Fire: Multiscale Detection, Forecasting & Response

• Fire is a combustion-driven phenomenon exhibiting ignition, spread, and thermal-acoustic dynamics across compartment, wildland, and urban environments.
• Recent studies leverage lightweight neural networks and multimodal sensing to enhance fire detection, localization, and event tracking in complex scenes.
• Innovative forecasting and digital twin systems integrate simulation models and real-time data to improve emergency response and ecological risk assessments.

Searching arXiv for the cited FIRE-related papers to ground the article and confirm relevant IDs. arXiv Search Query: FIRE fire detection wildfire spread digital twin FireNet Cell2Fire FireSentry Fire Dynamic Vision Fire as a Service DetectiumFire YOLO-FireAD FIDN FET PyroGuardian acoustics compartment fire ecology Fire is a combustion-driven physical process and a coupled environmental hazard whose scientific treatment spans ignition and spread, heat and smoke transport, ecological feedbacks, sensing, prediction, and response. In contemporary arXiv literature, fire is studied across distinct but connected regimes: as a binary visual event in embedded detection systems, as a propagating wildfire over heterogeneous landscapes, as a thermally and acoustically altered physical environment, and as an operational emergency context requiring digital twins, event tracking, robotics, and multimodal decision support. Taken together, this body of work treats fire not as a single phenomenon but as a hierarchy of interacting processes and representations, from pixel-level flame signatures to incident-level event histories and landscape-scale burn outcomes (Jadon et al., 2019).

1. Fire as a physical and environmental process

Fire research in the supplied literature distinguishes several physical settings. One is the compartment fire, where combustion creates temperature gradients and inhomogeneous time-varying temperature, density, and flow fields that alter wave propagation in enclosed spaces. In that setting, the dominant modeled mechanism is the fire-induced temperature field, which yields a time-varying sound-speed field; the reported consequences are that wave fronts arrive earlier, arrival times become more variable, room modes shift upward in frequency, and high-frequency modal structure is strongly attenuated, especially above about $2500\ \text{Hz}$ (Abbasi et al., 2021). Another is the wildland or forest fire, where spread is controlled by fuel, weather, moisture, and topography, and where simulation typically treats propagation over heterogeneous rasterized landscapes rather than as a single localized flame source (Pais et al., 2019). A third is the urban incident fire, whose main modeled consequence may be not perimeter growth but smoke transport and particulate exposure within a built environment, as in urban fire-and-smoke digital twins (Jiao et al., 2023).

At the level of basic propagation modeling, several abstractions recur. In the acoustic fire literature, sound speed increases with temperature according to $c=\sqrt{\gamma R T}$, so fire changes both travel times and modal structure in a room (Abbasi et al., 2021). In wildfire simulation, Cell2Fire assumes elliptical fire growth within each cell, with directional spread governed by quantities such as head rate of spread, back rate of spread, flank rate of spread, and length-to-breadth ratio, thereby embedding anisotropic spread into a raster message-passing simulator (Pais et al., 2019). In ecological fire systems, fire is represented as a stochastic disturbance whose return time depends on vegetation state, generating a fire–vegetation feedback in which plants both shape and respond to the fire regime (Magnani et al., 2023).

This suggests that fire is best understood as a multi-physics and multi-scale process. A plausible implication is that no single representation—flame image, perimeter polygon, hotspot point, or thermal field—is sufficient across all fire tasks. Different FIRE subfields instead choose observables matched to their operational target: sound-speed perturbations for compartment diagnostics, hotspot clusters for geostationary event tracking, raster burn states for landscape spread, or visual flame signatures for embedded alarms.

2. Sensing, perception, and visual recognition

A major branch of FIRE research concerns visual sensing. FireNet defines the core task as binary visual fire detection from images or video frames, classifying an input as fire or non-fire. Its design is explicitly specialized for low-cost embedded deployment and rejects fine-tuned large image-classification backbones as too bulky for practical fire-alarm products. The model uses a $64 \times 64 \times 3$ input, 3 convolution layers, pooling and dropout after each convolution, 2 hidden dense layers, and a 2-unit Softmax output, for a total of $646{,}818$ trainable parameters and an on-disk size of $\sim 7.45$ MB. On Raspberry Pi 3B, the paper reports 24 fps, which is the main empirical basis for its lightweight claim (Jadon et al., 2019).

Subsequent detector work shifts from binary classification to localization. FSDNet combines a DenseNet-inspired dense-connected feature extractor with a YOLOv3-style multi-scale detection head, plus SPP, Mosaic augmentation, and CIoU loss. It targets flame and smoke localization in complex scenes with scale variation, interference, partial occlusion, and dataset limitations. On the MS-FS dataset, it reports fire AP $=0.934$, smoke AP $=0.801$, and mAP $=0.868$, while on classification benchmarks it reports $99.82\%$ accuracy on DS1 and $91.15\%$ on DS2 (Zhu et al., 2023). YOLO-FireAD pushes the efficiency–accuracy balance further by inserting an Attention-guided Inverted Residual Block and a Dual Pool Downscale Fusion Block into a YOLOv8n-style detector. On the reported fire_detection benchmark, it reaches precision $c=\sqrt{\gamma R T}$0, recall $c=\sqrt{\gamma R T}$1, $c=\sqrt{\gamma R T}$2, $c=\sqrt{\gamma R T}$3, $c=\sqrt{\gamma R T}$4, with only $c=\sqrt{\gamma R T}$5M parameters, $c=\sqrt{\gamma R T}$6 GFLOPs, and model size $c=\sqrt{\gamma R T}$7 MB (Pan et al., 27 May 2025).

The literature also broadens from pure detection to multimodal understanding. DetectiumFire introduces a dataset with 22.5k fire-related images and 2.5k real-world fire-related videos, annotated with both bounding boxes and textual prompts describing burning object, environment, and severity. It explicitly includes controlled and uncontrolled flames, challenging hard negatives, and a four-level severity taxonomy, and reports that a LLaMA-3.2-11B-Vision-Instruct model fine-tuned on the dataset improves from $c=\sqrt{\gamma R T}$8 to $c=\sqrt{\gamma R T}$9 on burning-object recognition, from $64 \times 64 \times 3$0 to $64 \times 64 \times 3$1 on surrounding-environment recognition, and from $64 \times 64 \times 3$2 to $64 \times 64 \times 3$3 on fire-severity recognition (Liu et al., 4 Nov 2025). Fire Dynamic Vision addresses a different perception problem: not whether fire is present, but how fire fronts and plume boundaries can be segmented and tracked across scales using hybrid RGB-HSV or thermal thresholding, DBSCAN region separation, and $64 \times 64 \times 3$4-shape boundaries, thereby extracting physically interpretable front and plume dynamics from image sequences (Sagel et al., 2024).

A recurring misconception in this area is that visual fire detection and smoke detection are interchangeable. The supplied literature is explicit that FireNet itself is a fire detector, while smoke detection in the corresponding deployed system is sensor-assisted rather than learned visually by the same network (Jadon et al., 2019). Another misconception is that all vision models solve the same task. The papers instead separate classification, object detection, dynamic boundary tracking, and vision-language reasoning into distinct FIRE perception problems.

3. Datasets and benchmark construction

The FIRE literature repeatedly identifies data scarcity, redundancy, and annotation incompleteness as limiting factors. FireNet criticizes over-reliance on the Foggia dataset, noting that it contains many similar frames and may not be diverse enough for robust real-world generalization. To address this, it constructs a custom training dataset with 1,124 fire images and 1,301 non-fire images, split 70%/30% with no overlap, and evaluates on a realistic test set built from 46 fire videos, 16 non-fire videos, and 160 additional challenging non-fire images (Jadon et al., 2019). FSDNet likewise introduces MS-FS, a fire-and-smoke dataset containing 11,938 images with manual Pascal VOC annotation for flame and smoke after filtering by object area, SNR, and RGB-based background criteria (Zhu et al., 2023).

More recent work emphasizes benchmark design rather than only raw scale. DetectiumFire explicitly reduces duplicate and near-duplicate imagery and reports lower duplication ratios than D-Fire under both PHash and CNN-based similarity analyses, while also incorporating captions, severity, videos, and preference pairs for RLHF-style generative training (Liu et al., 4 Nov 2025). FireSentry moves from static image corpora to synchronized multimodal wildfire observations: visible and infrared UAV video, in-situ environmental measurements, and manually validated fire masks, collected at raw $64 \times 64 \times 3$5, 30 FPS, and downsampled for benchmarking to $64 \times 64 \times 3$6, 1 FPS, with 20,710 valid frames across five prescribed-burn regions (Zhou et al., 3 Dec 2025).

This suggests a structural shift in FIRE datasets from narrowly supervised image collections toward task-specific benchmark ecosystems. A plausible implication is that benchmark quality now depends not only on class balance and annotation count, but also on de-duplication, hard negatives, contextual labels, temporal continuity, and modality alignment. The same trend appears outside vision-only detection: WILDFIREIA defines a national-scale initial-attack-failure benchmark using fixed event units, chronological splits, forbidden-feature lists, and discovery-time-only public inputs to prevent leakage, showing that benchmark protocol is itself a central research object in FIRE prediction (Xu et al., 14 Jun 2026).

4. Forecasting spread, event evolution, and final burn outcomes

Wildfire forecasting appears in several distinct forms. Cell2Fire is a cell-based forest and wildland fire growth simulator in which a landscape is discretized into square cells, each with fuel, slope, elevation, location, and state. Fire spreads by send/receive message passing between adjacent cells under a Moore neighborhood, with within-cell propagation assumed elliptical and directional spread rates taken from an external fire spread model such as the Canadian FBP System. It is validated against Prometheus and reports, for example, mean $64 \times 64 \times 3$7, mean $64 \times 64 \times 3$8, and mean $64 \times 64 \times 3$9 across ten large British Columbia fires, while offering strong parallel performance for large-scale scenario simulation (Pais et al., 2019).

A different forecasting target is the final burnt area once ignition has already occurred. Fire-Image-DenseNet (FIDN) formulates this as an image-to-image prediction problem: given the first three daily burn observations plus environmental and meteorological drivers, predict the final binary burn scar. Using 303 fire events in the western United States and a DenseNet-based multi-resolution encoder–decoder, it reports $646{,}818$0, $646{,}818$1, and inference time $646{,}818$2 s on the benchmark comparison subset, with MSE about 82% lower than CA and about 67–68% lower than MTT baselines (Pang et al., 2024).

Event-based geostationary monitoring introduces another formulation. The Fire Event Tracker associates MTG-FCI hotspot detections into persistent events using DBSCAN with a 2 km radius, alpha-shape or simpler geometries depending on point count, and a 48-hour termination rule. It computes evolving AOI geometry, event-level FRP, postprocessed FRE, and a perimeter-based Forward Rate of Spread, and is used both retrospectively for Mediterranean 2025 analysis and operationally for Portugal and SILEX/EUBURN monitoring. For events above 400 ha, final FET AOI correlates with EFFIS burned area at $646{,}818$3 (Paugam et al., 4 Jun 2026).

At finer scales, FireSentry treats spread forecasting as short-horizon future video and mask prediction. Its FiReDiff paradigm first predicts future infrared sequences conditioned on past observations, land-vegetation, and environmental variables using a conditional diffusion objective,

$646{,}818$4

and then segments predicted infrared frames with SAM2. The paper reports substantial improvements over base generative models, including up to $646{,}818$5 in PSNR, $646{,}818$6 in SSIM, $646{,}818$7 in LPIPS, $646{,}818$8 in FVD, $646{,}818$9 in AUPRC, $\sim 7.45$0 in F1, $\sim 7.45$1 in IoU, and $\sim 7.45$2 in MSE when FiReDiff is applied to generative baselines (Zhou et al., 3 Dec 2025).

Across these works, “fire prediction” therefore spans at least four non-equivalent tasks: cell-based spread rollout, event-level hotspot association, final burnt-area inference, and short-horizon thermal video forecasting. This suggests that FIRE forecasting research is increasingly organized by operational horizon and output representation rather than by a single canonical predictive target.

5. Fire in coupled environments: acoustics, ecology, smoke, and digital twins

Fire also appears as a system-level modifier of its environment. In room acoustics, the compartment-fire study shows that temperature-induced sound-speed changes alone can explain earlier arrivals, delay spreading, low-frequency modal upshifts, and attenuation or disappearance of high-frequency structure. The work couples FDS temperature fields to COMSOL and Bellhop and argues that temperature variation is a major cause of observed acoustic changes, with practical implications for alarm audibility, source localization, and possible acoustic diagnostics (Abbasi et al., 2021).

In ecology, Magnani et al. model fire-prone plant communities through competition, stochastic fires, and fire–vegetation feedback. Fire return time depends on vegetation flammability, and post-fire cover is reduced by a fire-response parameter $\sim 7.45$3. The paper argues that the fire response of the strongest competitor determines whether a system is monostable or exhibits alternative ecological states, with examples spanning Mediterranean forests, humid tropical savanna–forest mosaics, and boreal forests (Magnani et al., 2023). This shifts FIRE from a disturbance external to vegetation toward a co-produced ecosystem regime.

Urban smoke modeling introduces yet another coupled framework. FireCOM builds a fire-and-smoke digital twin by combining live incident feeds, city geometry, weather inputs, VSmoke for fast Gaussian-plume-style approximation, Blender + MantaFlow for geometry-aware 3D smoke simulation, and browser deployment. For Austin, it reports under-1-second 2D forecast generation and about 30-second 3D forecast generation up to three hours ahead, while also concluding that existing urban air-quality sensor density is insufficient to validate urban fire presence (Jiao et al., 2023).

In robotics, Fire as a Service augments robot simulators with asynchronous co-simulation of thermally and optically meaningful fire dynamics. Fire-X provides GPU-accelerated reactive flow and smoke, while the added framework computes thermal hazard, renders smoke/flame aligned to robot cameras, and exposes a thermal costmap for planning. A central hazard metric is the accumulated thermal dose,

$\sim 7.45$4

approximated discretely from incident thermal radiation measurements along a trajectory (Wagner et al., 19 Mar 2026).

These works collectively indicate that FIRE is not only an object to detect or forecast, but also a medium that transforms other physical and operational subsystems: acoustics, vegetation dynamics, urban air quality, and robot perception. A plausible implication is that future FIRE research will increasingly require cross-domain co-modeling rather than isolated fire-only predictors.

6. Emergency response, IoT systems, and operational support

A strong applied thread in the literature concerns complete response systems rather than standalone models. FireNet is integrated into an IoT-enabled fire-and-smoke detection unit on Raspberry Pi 3B, connected to a camera, smoke sensor, microcontroller, two distinct alarms, Amazon S3 for visual upload, and Twilio for SMS/MMS alerts. In that system, visual fire detection complements physical smoke sensing to reduce false triggering and detection delay (Jadon et al., 2019). PyroGuardian takes a different operational role: wearable firefighter monitoring. Using LoRa-linked PyroHelm and PyroStrap modules plus the PyroPortal Android interface, it tracks heart rate, blood oxygen saturation, body temperature, ambient temperature, GPS, and orientation. The reported maximum communication range is 610 meters with obstacles and walls, battery life is estimated at about 8 hours, and survey responses after a live burn-building demonstration average 4.5/5 for capability fit and 4.18/5 for ease of use (Kaplan et al., 2024).

FireFly extends FIRE response into conceptual autonomous drone first response. It proposes a squad of drones dispatched after in-building sensor detection, using GPS, PX4 autopilot with Raspberry Pi, LoRa communication, thermal imaging, 3D radar, gas sensing, Jetson Nano inference, rescue-kit delivery, window breaching, and fire extinguishing balls. The paper is explicit that this is a conceptual system proposal rather than an experimentally validated platform; still, it specifies a 45-minute total fly time, 20 minutes for transit, 25 minutes remaining for rescue and fire control, and a minimum payload of 5. Kgs (Mnaouer et al., 2021).

At the monitoring-and-command level, event-based FCI tracking and urban digital twins both support incident intelligence rather than direct suppression. FET products support re-intensification detection, active-area identification, and plume-model initialization (Paugam et al., 4 Jun 2026), while FireCOM provides public-facing and operational smoke-path forecasts for active urban incidents (Jiao et al., 2023). This suggests a layered FIRE response stack: detection, crew monitoring, incident tracking, smoke forecasting, and in some cases active robotic or drone intervention.

A common misconception is that FIRE response systems are now fully autonomous. The supplied papers do not support that reading. FireFly is conceptual and lacks controlled validation (Mnaouer et al., 2021), PyroGuardian is a commander-assistance platform rather than an autonomous decision-maker (Kaplan et al., 2024), and many deployed detectors remain narrow classifiers embedded within larger sensor and notification architectures rather than general emergency-response agents (Jadon et al., 2019).

7. Research trends, limitations, and future directions

Several cross-cutting limitations recur. Reproducibility is often incomplete in detector papers: FireNet omits optimizer, learning rate, batch size, epochs, and exact augmentation operations (Jadon et al., 2019), and FSDNet does not report full layerwise tensors, anchor settings, input resolution, optimizer, or hardware (Zhu et al., 2023). Benchmark leakage and target mismatch are active concerns in wildfire prediction, motivating the strict discovery-time feature contract and chronological split in WILDFIREIA (Xu et al., 14 Jun 2026). Dataset realism is also a repeated issue: SCU-CGAN explicitly addresses the scarcity of indoor fire data by generating localized synthetic fire imagery from non-fire images and shows improvements such as a reported $\sim 7.45$5 increase in YOLOv5n $\sim 7.45$6 after augmentation, but its training data are largely SDXL-generated indoor scenes and the paper itself notes the need for broader real-world indoor and outdoor fire datasets (Kim et al., 9 Dec 2025).

Another recurring limitation is that many FIRE systems remain strong in one dimension but narrow in scope. Vision detectors may not model smoke physics; event trackers may overestimate AOI due to coarse resolution; fine-grained wildfire benchmarks may still be limited to prescribed burns in one province; firefighter monitoring systems may lack indoor localization; urban smoke twins may lack dense validation data (Zhou et al., 3 Dec 2025). This suggests that FIRE research is now less constrained by isolated algorithmic novelty than by integration across modalities, scales, and operational constraints.

The emerging directions are correspondingly integrative. DetectiumFire points toward multimodal vision-language fire reasoning and preference-aligned synthetic generation (Liu et al., 4 Nov 2025). FireSentry points toward thermal-video-first forecasting rather than mask-only prediction (Zhou et al., 3 Dec 2025). Fire as a Service points toward thermally grounded robot simulation and policy learning in hazardous scenes (Wagner et al., 19 Mar 2026). Event-based geostationary fire tracking points toward continuous operational incident histories rather than isolated hotspots (Paugam et al., 4 Jun 2026). A plausible implication is that the next phase of FIRE research will center on unified systems that connect sensing, simulation, reasoning, and action while preserving strict benchmark discipline and deployment realism.