PaperlayerView exact revision on arXiv

Source-linked AI summary

A Simulation-Grounded Agentic VLM Framework for Wildfire Monitoring and Reporting

Duowen Chen, Yuchen Sun, Zhiqi Li, Yuxuan Liao, Sinan Wang, Bart van Bloemen Waanders, Bo Zhu

arXiv:2610.02451v1cs.CVcs.GR

TL;DR

Real wildfire videos with synchronized physical annotations are scarce, and simulator outputs differ from aerial-view video. The paper converts simulations into labeled video memory for training-free multi-agent VLM reporting. Video memory achieves 51.5% exact four-tag accuracy versus 22.6% for direct VLM querying and 16–17% for text-only memory, while the full reporter reaches 77.3% on six simulator-derived fields.

  • Problem

    Real wildfire videos are visually rich but usually lack synchronized physical-state labels, while simulators expose physical conditions in outputs unlike the aerial-view videos VLMs must interpret.

    Source: §1 · p. 1

  • Method

    The framework automatically maps 2D wildfire simulations into proxy-conditioned labeled videos and uses them as multimodal memory for retrieval-augmented, training-free multi-agent VLM reporting.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    51.5% exact four-tag accuracy for video memory versus 22.6% for direct VLM querying and 16–17% for text-only memory; the full reporter reaches 77.3% on six simulator-derived fields.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Controlled, cross-generator, and real-UAV perceptual tests support using simulation experience for VLM wildfire monitoring under scarce physical annotations.

    Source: §8 · p. 9, §7 · p. 9

    Real-incident environmental-driver estimates still require validation against synchronized measurements; real-UAV tests assess smoke/fire detection and visible fire area, not physical-label targets.

    Source: §7 · p. 9

Abstract

from arXiv · show

Effective wildfire monitoring requires relating visual evidence to physical fire dynamics, yet real videos with synchronized physical annotations are scarce and high-fidelity 3D simulation is costly. We present a simulation-grounded vision-language model (VLM) framework that automatically converts 2D wildfire simulations into labeled video episodes. A fixed Blender mapping produces low-detail 3D proxies aligned with simulator terrain, fuel layout, fire activity, and wind cues; controllable video generation supplies richer appearance. The proxies are intermediate representations rather than finely rendered final scenes. Generated videos and simulator labels form reusable multimodal memory for a training-free multi-agent VLM system that retrieves reference episodes, reconciles visual and memory-based predictions, and produces structured wildfire reports. On held-out generated episodes, video memory achieves 51.5% exact four-tag accuracy, compared with 22.6% for direct VLM querying and 16-17% for text-only memory; the complete system achieves 77.3% accuracy on six simulator-derived report fields. Component ablations, cross-generator tests, and three real-UAV evaluations assess retrieval, reporting, generator changes, and observable monitoring tasks. The framework connects automatic simulation-to-proxy conversion with memory-based VLM reasoning under scarce real-world physical annotations.

1 INTRODUCTION

The paper addresses the scarcity of wildfire videos with synchronized physical labels and the mismatch between simulator outputs and aerial-view observations. It automatically converts simulations into labeled visual episodes and uses them as memory for training-free VLM reporting.

1 INTRODUCTION

  • Real wildfire videos rarely pair temporally rich visual evidence with physical-state labels, while simulators expose physical variables through outputs unlike aerial-view video.Synchronized annotations are hazardous, expensive, and difficult to control; full 3D fire simulation is costly across varied conditions.Source: §1 · p. 1
  • A fully automatic Blender mapping lifts 2D simulations over real terrain and fuel data into lightweight proxies that guide richer video generation.The fixed mapping encodes terrain, fuel layout, fire activity, and wind cues; the proxy is an intermediate representation, not a finely rendered final scene.Source: §1 · p. 2
  • Generated clips retain simulator parameters, environmental tags, and fire-behavior labels as reusable multimodal memory for reasoning about new videos without updating VLM weights.Retrieval complements direct visual reasoning in a system combining complementary reporters and selective disagreement adjudication.Source: §1 · p. 2
  • 51.5% exact four-tag accuracy from video memory exceeds 22.6% for direct VLM querying and 16–17% for text-only memory; full reporting reaches 77.3% on six simulator-derived fields.The evaluation also includes ablations, cross-generator tests, and real-UAV perceptual evaluations.Source: §1 · p. 2
  • The framework connects simulation-grounded visual episode construction with memory-augmented agentic reporting under scarce real-world physical annotations.Real-UAV evaluations assess smoke/fire detection and visible fire-area assessment.Source: §1 · p. 2

2 RELATED WORK

Related work spans wildfire simulation and sensing, wildfire-focused VLM applications, and retrieval-augmented multimodal reasoning. The paper positions its contribution as simulator-linked visual reference episodes for VLM monitoring and reporting.

2 RELATED WORK

  • Wildfire prediction combines spread models with satellite and UAV sensing; this work translates simulator states into visual reference episodes rather than replacing simulators or sensors.The cited examples include Rothermel-based spread, level-set methods, and coupled spread-atmosphere modeling.Source: §2 · p. 2
  • Wildfire VLM studies cover recognition, detection, UAV awareness, satellite risk reasoning, and planning, while scientific VLM benchmarks expose domain-specific challenges.These lines of work motivate physically grounded video data and task-specific evaluation.Source: §2 · p. 2
  • Prior multimodal memories retrieve visual, textual, or scene evidence, whereas this framework pairs reference videos with simulator-controlled conditions and fire-behavior labels for inference across episodes.The overall workflow combines simulation-grounded visual construction with memory-augmented multi-agent reporting.Source: §2 · p. 3, Figure 1 · p. 3

3 METHOD

The method turns simulator states into labeled aerial-view video episodes, then uses retrieved episodes alongside direct visual evidence to produce structured wildfire reports. Its stages span simulation, proxy construction, video generation, multimodal memory, and agentic reporting.

3 METHOD

  • The framework has two components: constructing simulation-grounded visual evidence and using a memory-augmented multi-agent VLM system for monitoring and reporting.The visual-evidence stage retains 2D simulation annotations while lifting diverse scenarios into procedural 3D proxies and aerial-view clips.Source: §3 · p. 3, §3.1 · p. 3

3.1 SIMULATION-GROUNDED VISUAL EVIDENCE CONSTRUCTION

  • The first stage links each visual episode to its originating scenario through simulator-derived labels and controls; cue alignment is evaluated separately.Simulation-grounded episodes are intended to reduce reliance on scarce real footage with synchronized physical annotations.Source: §3.1 · p. 3

3.1.1 PHYSICS-GROUNDED 2D WILDFIRE SIMULATION

  • SimFire episodes use real terrain and fuel data over approximately 3 km regions, with randomized wind and moisture settings and tags for four environmental factors.The factors are wind speed, moisture, slope angle, and vegetation density; each is discretized into low and high levels for retrieval and reasoning.Source: §3.1.1 · pp. 3–4

3.1.2 PROCEDURAL 3D PROXY

  • A procedural Blender proxy maps elevation, fuel categories, burning cells, and wind-conditioned smoke into a lightweight scene encoding simulator-derived spatial structure.The proxy is not intended to be photorealistic; it represents terrain, vegetation, active fire location, and smoke direction.Source: §3.1.2 · p. 4

3.1.3 CONTROLLABLE VIDEO GENERATION

  • Video generation uses depth, scribble, and text controls to guide geometry and appearance, while each generated clip retains its originating simulator labels.Prompts derive from simulation metadata and are refined by a VLM; parameter-conditioned negative prompts suppress conflicting attributes, and cue agreement is assessed rather than assumed.Source: §3.1.3 · p. 4

3.2 MEMORY-AUGMENTED AGENTIC VLM MONITORING AND REPORTING

  • The reporting stage retrieves visually and semantically similar episodes, uses their simulator labels as reference priors, and combines them with direct visual reasoning.New episodes update the memory bank rather than the VLM weights.Source: §3.2 · pp. 4–5, Figure 3 · p. 5

3.2.1 PHYSICS-LABELED MULTIMODAL MEMORY

  • Each multimodal memory entry pairs a video-and-description embedding key with the clip, prompt, simulation parameters, environmental tags, and fire-behavior labels.Cosine similarity retrieves top-k candidates; their labels inform query-tag estimates, with candidate reranking added because neither the agent nor embedder alone is sufficiently accurate across attributes.Source: §3.2.1 · p. 5

3.2.2 MEMORY-AUGMENTED MULTI-AGENT REPORTING

  • Complementary reporters use query frames, retrieved tag priors, and simulator summaries; agreement is copied into the report, while only disputed fields go to a judge.The resolved predictions are compiled into an HTML report recording visual evidence, labels, routing status, rationales, and memory summaries.Source: §3.2.2 · pp. 5–6, Figure 3 · p. 5

4 ABLATION STUDY

The ablations compare retrieval choices, memory modalities, and reporting agents; video memory and the full reporter outperform their stated alternatives.

Experimental setup

  • Figure 5 presents two generated video clips for each simulation tag.The ablation experiments use 400 memory videos and 100 held-out videos from 500 disjoint locations.Source: Figure 5 · p. 6, §4 · p. 6

Multimodal embedding model

  • Qwen3-2B-VL has the strongest tag-retrieval accuracy in most comparisons and is selected for the final memory bank.The alternatives are Cosmos-Embed1, Rzen-v2-7B, and Ops-MM-7B.Source: §4 · p. 7

Reranking retrieved candidates

  • GPT-5 reranking raises exact four-tag accuracy from 45.5% to 51.5%, while increasing k from 1 to 4 raises retrieval hit rate from 51.52% to 69.70%.The hit-rate result measures retrieval coverage, not final prediction accuracy.Source: §4 · p. 7

Importance of Video Memory

  • 51.5% exact four-tag accuracy with full video memory exceeds 16–17% for text-only variants, with stronger accuracy for every individual tag.The comparison supports stored video exemplars over prompt-only memory.Source: §4 · p. 7

Multi-Agent Effects on Downstream Task Accuracy

  • 77.3% accuracy on six simulator-derived fields is achieved by the full system, versus 72.2% for the combined-memory reporter and 72.6% for majority vote.The 82.3% oracle counts fields as correct when any reporter is correct; the ablations are not context- or compute-matched.Source: §4 · p. 7

5 VALIDATION AND RESULTS

The evaluation tests proxy–video alignment, environmental-driver prediction, memory growth, six-field reporting, and transfer across generators and to real UAV footage.

5 VALIDATION AND RESULTS

  • The validation evaluates proxy–video alignment, environmental-driver prediction, and six-field reporting on generated episodes with known simulator targets, then tests generator changes and real-UAV perception.Real-UAV tasks are smoke/fire detection and visible fire-area assessment.Source: §5 · p. 7

5.1 SIMULATION-TO-OBSERVATION RESULTS

  • Correct-pair SSIM and LPIPS win rates are 76.6% and 99.1%, while mean smoke-plume-axis and growth-front-axis errors are 7.6° and 35.8°.These results support partial visual-cue alignment, not preservation of physical spread rates or burned-area trajectories.Source: §5.1 · p. 7

5.2 MEMORY-BASED ENVIRONMENTAL DRIVER PREDICTION

  • Reranked memory reaches 51.5% exact four-tag accuracy, versus 22.6% for the strongest query-only VLM; with 16 labeled references, the strongest example-augmented baseline reaches 27.27%.Query simulator metadata is withheld.Source: §5.2 · p. 7, Table 3 · p. 8

5.3 MEMORY GROWTH EFFECTS

  • Increasing memory from 100 to 400 episodes raises exact four-tag accuracy from 26.3% to 51.5%.This adds simulator-grounded reference episodes without updating VLM weights.Source: §5.3 · p. 7, Figure 7 · p. 8

5.4 DOWNSTREAM FIRE-BEHAVIOR PREDICTION

  • The system leads on all six simulator-derived report fields when query simulator quantities are withheld, including Fast Spread Risk at 75.3% versus 55.7% for direct GPT.Reports combine predictions, visual evidence, retrieved summaries, tags, and rationales.Source: §5.4 · p. 8, Table 4 · p. 8

5.5 CROSS-GENERATOR AND REAL-UAV EVALUATION

  • The method achieves 41.00% for LTX queries against VACE memory and 44.44% in the reverse direction, versus direct-VLM baselines of 28.00% and 27.27%.The advantage persists across generators, although accuracy is below the within-generator result.Source: §5.5 · pp. 8–9, Table 5 · p. 8
  • Real-UAV tests cover smoke detection, binary fire presence, and visible fire-area assessment; the method has the highest balanced accuracy on FireMan-UAV-RGBT and NDCG on FLAME.It ties the best FLAME Hit@1 but not mean AP; several methods reach 100% on Boreal Forest Fire, and the annotations do not jointly specify all four environmental drivers.Source: Table 5 · p. 8, §5.5 · p. 9

6 DISCUSSION

The discussion distinguishes fixed simulator-to-proxy conversion from agent-authored scene work and positions the proxy as an aligned, low-overhead intermediate representation.

Relationship to Code World Model

  • The fixed converter maps prescribed wildfire trajectories to aligned aerial-view episodes, which become multimodal memory for interpreting query videos.Code World Model instead uses a coding agent to maintain persistent state and revise interaction mechanisms.Source: §6 · p. 9

Fixed conversion versus scene authoring

  • The proxy prioritizes simulator alignment and low rendering overhead, leaving appearance detail to video generation.Reuse avoids per-episode scene-authoring calls and token costs, but other inference costs remain; no speed or cost advantage is measured.Source: §6 · p. 9

7 LIMITATIONS AND FUTURE WORK

The framework's physical-driver claims are limited by simulation-derived targets and a domain-specific proxy; future work calls for real-footage calibration and broader validation.

Limitations

  • Real-incident environmental-driver estimates remain unvalidated against synchronized measurements; real-UAV tests assess perception, while physical-label targets come from simulation.The fixed Blender mapping may need extensions for new simulator fields or scene types.Source: §7 · p. 9
  • Generated videos may not preserve proxy dynamics, and the coarse simulator omits coupled fire–atmosphere dynamics, spotting, suppression, and changing weather.Footage may also omit sensor noise, camera motion, long-horizon evolution, or rare extremes.Source: §7 · p. 9

Future Work

  • Future work calls for calibrating generated observations against physically annotated real footage, expanding scenario coverage, and involving wildfire experts.The paper also proposes isolating prompt conditioning and comparing reporting with matched evidence and compute.Source: §7 · p. 9

8 CONCLUSION

The framework converts 2D wildfire states into simulation-aligned proxies and labeled episodes, then uses retrieved memory and complementary reporters for structured reporting without updating VLM weights.

8 CONCLUSION

  • Controlled, cross-generator, and real-UAV perceptual tests support using simulation experience under scarce physical annotations.Real-world physical calibration and expert validation remain important next steps.Source: §8 · p. 9
  • A fully automatic Blender conversion expresses 2D wildfire states as lightweight, simulation-aligned proxies for controlled video generation.The resulting labeled episodes supply multimodal memory.Source: §8 · p. 9
  • Retrieved references and complementary reporters support environmental-driver estimation and structured reporting without updating VLM weights.Source: §8 · p. 9

A DOWNSTREAM TASK AND PHYSICS LABELS

This appendix defines simulator-derived binary report fields and describes the fire-spread model and fuel-profile assets used to generate episodes.

A DOWNSTREAM TASK AND PHYSICS LABELS

  • Six binary report fields are defined from simulator-derived episode quantities, including maximum spread rate, peak and final burned area, and mean canopy cover.These episode-level targets are withheld from query inference and are not validated operational forecasts.Source: A DOWNSTREAM TASK AND PHYSICS LABELS · p. 15, Table 8 · p. 16

B.2 ANDERSON 13 FUEL PROFILES

B WILDFIRE MODELS

  • The appendix identifies the Rothermel surface fire spread model as SimFire’s backend, predicting directional spread from local fuel, moisture, wind, and terrain conditions.The model applies directional rates from an actively burning cell to candidate neighboring cells.Source: §B.1 · p. 15

B.1 ROTHERMEL SURFACE FIRE SPREAD MODEL

  • The directional spread calculation combines reaction intensity and propagating flux with wind and slope corrections, while fuel moisture affects heat of preignition.Wind is projected along the candidate spread direction; signed slope correction makes uphill spread faster and downhill spread slower.Source: §B.1 · pp. 15–17

C IMPLEMENTATION DETAILS

Implementation uses SimFire episodes over real terrain and fuel data, Blender proxies and controlled video generation, then retrieval and adjudication to produce structured reports.

2D Simulation

  • The 2D simulation uses SimFire on a 3 km domain with approximately 30 m cells and a one-hour propagation horizon.The queried physical locations are shown for memory construction and held-out evaluation in Figure 9.Source: C IMPLEMENTATION DETAILS · p. 19, Figure 9 · p. 19

3D Rendering

  • Blender converts LANDFIRE elevation into terrain surfaces, represents fuel types with billboards, and renders fire and wind-driven smoke from burning cells.Aerial viewpoints resemble UAV footage; rendering uses Eevee.Source: C IMPLEMENTATION DETAILS · p. 20

3D Generation

  • VACE generates videos from depth and scribble controls derived from Blender proxies, with low depth-alignment scores triggering retries and possible location resampling.The pipeline retries up to three times when the Spearman depth-correlation score is below 0.35.Source: C IMPLEMENTATION DETAILS · p. 20

Memory System

Multi-Agent Reporting

  • For six-field reporting, five retrieved episodes inform consensus and metadata summaries; unanimous reporter outputs are retained and only disputed fields go to the judge.Four-tag prediction instead uses the top one retrieved episode; model weights are not updated.Source: Figure 9 · p. 19, C IMPLEMENTATION DETAILS · p. 20

D SIMULATION TO OBSERVATION ALIGNMENT NUMERICAL METRICS

The appendix evaluates how generated videos align with proxy renders and documents prompting, reference-video baselines, and reranker choices, while distinguishing visual agreement from physical fidelity.

D SIMULATION TO OBSERVATION ALIGNMENT NUMERICAL METRICS

D.1 VLM-BASED OBSERVABLE CUE AGREEMENT

  • GPT-5.6-luna rates proxy–generated agreement across fire-front location, fuel distribution, burned area, and terrain; identity-pair averages of approximately 95 for proxy pairs and 90 for generated pairs contextualize these scores.The ratings are not calibrated percentages of physical fidelity.Source: §D.1 · p. 21, Table 14 · p. 21

F.1 VIDEO GENERATION PROMPT FROM SIMULATION DATA

  • Simulation metadata controls prompt appearance: wind affects motion and smoke wording, while fuel maps and barriers are translated into land-cover and fire-break descriptions.Spotting language is an appearance control, not evidence that ember trajectories or spotting dynamics are simulated.Source: §F.1 · pp. 23–24

F.2.1 SYSTEM PROMPT FOR PROMPT REFINEMENT

  • Prompt refinement is instructed to preserve explicit wind-strength and scenario cues while improving clarity and generation stability.The template calls for a compact paragraph and avoids coordinates, fuel-model names, numeric fuel codes, and brittle shot-by-shot instructions.Source: §F.2.1 · pp. 24–25, §F.2 · p. 24

G.5 BASELINES WITH LABELED REFERENCE VIDEOS

  • With 16 labeled reference videos per query, the strongest exact-match baseline is GPT-5 mini at 27.27%, versus 51.5% for the full memory system; some cross-group model versions differ.The differing versions mean this is not a matched-model ablation of adding examples.Source: §G.5 · p. 28, Table 20 · p. 28

H.1 OUR RERANKING METHOD

  • The downstream reranking build retrieves with Qwen embeddings and then uses GPT to score candidates from the retrieved shortlist.Source: §H.1 · p. 28

H.3 RERANKER ACCURACY AND LATENCY

  • GPT-5 and Gemini 2.5 Pro both achieve 51.52% exact four-tag accuracy, while local Qwen3-VL-Reranker-2B is faster but less accurate in the tested configurations.API and local execution are distinct deployment conditions, and reported scorer latency is not end-to-end pipeline runtime.Source: Table 21 · p. 29, §H.3 · p. 29

H.4 RETRIEVAL-SIZE ABLATION

This section examines retrieval-size effects, cross-generator results, and the associated specialist prompts, judging, report-generation, and output-schema procedures. It also presents generated reports as illustrative outputs requiring expert validation, not evidence of operational validity.

H.4 RETRIEVAL-SIZE ABLATION

  • Exact four-tag hit rate rises from 51.52% to 69.70% as the number of retrieved videos increases from one to four; these are retrieval-coverage measurements, not downstream predictions.Table 22 reports retrieval metrics, not final predictions from a multi-agent system using larger retrieval sets.Source: §H.4 · p. 29, Table 22 · p. 29

I DOWNSTREAM SIX-TASK SPECIALIST PROMPTS

  • The fixed pipeline uses direct-visual, lookup-memory, and combined-memory specialists; unanimous fields pass through and the judge resolves only disputed fields.The judge receives the disputed task keys while retaining all six specialist labels and reasons as context.Source: I DOWNSTREAM SIX-TASK SPECIALIST PROMPTS · p. 29, §I.8 · p. 35

I.1 DOWNSTREAM TASK LIST

  • The six downstream tasks cover spread speed, peak burning area, final burnscar, canopy cover, rapid growth with large burn, and severe fire behavior, with composite labels tied to stated thresholds.The thresholds include peak burning area of 90,000 m^2, final burnscar area of 300,000 m^2, and canopy cover above the train-set median of about 2.61%.Source: §I.1 · pp. 29–30

I.2 VISUAL REPORT AGENT

  • The visual specialist relies on query frames and conservative visible-evidence criteria, while memory specialists treat retrieved priors as soft context and defer to conflicting query evidence.The visual rubric rejects inferring high burnscar from smoke alone; combined memory includes the retrieved prompt and environmental and fire-behavior metadata.Source: §I.2 · p. 30, §I.3 · p. 31, §I.5 · p. 32

I.7 DOWNSTREAM TASK PREDICTION PROMPT FOR GPT, QWEN2-WILDFIRE-2B AND FORESTFIREVLM

  • Direct GPT Query, Qwen2-Wildfire-2B, and ForestFireVLM receive their own predicted four-tag state as soft context alongside video frames for six-task prediction.The prompt instructs models to prioritize visible query evidence when it conflicts with the predicted tags.Source: §I.7 · p. 34

I.8 MULTI-AGENT JUDGE PROMPTS

  • The judge prompt directs task-specific use of visual evidence and memory consensus, but its requested “calibrated” behavior is not empirically validated uncertainty calibration.The reported fixed pipeline routes only disputed task keys to the judge; unanimous specialist fields are retained.Source: §I.8 · p. 35
  • The judge treats top-k hybrid-memory consensus as a soft prior, favoring query frames for visible spread and memory for less-visible canopy and final-burnscar outcomes.It is instructed not to default to any single specialist and to explain acceptance or rejection of strong memory evidence against conservative visual predictions.Source: §I.8 · pp. 36–37

I.8.1 LEAVE-ONE-OUT JUDGE PROMPT VARIANT

  • Leave-one-out ablations change only the specialist-roster sentence; the judge uses frames for visible spread and footprint evidence and memory as calibrated, non-ground-truth evidence.Task-specific reliability rules give stronger memory consensus more weight for area thresholds and canopy, and require broad visual evidence or strong memory support for high labels.Source: §I.8.1 · pp. 38–39

J.3 STRUCTURED REPORT EXAMPLE

  • Reports combine visual evidence, predicted tags and downstream labels, retrieved metadata, reasoning, and suggestions, but illustrate format rather than operational validity.The narrative prompt omits internal retrieval and routing details; reports and response suggestions require expert validation before operational use.Source: J HTML REPORT GENERATION PROMPT · p. 39, §J.3 · p. 40

J.2 USER PROMPT TEMPLATE FOR REPORT GENERATION

  • Per case, the prompt supplies predicted location, severity, four latent tags, and six downstream task labels, requiring strict JSON with exactly one fire_spread_prediction key whose value is a compact 2–4-sentence paragraph.It prohibits mentioning ground truth, validation, agents, routing, retrieval, memory, model internals, or confidence scores.Source: §J.2 · pp. 39–40

K.1 FOUR-TAG CLASSIFICATION SCHEMA

  • The four-tag schema assigns binary low/high labels with short visible-evidence reasons to wind, moisture, slope, and vegetation regimes.Source: §K.1 · p. 40

K.2 CANDIDATE-RERANKING SCHEMA

  • Candidate reranking returns two candidates, each with an item ID, 0–100 match score, and short reason.Source: §K.2 · p. 40

K.3 DOWNSTREAM SPECIALIST AGENT / JUDGE SCHEMA

  • The downstream specialist and judge schema defines six label-and-reason outputs covering spread, peak area, burnscar, canopy cover, rapid growth with large burn, and severe behavior.Source: §K.3 · p. 41

L.1 CROSS-GENERATOR EVALUATION

  • 41.00% and 44.44% accuracy versus 28.00% and 27.27% for the strongest direct VLM baselines in the two cross-generator query sets.The evaluation swaps LTX-2.3 + AVControl and VACE as video-generation backbones, without changing the simulator or proxy renderer.Source: §L.1 · p. 41

L.2 EXTERNAL UAV EVALUATION

Real-UAV evaluations assess smoke detection, fire presence, and visible fire-area ranking, while their annotations do not jointly validate recovery of environmental drivers. Additional visual results provide generated-video examples and a per-attribute comparison.

L.2 EXTERNAL UAV EVALUATION

  • The real-UAV tests cover smoke detection, fire presence, and visible fire-area assessment, but their annotations do not jointly validate four environmental drivers.They assess observable monitoring tasks rather than physical-driver recovery.Source: §L.2 · p. 41
  • On Boreal Forest Fire, our method, Claude 4, and ForestFireVLM-7B each reach 100% accuracy, precision, recall, and F1.This evaluation does not distinguish among those methods.Source: §L.2 · p. 41, Table 23 · p. 41
  • FireMan-UAV-RGBT evaluation measures binary fire presence with overall and balanced accuracy to account for class imbalance.The supplied passages do not report the per-method outcomes.Source: §L.2 · p. 41, Table 24 · p. 42
  • On FLAME, our method has the highest NDCG and ties Claude Sonnet 4.6 on Hit@1, while Gemini 2.5 Pro has the highest mean AP.These annotations concern visible fire regions, not a complete physical characterization of fire severity.Source: §L.2 · p. 42, Table 25 · p. 42

M ADDITIONAL VISUAL RESULTS

  • Additional visuals include generated-video frames (Figure 10) and a per-attribute comparison complementing Table 3 (Figure 8).Figure 10 shows additional videos generated from 2D simulations through procedural 3D proxies.Source: M ADDITIONAL VISUAL RESULTS · p. 42, Figure 10 · p. 42

Related papers

Loading 2610.02451v1…