaching a perceiver-style decoder directly to Qwen's last embedding layer ("Qwen+decoder") also performs worse, indicating the gains do not stem from the raw representational capacity of the VLM.

• Reasoning traces carry recoverable signal for experts. In the expert study on 50 sampled test areas, experts scored QWK 0.33 and 0.11 when classifying risk from Oracle's interpretations, compared with 0.50 and 0.59 from "golden" traces generated by GPT-5 given the correct classification. One expert thus reached nearly 70% of their golden-trace score; the other's 0.11 is above zero.

• Raster generation is faithfully and consistently guided by the CoT. Automated metrics report fidelity 0.33 and consistency 0.91 for Oracle. Near-perfect consistency means paraphrased CoTs barely change the generated raster, while fidelity 0.33 means pixel-wise predicted risk shifts by 33% toward the opposite scale on average when the CoT is perturbed.

• Error reduction grows with distance from the training distribution. Figure 7 reports that FireScope's improvement across Europe is proportional to the distance from the training data distribution, reaching its maximum at the 60th latitude and in the year 2018.

Methodology in Plain English

The authors first built a dataset. Wildfire risk targets come from Wildfire Risk to Communities, a probabilistic wildfire modeling project provided at 30 m per pixel across the United States; the target variable used is "Risk to Potential Structures," which combines wildfire likelihood with expected consequences and is normalized via a quantile transform into a [0,1] relative risk measure. These rasters are tiled into 341×341 pixel images of roughly 100 km² each, producing 40K training, 4K validation, and 4K testing samples across the mainland US and Hawaii, plus a smaller 1K/100/100 dataset for faster experimentation; large training tiles are grouped into 2×2 supertiles for runtime augmentation.

Predictors are deliberately restricted to raw, scalable sources: Sentinel-2 level 2A imagery at 10 m resolution (1024×1024 pixel tiles), mosaicked from cloud-free summer imagery between June 22 and September 22, plus monthly NASA POWER climatology at 50 km resolution covering near-surface temperature, precipitation, humidity, wind speed, and wind direction for each tile centroid, forming a 60-dimensional vector. Crucially, European wildfire events use imagery from the year preceding the event so the model must predict rather than detect fire. Evaluation uses 3K European wildfire events from the EFFIS burnt areas collection (2018–2025, events under five square kilometers filtered out), balanced by limiting events per country proportionally to area, plus 2K inland control locations with no detected wildfires. USA events from MTBS are included for completeness.

The framework has two stages. In stage one, an Oracle — Qwen2.5-VL-7B-Instruct — is fine-tuned with GRPO, a critic-free reinforcement learning algorithm, using a reward that is a weighted sum of accuracy and format terms: R = 0.9·R_acc + 0.1·R_fmt, with frequency-weighted aggregation to handle label imbalance. The authors deliberately do not steer the intermediate reasoning, letting it evolve naturally. In stage two, the Oracle's scalar outputs for the training set are generated deterministically and injected into a vision encoder-decoder through FiLM layers applied before each trainable block. The decoder regresses a normalized risk raster in [-1,1] at 341×341 resolution by minimizing a weighted sum of three terms: a Smooth-L1 reconstruction loss with β = 1.0, 0.5 times (1 − SSIM) with an 11×11 Gaussian window for structure, and 0.2 times an L1 loss on first-order finite differences for sharper edges. Encoders compared are SegFormer MiT-B5 and AlphaEarth (both frozen) and a lightweight U-Net trained from scratch. Ablations include a 40-times-larger-data U-Net and a Qwen+decoder variant. Interpretability is probed by generating default and "golden" CoTs on 50 test areas, summarizing them with Qwen/Qwen3-4B-Instruct-2507, asking two wildfire experts from the Wildfire Risk to Communities project to classify risk from anonymized interpretations, and by computing fidelity and consistency from synthetically perturbed or paraphrased CoTs.

Why This Matters

Impact on research. The paper reframes structured spatial prediction (raster generation) as a reasoning-to-generation problem rather than direct correlation learning, and provides a benchmark with an explicit cross-continental OOD split. Its central empirical claim — that a scalar, language-mediated signal from a reasoning model improves per-pixel generalization more than data scaling does — challenges the assumption that more pixels and more parameters are the primary route to robustness in geospatial vision, and offers a reusable evaluation protocol for reasoning in image generation rather than question answering.

Real-world applications:

• Operational wildfire risk mapping at 30 m resolution across regions and continents, producing continuous risk fields rather than coarse meteorological indices.
• Cross-border or cross-continental deployment, where a model trained in one continent (here, the USA) must be applied to another (here, Europe) without retraining — a setting where purely visual or correlation-based models often fail.
• Insurance, land-use planning, and infrastructure siting, where the target variable is explicitly tied to expected consequences for structures.
• Decision support with an audit trail: risk predictions come with human-readable reasoning that domain experts can inspect, critique, and partially reproduce.

Industry relevance. Wildfire risk is a direct input to property insurance underwriting, utility asset management and wildfire mitigation planning, forestry and utility operations, and public-sector emergency preparedness. A model that provides both high-resolution risk rasters and interpretable reasoning traces addresses two industry requirements at once — spatial precision and explainability for regulatory or expert review — and the benchmark's OOD design maps onto the real constraint that risk models must work in geographies where labeled historical data is scarce.

Future Directions

• Remove the scalar bottleneck. The paper identifies as its main limitation that communication between the encoder-decoder and the Oracle passes through a single scalar, which restricts the spatial granularity of the reasoning's influence. It proposes exploring token-level or region-aware embeddings, or multi-dimensional conditioning, so the Oracle can convey CoT-enhanced, spatially resolved information.
• Extend reasoning-guided geospatial modeling beyond wildfire. The framework's design — reasoning model as prior, lightweight vision decoder as generator — is presented as a general template for structured prediction, raising the question of which other continuous spatially aligned variables (for example, other hazard or environmental rasters) benefit from the same treatment.
• Explain and improve the mismatch between ID and OOD behavior. Climate-conditioned models perform strongly in-distribution but fail OOD, which the paper attributes to memorization of coarse regional climate signatures. Understanding and reducing this overfitting tension, as probed by the benchmark, is an open research question.
• Strengthen interpretability evaluation. The authors note that measuring how faithful a CoT is remains largely unsolved, that fidelity is expected to lag consistency because CoTs can only be perturbed so far without altering facts, and that expert usefulness of the traces appears somewhat subjective, unlike the golden traces.

Target Audience

Researchers and graduate students in computer vision, multimodal learning, and geospatial/Earth observation who are interested in reasoning-augmented generation, cross-domain generalization, or interpretable structured prediction. Also relevant to environmental scientists, wildfire risk modelers, and practitioners in insurance, utilities, and public safety who need high-resolution, explainable risk maps — particularly those working on models intended to transfer between continents. Readers without background in VLMs, reinforcement learning, or raster metrics will need to consult the cited literature, since the paper assumes fluency in all three.

Authors’ abstract

Predicting wildfire risk is a reasoning-intensive spatial problem that requires the integration of visual, climatic, and geographic factors to infer continuous risk maps. Existing methods lack the causal reasoning and multimodal understanding required for reliable generalization. We introduce FireScope-Bench, a large-scale dataset and benchmark that couples Sentinel-2 imagery and climate data with expert-defined risk rasters across the USA, and real wildfire events in Europe for cross-continental evaluation. Building on this dataset, we propose FireScope, a VLM-based reasoning-to-generation framework that learns from both reinforcement learning and visual supervision to predict risk rasters with complementary reasoning traces. When trained in the USA and tested in Europe, FireScope achieves substantial performance gains, while expert feedback and automated analysis confirm that its reasoning traces are faithful and semantically meaningful. Our findings demonstrate that reasoning can ground raster prediction models, improving both generalization and interpretability. To our knowledge, this is the first framework to (1) demonstrate that language-based reasoning can improve generalization in visual generation, (2) propose a high-resolution wildfire risk model that can be applied across continents, and (3) enable systematic studies of robust cross-continental generalization for multimodal fire risk models. We believe that FireScope-Bench has the potential to serve as a foundation for advancing reasoning-driven, interpretable and generalizable spatial modeling. Data and source code will be made publicly available.