FireScope: Wildfire Risk Raster Prediction with a Chain-of-Thought Oracle
Pith reviewed 2026-05-17 20:41 UTC · model grok-4.3
The pith
A vision-language model with chain-of-thought reasoning predicts wildfire risk rasters that transfer from US training to European testing.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
FireScope demonstrates that integrating language-based chain-of-thought reasoning into a vision-language model for raster generation allows the model to learn transferable causal reasoning about wildfire risk factors, leading to better performance when moving from expert-defined supervision in the United States to real-world fire events in Europe.
What carries the argument
The chain-of-thought oracle in the VLM that generates complementary reasoning traces during risk raster prediction, enabling integration of visual, climatic, and geographic factors.
If this is right
- Reasoning traces improve both the accuracy and interpretability of generated risk rasters.
- Models can generalize across continents when trained on aligned supervision signals from expert rasters and real events.
- Similar reasoning-to-generation frameworks could apply to other continuous spatial prediction problems.
- High-resolution risk models become feasible for cross-continental use without region-specific retraining.
Where Pith is reading between the lines
- Extending this to real-time data streams could enable dynamic wildfire risk forecasting during fire seasons.
- The method might help in other domains like urban planning or agricultural risk assessment where spatial reasoning is key.
- Validating the reasoning traces further could lead to hybrid human-AI systems for environmental monitoring.
Load-bearing premise
Expert-defined risk rasters in the US and real wildfire events in Europe provide sufficiently aligned supervision for learning transferable causal reasoning across continents.
What would settle it
Observing no performance improvement or unfaithful reasoning traces when the model is applied to wildfire data from a third continent like Australia would falsify the claim of robust cross-continental generalization.
Figures
read the original 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.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript introduces FireScope-Bench, a large-scale dataset pairing Sentinel-2 imagery and climate data with expert-defined risk rasters across the USA and real wildfire events in Europe for cross-continental evaluation. It proposes FireScope, a VLM-based reasoning-to-generation framework trained via reinforcement learning and visual supervision to output risk rasters together with chain-of-thought reasoning traces. The central claims are that US-trained models achieve substantial performance gains on European test data, that the reasoning traces are faithful and semantically meaningful per expert feedback and automated analysis, and that the work is the first to demonstrate language-based reasoning improving generalization in visual generation, a high-resolution cross-continental wildfire model, and systematic studies of robust multimodal generalization.
Significance. If the performance and faithfulness claims hold after addressing alignment and reporting gaps, the work would contribute a useful benchmark and an interpretable reasoning-driven approach to geospatial raster prediction. The planned public release of FireScope-Bench and code would support reproducibility and further studies in cross-domain spatial modeling.
major comments (2)
- [§3.2] §3.2 (Dataset Construction): the cross-continental supervision setup assumes US expert-defined risk rasters and European real-event maps supply aligned signals for transferable causal reasoning, yet no quantitative alignment analysis (mutual information, covariate-shift statistics, or raster-event overlap metrics) is reported. This is load-bearing for the generalization claim, as differing label semantics, spatial density, and correlation structures with Sentinel-2/climate covariates could produce domain-adaptation artifacts rather than genuine reasoning transfer.
- [§5] §5 (Experiments): the abstract asserts 'substantial performance gains' and 'faithful reasoning traces' but the visible description supplies no specific quantitative metrics, ablation results, or error analysis relative to baselines. This leaves the central empirical claim without sufficient visible support and requires detailed tables or figures showing effect sizes and controls.
minor comments (1)
- [Abstract] Abstract: the triple 'first framework' claim would be strengthened by a concise literature comparison rather than a broad assertion.
Simulated Author's Rebuttal
We thank the referee for the detailed and constructive feedback on our manuscript. We address each of the major comments below and outline the revisions we plan to make to strengthen the paper.
read point-by-point responses
-
Referee: [§3.2] §3.2 (Dataset Construction): the cross-continental supervision setup assumes US expert-defined risk rasters and European real-event maps supply aligned signals for transferable causal reasoning, yet no quantitative alignment analysis (mutual information, covariate-shift statistics, or raster-event overlap metrics) is reported. This is load-bearing for the generalization claim, as differing label semantics, spatial density, and correlation structures with Sentinel-2/climate covariates could produce domain-adaptation artifacts rather than genuine reasoning transfer.
Authors: We agree that providing quantitative evidence of alignment between the US expert-defined risk rasters and the European real-event maps is crucial to support our generalization claims. In the revised manuscript, we will add a new subsection in §3.2 that includes mutual information analysis between the label distributions, covariate shift statistics (e.g., using maximum mean discrepancy or KL divergence on the Sentinel-2 and climate feature distributions), and metrics for raster-event overlap. This analysis will help demonstrate that the performance gains arise from transferable causal reasoning rather than mere domain adaptation effects. revision: yes
-
Referee: [§5] §5 (Experiments): the abstract asserts 'substantial performance gains' and 'faithful reasoning traces' but the visible description supplies no specific quantitative metrics, ablation results, or error analysis relative to baselines. This leaves the central empirical claim without sufficient visible support and requires detailed tables or figures showing effect sizes and controls.
Authors: We acknowledge that the experiments section would benefit from more explicit quantitative details to make the claims fully supported. We will expand §5 with a comprehensive table reporting specific metrics such as IoU, F1-score, and RMSE for FireScope against relevant baselines on the European test set. Additionally, we will include ablation studies isolating the contribution of the chain-of-thought reasoning and visual supervision components, along with an error analysis highlighting failure cases and their relation to the reasoning traces. These additions will provide the necessary effect sizes and controls. revision: yes
Circularity Check
No circularity: empirical training and evaluation framework
full rationale
The paper introduces FireScope-Bench and a VLM-based framework trained via reinforcement learning plus visual supervision to predict wildfire risk rasters. Central claims rest on reported performance gains when training on US expert rasters and testing on European wildfire events, plus qualitative confirmation of reasoning traces. No mathematical derivation chain, first-principles result, or fitted parameter is presented that reduces by construction to its own inputs. The work is self-contained as standard empirical ML with cross-continental evaluation and does not rely on load-bearing self-citations or ansatzes that collapse into the target claim.
Axiom & Free-Parameter Ledger
free parameters (1)
- RL reward scaling and supervision loss weights
axioms (1)
- domain assumption Vision-language models can integrate visual, climatic, and geographic cues into faithful causal reasoning for spatial risk tasks.
invented entities (1)
-
Chain-of-Thought Oracle
no independent evidence
Lean theorems connected to this paper
-
IndisputableMonolith/Foundation/RealityFromDistinction.leanreality_from_one_distinction unclear?
unclearRelation between the paper passage and the cited Recognition theorem.
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
-
IndisputableMonolith/Cost/FunctionalEquation.leanwashburn_uniqueness_aczel unclear?
unclearRelation between the paper passage and the cited Recognition theorem.
Reward Design. Following common practice we define our reward as the weighted sum of two components: R = 0.9·R_acc + 0.1·R_fmt
What do these tags mean?
- matches
- The paper's claim is directly supported by a theorem in the formal canon.
- supports
- The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
- extends
- The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
- uses
- The paper appears to rely on the theorem as machinery.
- contradicts
- The paper's claim conflicts with a theorem or certificate in the canon.
- unclear
- Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.
Reference graph
Works this paper leans on
-
[1]
Matej Arlovic, Franko Hrzic, Mitesh Patel, Tomasz Bednarz, and Josip Balen. Evaluation of synthetic data impact on fire segmentation models performance.Scientific Reports, 15(1): 16759, 2025. 1
work page 2025
-
[2]
Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. SegNet: A deep convolutional encoder-decoder architecture for image segmentation.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence (TPAMI), 39(12):2481–2495,
-
[3]
Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao- hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-VL technical report....
work page internal anchor Pith review Pith/arXiv arXiv
-
[4]
eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers
Yogesh Balaji and et al. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers. Inpreprint arXiv:2211.01324, 2022. 3
work page internal anchor Pith review Pith/arXiv arXiv 2022
-
[5]
SatlasPretrain: A large-scale dataset for remote sensing image understanding
Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Fer- dinando, and Aniruddha Kembhavi. SatlasPretrain: A large-scale dataset for remote sensing image understanding. preprint arXiv:2211.15660, 2023. 3
-
[6]
Recognition in terra incognita
Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. InEuropean Conference on Computer Vision (ECCV), 2018. 1, 2, 3
work page 2018
-
[7]
Stephanie Bohlmann and Marko Laine. Statistical calibra- tion of probabilistic medium-range fire weather index fore- casts in europe.Natural Hazards and Earth System Sciences, 24:4225–4235, 2024. 1, 3
work page 2024
-
[8]
G. W. Brier. Verification of forecasts expressed in terms of probability.Monthly Weather Review, 78(1):1–3, 1950. 4, 12
work page 1950
-
[9]
Christopher F. Brown, Michal R. Kazmierski, Valerie J. Pasquarella, William J. Rucklidge, Masha Samsikova, Chen- hui Zhang, Evan Shelhamer, Estefania Lahera, Olivia Wiles, Simon Ilyushchenko, Noel Gorelick, Lihui Lydia Zhang, Sophia Alj, Emily Schechter, Sean Askay, Oliver Guinan, Rebecca Moore, Alexis Boukouvalas, and Pushmeet Kohli. AlphaEarth foundatio...
work page internal anchor Pith review arXiv 2025
-
[10]
Jeremy Buch, Erich Fischer, Jorge Pe˜na, et al. SMLFire1.0: a stochastic machine learning model for fire frequency and size distributions across the western united states.Geoscientific Model Development, 16:3407–3432, 2023. 1, 3
work page 2023
-
[11]
R2I- Bench: Benchmarking reasoning-driven text-to-image gen- eration.preprint arXiv:2505.23493, 2025
Kaijie Chen, Zihao Lin, Zhiyang Xu, Ying Shen, Yuguang Yao, Joy Rimchala, Jiaxin Zhang, and Lifu Huang. R2I- Bench: Benchmarking reasoning-driven text-to-image gen- eration.preprint arXiv:2505.23493, 2025. 3
-
[12]
Encoder-decoder with atrous separable convolution for semantic image segmentation
Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In European Conference on Computer Vision (ECCV), 2018. 3
work page 2018
-
[13]
Jacob Cohen. Weighted kappa: Nominal scale agreement with provision for scaled disagreement or partial credit.Psy- chological Bulletin, 70(4):213–220, 1968. 4, 6, 12
work page 1968
-
[14]
Copernicus. EFFIS burnt areas (by MODIS) was accessed on 24.10.2025 from https://forest-fire.emergency.copernicus.eu, . Accessed 24.10.2025. 3
work page 2025
-
[15]
Sentinel-2 was accessed on 24.10.2025 from https://registry.opendata.aws/sentinel-2,
Copernicus. Sentinel-2 was accessed on 24.10.2025 from https://registry.opendata.aws/sentinel-2, . Accessed 24.10.2025. 4
work page 2025
-
[16]
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
DeepSeek-AI. DeepSeek-R1: Incentivizing reasoning ca- pability in LLMs via reinforcement learning.preprint arXiv:2501.12948, 2025. 5
work page internal anchor Pith review Pith/arXiv arXiv 2025
-
[17]
Global data-driven prediction of fire ac- tivity.Nature Communications, 16(1):58097, 2025
Francesca Di Giuseppe, Joe McNorton, Anna Lombardi, and Fredrik Wetterhall. Global data-driven prediction of fire ac- tivity.Nature Communications, 16(1):58097, 2025. 1, 3
work page 2025
-
[18]
Tam- ing transformers for high-resolution image synthesis
Patrick Esser, Robin Rombach, and Bj ¨orn Ommer. Tam- ing transformers for high-resolution image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3
work page 2021
-
[19]
T. Fawcett. An introduction to ROC analysis.Pattern Recog- nition Letters, 27(8):861–874, 2006. 4, 12
work page 2006
-
[20]
Soft Actor-Critic Algorithms and Applications
Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, and Sergey Levine. Soft actor-critic algorithms and applications.preprint arXiv:1812.05905, 2019. 5, 13
work page internal anchor Pith review Pith/arXiv arXiv 2019
-
[21]
On misconceptions about the brier score in binary prediction models.preprint arXiv:2504.04906v4,
Linard Hoessly. On misconceptions about the brier score in binary prediction models.preprint arXiv:2504.04906v4,
-
[22]
P. Jaccard. The distribution of the flora in the alpine zone. New Phytologist, 11(2):37–50, 1912. 4, 12
work page 1912
-
[23]
Perceiver IO: A General Architecture for Structured Inputs & Outputs
Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Kop- pula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier H´enaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, and Jo ¯ao Carreira. Perceiver IO: A general architecture for structured inputs & outputs.preprint arXiv:2107.14795, 2022. 6
work page internal anchor Pith review Pith/arXiv arXiv 2022
-
[24]
Instruction reasoning dataset for ad- vanced image editing.preprint arXiv:2405.11190, 2024
Ying Jin, Pengyang Ling, Xiaoyi Dong, Pan Zhang, Jiaqi Wang, and Dahua Lin. Instruction reasoning dataset for ad- vanced image editing.preprint arXiv:2405.11190, 2024. 3
-
[25]
Evaluating numerical reasoning in text-to- image models.preprint arXiv:2406.14774, 2024
Ivan Kaji ´c et al. Evaluating numerical reasoning in text-to- image models.preprint arXiv:2406.14774, 2024. 3
-
[26]
Wilds: A benchmark of in-the-wild distribution shifts
Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akhil Balsubramani, 9 Weihua Hu, Michihiro Yasunaga, Percy Liang, Yair Carmon, et al. Wilds: A benchmark of in-the-wild distribution shifts. InInternational Conference on Machine Learning (ICML),
-
[27]
Large Language Models are Zero-Shot Reasoners
Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners.preprint arXiv:2205.11916, 2022. 3
work page internal anchor Pith review Pith/arXiv arXiv 2022
-
[28]
Spyros Kondylatos, Ioannis Prapas, Michele Ronco, Ioannis Papoutsis, Gustau Camps-Valls, Mar´ıa Piles, Miguel-´Angel Fern´andez-Torres, and Nuno Carvalhais. Wildfire danger prediction and understanding with deep learning.Geophysi- cal Research Letters, 49(17):e2022GL099368, 2022. 3
work page 2022
-
[29]
Uncertainty-aware deep learning for wildfire danger forecasting.preprint arXiv:2509.25017, 2025
Spyros Kondylatos, Gustau Camps-Valls, and Ioannis Pa- poutsis. Uncertainty-aware deep learning for wildfire danger forecasting.preprint arXiv:2509.25017, 2025. 1
-
[30]
Measuring Faithfulness in Chain-of-Thought Reasoning
Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, Kamil ˙e Lukoˇsi¯ut˙e, Karina Nguyen, Newton Cheng, Nicholas Joseph, Nicholas Schiefer, Oliver Rausch, Robin Larson, Sam McCandlish, Sandipan Kundu, Saurav Kadavath, Shan- non Yang, Thomas Henighan, Timothy...
work page internal anchor Pith review Pith/arXiv arXiv 2023
-
[31]
Evaluating text-to-visual generation with image-to-text generation
Ziqiu Lin et al. Evaluating text-to-visual generation with image-to-text models.preprint arXiv:2404.01291, 2024. 3
-
[32]
MM-REACT: Prompting ChatGPT for Multimodal Reasoning and Action
Haotian Liu, Chunyuan Li, Pengchuan Zhang, and Yong Jae Lee. MM-ReAct: Prompting ChatGPT for multimodal rea- soning and action.preprint arXiv:2303.11381, 2023. 3
work page internal anchor Pith review Pith/arXiv arXiv 2023
-
[33]
Jia Liu, Yukuan Wang, Yafeng Lu, Pengguo Zhao, Shunjiu Wang, Yu Sun, and Yu Luo. Application of remote sensing and explainable artificial intelligence for wildfire risk zon- ing in the mountainous region of Southwest China.Remote Sensing, 16(19):3602, 2024. 3
work page 2024
-
[34]
Fully convolutional networks for semantic segmentation
Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3431–3440, 2015. 3
work page 2015
-
[35]
Decoupled Weight Decay Regularization
Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.preprint arXiv:1711.05101, 2017. 14
work page internal anchor Pith review Pith/arXiv arXiv 2017
-
[36]
SGDR: Stochastic gradi- ent descent with warm restarts
Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradi- ent descent with warm restarts. InInternational Conference on Learning Representations (ICLR), 2017. 14
work page 2017
-
[37]
A global probability-of-fire (PoF) forecast.Geophysical Research Letters, 51:e2023GL107929, 2024
Joe Ramu McNorton, Francesca Di Giuseppe, Ewan Mark Pinnington, Matthew Chantry, and Chris Barnard. A global probability-of-fire (PoF) forecast.Geophysical Research Letters, 51:e2023GL107929, 2024. 1, 3
work page 2024
-
[38]
PhyBench: A physical com- monsense benchmark for evaluating text-to-image models
Fanqing Meng, Wenqi Shao, Lixin Luo, Yahong Wang, Yi- ran Chen, Quanfeng Lu, Yue Yang, Tianshuo Yang, Kaipeng Zhang, Yu Qiao, and Ping Luo. PhyBench: A physical com- monsense benchmark for evaluating text-to-image models. preprint arXiv:2406.11802, 2024. 3
-
[39]
M. P. Naeini, G. F. Cooper, and M. Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. InAAAI Conference on Artificial Intelligence, 2015. 4, 12
work page 2015
-
[40]
NASA. Data obtained from national aeronautics and space administration (NASA) Langley Research Center’s predic- tion of worldwide energy resources (POWER), NASA Earth Science Division, . Accessed 24.10.2025. 4
work page 2025
-
[41]
Data obtained from the POWER project’s climatol- ogy,
NASA. Data obtained from the POWER project’s climatol- ogy, . Accessed 24.10.2025. 4
work page 2025
- [42]
-
[43]
Marc-Andr ´e Parisien and Max A. Moritz. Environmental controls on the distribution of wildfire at multiple spatial scales.Ecological Monographs, 79(1):127–154, 2009. 4, 12
work page 2009
-
[44]
FiLM: Visual reasoning with a general conditioning layer
Ethan Perez, Florian Strub, Harm de Vries, Vincent Du- moulin, and Aaron Courville. FiLM: Visual reasoning with a general conditioning layer. InAAAI Conference on Artificial Intelligence, 2018. 2, 5
work page 2018
-
[45]
High-resolution image syn- thesis with latent diffusion models
Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3
work page 2022
-
[46]
U- Net: Convolutional networks for biomedical image segmen- tation
Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- Net: Convolutional networks for biomedical image segmen- tation. InMedical Image Computing and Computer-Assisted Intervention (MICCAI), 2015. 3, 5, 13
work page 2015
-
[47]
J. San-Miguel-Ayanz, Ernst Schulte, Guido Schmuck, An- drea Camia, Peter Strobl, Giorgio Libert `a, Cristiano Gio- vando, Roberto Boca, Fernando Sedano, Pieter Kempeneers, Daniel McInerney, Ceri Withmore, Sandra Oliveira, Mar- cos Rodrigues, Tracy Durrant, Paolo Corti, Friderike Oehler, Lara Vilar, and Giuseppe Amatulli. Comprehensive monitor- ing of wild...
work page 2012
-
[48]
Proximal Policy Optimization Algorithms
John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Rad- ford, and Oleg Klimov. Proximal policy optimization algo- rithms.preprint arXiv:1707.06347, 2017. 5, 13
work page internal anchor Pith review Pith/arXiv arXiv 2017
-
[49]
A. Sengupta et al. Recent advances in explainable machine learning models for wildfires: From forecasting to burned area estimation.Environmental Data Science, 2025. In press. 3
work page 2025
-
[50]
USDA Forest Service. Wildfire risk to communities. https://wildfirerisk.org. Accessed 24.10.2025. 3, 7
work page 2025
-
[51]
Dmitrii Shadrin, Svetlana Illarionova, Fedor Gubanov, Kse- nia Evteeva, Maksim Mironenko, Ivan Levchunets, Roman Belousov, and Evgeny Burnaev. Wildfire spreading pre- diction using multimodal data and deep neural network ap- proach.Scientific Reports, 14:2606, 2024. 1, 3
work page 2024
-
[52]
DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
Zhihang Shao, Ziyu Wang, Yuxin Zhang, Zihan Zheng, Yao Liu, Zihan Liu, Yibo Shang, Linyang Xu, Tianyang Zhang, Lingpeng Chen, et al. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models.preprint arXiv:2402.03300, 2024. 5, 13
work page internal anchor Pith review Pith/arXiv arXiv 2024
-
[53]
ViperGPT: Visual inference via python execution for reasoning
D ´avid Sur´ıs, Sachit Menon, and Carl V ondrick. ViperGPT: Visual inference via python execution for reasoning. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3
work page 2023
-
[54]
C. E. Van Wagner. Development and structure of the cana- dian forest fire weather index system. Technical Report 10 Forestry Technical Report 35, Canadian Forestry Service, Petawawa National Forestry Institute, Chalk River, Ontario,
-
[55]
Self-Consistency Improves Chain of Thought Reasoning in Language Models
Xuezhi Wang, Jason Wei, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Self-consistency improves chain-of-thought reasoning in language models. preprint arXiv:2203.11171, 2022. 3
work page internal anchor Pith review Pith/arXiv arXiv 2022
-
[56]
Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4): 600–612, 2004. 4, 5, 12
work page 2004
-
[57]
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models.preprint arXiv:2201.11903, 2022. 3
work page internal anchor Pith review Pith/arXiv arXiv 2022
-
[58]
Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transform- ers. InAdvances in Neural Information Processing Systems (NeurIPS), 2021. 5, 13
work page 2021
-
[59]
Zhengsen Xu, Jonathan Li, Sibo Cheng, Xue Rui, Yu Zhao, Hongjie Heand Haiyan Guan, Aryan Sharma, Matthew Erxleben, Ryan Chang, and Linlin Xu. Deep learning for wildfire risk prediction: Integrating remote sensing and en- vironmental data.ISPRS Journal of Photogrammetry and Remote Sensing, 2025. Early access. 3
work page 2025
-
[60]
Mmmu: A massive multi-discipline multimodal understand- ing and reasoning benchmark for expert agi
Xiang Yue, Yuansheng Ni, Tianyu Zheng, Kai Zhang, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren- liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understand- ing and reasoning benchmark for...
work page 2024
-
[61]
The Dawn of LMMs: Preliminary Explorations with GPT-4V(ision)
Renrui Zhang, Zheng Li, Hongyang Li, Yu Qiao, and Peng Gao. Visual chain-of-thought reasoning for multimodal large language models.preprint arXiv:2309.17421, 2023. 3
work page internal anchor Pith review Pith/arXiv arXiv 2023
-
[62]
Xinshen Zhang, Zhen Ye, and Xu Zheng. Towards omnidi- rectional reasoning with 360-r1: A dataset, benchmark, and GRPO-based method.preprint arXiv:2505.14197, 2025. 5
-
[63]
Weixiang Zhao, Xingyu Sui, Jiahe Guo, Yulin Hu, Yang Deng, Yanyan Zhao, Bing Qin, Wanxiang Che, Tat-Seng Chua, and Ting Liu. Trade-offs in large reasoning models: An empirical analysis of deliberative and adaptive reasoning over foundational capabilities.preprint arXiv:2503.17979,
-
[64]
Miguel ´Angel Torres-V ´azquez, Sixto Herrera, And- rina Gincheva, Amar Halifa-Mar ´ın, Leone Cavicchia, Francesca Di Giuseppe, Juan Pedro Mont ´avez, and Marco Turco. Enhancing seasonal fire predictions with hybrid dy- namical and random forest models.Natural Hazards, 2,
-
[65]
1, 3 11 FireScope: Wildfire Risk Prediction with a Chain-of-Thought Oracle Supplementary Material
-
[66]
Detailed Metrics In-distribution (ID).As we have ground truth continuous risk rasters in the US, we use three metrics for evaluation: Mean Squared Error (MSE)to quantify per-pixel predic- tion error: MSE= 1 N NX i (xi −y i)2 (5) . Mean Absolute Error (MAE)to quantify per-pixel pre- diction error: MAE= 1 N NX i |xi −y i|(6) Structural Similarity Index (SSI...
-
[67]
Experiments Configurations 10.1. Models Oracles. We select Qwen2.5-VL-7B-Instruct [3] as our Or- acle for its strong open-source performance across multi- modal reasoning benchmarks and efficiency to train and de- ploy at only 7B parameters. We train two versions of it: one with CoT reasoning as outlined in Section 10.2, and one trained with supervised fi...
work page 2016
-
[68]
Ablation results Metrics of ablation methods are reported in Table 5 Table 5. Metrics of ablation methods. Method OOD wildfire eventsOOD wildfire pixelsID wildfire risk rastersConditioning EncoderBrier↓ROC AUC↑ECE↓ROC AUC↑IoU@0.5↑MSE↓SSIM↑MAE↓n/a VLM-enc 0.200 0.738 0.0690.647 0.176 0.050 0.468 0.182Image only Unet*0.208 0.699 0.0580.619 0.173 0.019 0.620...
-
[69]
General-Purpose Capabilities of CoT Ora- cle (FireScope) After prompting CoT Oracle to provide its reasoning and overall risk score, it may be of interest to users to acquire answers to follow-up questions or continue conversing. For this reason, we perform simple quantitative and qualitative analyses of CoT Oracle’s general purpose capabilities. In this ...
-
[70]
Dense vegetation can increase fire risk due to more fuel available for fires to spread
**Vegetation Density**: The satellite image shows areas with 14 varying shades of green, indicating different levels of vegetation density. Dense vegetation can increase fire risk due to more fuel available for fires to spread
-
[71]
**Climate Data**: - **Average Temperature**: Lower temperatures generally reduce fire risk because fires require higher temperatures to ignite and sustain. However, the average temperatures in the image region are relatively low, suggesting that fire risk might be moderate but not extremely high. - **Average Humidity**: Higher humidity reduces fire risk b...
-
[72]
Slopes can increase fire risk as they allow fires to spread more easily uphill
**Slope**: The satellite image shows areas with varying eleva- tions, which could indicate slopes. Slopes can increase fire risk as they allow fires to spread more easily uphill. Analysis: - **Vegetation Density**: The image shows a mix of dense and less dense vegetation. Dense vegetation increases fire risk. - **Climate Data**: - **Temperature**: Low tem...
-
[73]
**Avoid Areas Near Vegetation**: Stay away from areas with dense vegetation, as these are more likely to catch fire and spread quickly
-
[74]
**Consider Elevation**: Higher elevation areas may have better air circulation, which can help dissipate heat and reduce the risk of fire
-
[75]
**Avoid Low-Lying Areas**: Low-lying areas can accumulate heat and moisture, making them more susceptible to fire. Recommended Area: Based on the satellite image, a good recommendation would be to build the house in the upper part of the image, specifically in the lighter green areas that appear to be less dense with vegetation. These areas are likely to ...
-
[76]
This can make the area more comfortable year-round
**Better Air Circulation**: Higher elevation areas typically ex- perience better air circulation, which can help cool the surrounding environment and reduce the risk of heat accumulation. This can make the area more comfortable year-round
-
[77]
Additionally, water can be used for firefighting purposes if needed
**Proximity to Water**: The presence of a body of water nearby can provide a natural buffer zone, reducing the risk of direct fire threats. Additionally, water can be used for firefighting purposes if needed
-
[78]
**Scenic Views**: Higher elevation areas often offer panoramic views of the landscape, which can be visually appealing and provide a sense of tranquility
-
[79]
**Reduced Noise Pollution**: Higher elevation areas are usu- 15 ally quieter than lower-lying areas, as they are farther from urban centers and industrial noise sources
-
[80]
**Potential for Wildlife**: Depending on the local ecosystem, higher elevation areas might attract more wildlife, offering opportunities for nature observation and enjoyment. Disadvantages:
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.