---
title: 'SmokeSeer: Multi-Method Smoke Analysis'
url: https://www.emergentmind.com/topics/smokeseer
type: topic
---

# SmokeSeer: Multi-Method Smoke Analysis

Searching arXiv for papers related to “SmokeSeer” and closely related smoke-detection/reconstruction systems.
SmokeSeer is a name applied in the arXiv literature to multiple smoke-analysis systems rather than to a single canonical framework. The name has been used for a 2018 citizen-science platform for detecting and documenting industrial fugitive emissions from timelapse imagery, a 2022 camera-based wildfire smoke early-warning and alert system, and a 2025 multimodal inverse-rendering method for smoke removal and 3D scene reconstruction from RGB-thermal multi-view video [1809.06263] [2211.01629] [2509.17329]. This suggests that “SmokeSeer” is best understood as a recurrent label for systems that operationalize smoke perception under distinct problem formulations: event discovery, early warning, and scene restoration.

## 1. Terminological scope and research usage

Within the arXiv record, SmokeSeer does not denote one stable codebase, benchmark, or model family. Instead, it appears across at least three technically different systems with different sensing modalities, supervision regimes, and operational objectives [1809.06263] [2211.01629] [2509.17329].

| Usage | Task domain | Characteristic method |
|---|---|---|
| SmokeSeer (2018) | Industrial smoke monitoring | Heuristic detection, timeline visualization, fast-forwarding, animated smoke documentation |
| SmokeSeer (2022) | Wildfire early warning | ResNet-18 backbone, FPN, deformable convolutions, FCOS-like detection, human-confirmed alerts |
| SmokeSeer (2025) | Smoke removal and 3D reconstruction | RGB-thermal 3D Gaussian Splatting with explicit smoke/non-smoke decomposition |

The broader research landscape around these systems is larger than the name itself. Closely related work includes early smoke segmentation with Focus, Separation, and Domain Fusion modules [2306.04474], Bayesian smoke segmentation with explicit epistemic and aleatoric uncertainty [2303.00900], optical-flow-guided uncertainty-aware smoke detection [2508.14597], hyperspectral smoke segmentation with a mixture-of-prototypes architecture [2602.10858], and wildfire-to-smoke surrogate modeling with multilinear operators [2605.04164]. The result is a heterogeneous field in which “SmokeSeer” sits at the intersection of computer vision, environmental monitoring, inverse rendering, and decision support.

## 2. Industrial emissions monitoring and citizen-science workflow

The earliest explicit SmokeSeer system is a software platform for helping citizen scientists detect, inspect, and document industrial smoke emissions from long timelapse camera streams in Pittsburgh, where a local community monitored fugitive emissions from a coke refinery [1809.06263]. Its central problem was scale: the live camera captured a frame every 5 seconds, so one day of daytime imagery contained nearly 17,000 frames, making manual review impractical when the goal was not only detection but also the production of communicable evidence.

The system was organized as a four-stage workflow. First, a smoke detection algorithm identified candidate smoke frames and grouped them into events. Second, an interactive timeline visualization highlighted those events for rapid seeking. Third, an autonomous fast-forward mode skipped uninteresting frames and played only likely smoke segments. Fourth, the system automatically generated animated smoke images that could be inserted into reports or documents. The design was explicitly framed around citizen science, with the purpose of shifting effort from frame-by-frame review toward interpretation, documentation, and storytelling [1809.06263].

At the algorithmic core, SmokeSeer used a heuristic detector for daytime imagery from a static camera. Background estimation relied on a running median over the previous 60 frames,
$$
B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),
$$
after converting 8-bit RGB frames to floating point in \([0,1]\). The detector then combined two complementary change-detection paths. A high-frequency path applied difference-of-Gaussians filtering and normalized background subtraction,
$$
bgSub(I,B)=\frac{I-B}{\max(I+B,0.1)},
$$
followed by entropy filtering and smoothing. A second path analyzed intensity changes after CLAHE enhancement using current-vs-background and current-vs-previous comparisons, then AND-combined their thresholded masks. These outputs were merged into a change-detection mask and passed to a texture-segmentation stage based on Laws’ texture energy measures and 125-dimensional per-pixel descriptors. PCA preserved 98% of the energy before accelerated \(k\)-means++ clustering partitioned pixels into textons and image regions [1809.06263].

Region filtering then rejected obvious non-smoke candidates using bounding-box heuristics, median color constraints, size tests, change consistency, and a shadow-rejection test based on kernel density estimation over background-subtracted values. Event detection treated the number of smoke pixels per frame as a time-series signal, detected peaks and peak widths, and merged nearby segments into emission events. This event-centric representation was significant because the operational question was when a plume started and ended, not merely which frames were positive [1809.06263].

Evaluation used 21 days of 2015 daytime timelapse footage, each day containing 16,838 frames, with processing on a 496-by-528 pixel window taking about 30 minutes per day on a workstation with two hex-core Intel Xeon X5670 CPUs using all cores. A predicted segment was counted as a true positive if more than 30% of its frames overlapped ground truth. The paper reported persistent false positives from steam and fast-moving shadows, false negatives for low-opacity or small smoke, and an inability to handle nighttime imagery. These limitations were not incidental: they defined the boundary between a practical heuristic system and a general smoke-recognition model [1809.06263].

## 3. Wildfire early warning and camera-based alerting

A later explicit use of the name appears in a real-time wildfire smoke early detection and alert system deployed in the United States to monitor near real-time feeds from hundreds of cameras every day [2211.01629]. In this formulation, SmokeSeer was an object-detection-style pipeline for spotting wildfire smoke before a fire became large and uncontrollable. The system consumed live camera images, classified regions as wildfire smoke versus non-wildfire, drew bounding boxes around detected smoke, and issued alerts. The operational logic was deliberately human-in-the-loop: a smoke detection alert was first sent to a human operator, and only after confirmation was it propagated to other users such as first responders.

The training corpus was built from the Alert Wildfire camera network. Images were collected from 400 different cameras across diverse terrains, sampled monthly from June 2021 through March 2022, producing 90,000 labeled images with bounding boxes and wildfire/non-wildfire annotations. The appendix reported 46,416 training images and 41,134 validation images. Architecturally, the detector used a ResNet-18 backbone, an FPN for multi-scale smoke representation, deformable convolutions inside the pyramid, and a prediction head composed of \(1\times 1\) convolutions producing classification and bounding-box regression outputs. The paper described the model as similar in spirit to FCOS [2211.01629].

The training objective adapted the FCOS loss:
$$
loss = \frac{1}{N_{pos}}\Big(\sum_{x, y}L_{cls}(p_{x, y}, c_{x, y}) + \sum_{x, y}\mathbb{1}_{\{c_{x, y} = 1\}}L_{reg}(t_{x, y}, t^{*}_{x, y}) +  \sum_{x, y}\mathbb{1}_{\{c_{x, y} = 1\}}L_{cen}(t^{*}_{x, y})\Big),
$$
with focal loss for classification, IoU loss for regression, and centerness loss for positive samples only. A modified adaptive training sample selection procedure scored candidate points inside each ground-truth box by
$$
Score_{x, y} = (IOU)_{x,y} * p_{x, y},
$$
then retained points above a mean-plus-standard-deviation threshold as positives while ignoring the rest. The paper stated that this modified ATSS selected a greater number of relevant training samples than the original ATSS formulation [2211.01629].

Validation results reported \(TP=10690\), \(TN=27005\), \(FP=2205\), and \(FN=1234\), corresponding to accuracy \(0.916\), precision \(0.829\), recall \(0.896\), and false positive rate \(0.075\). Operational timing was a major contribution: across 869 wildfire smoke events observed from June 2022 to July 2022, 60.5% were detected within 60 seconds from the start of smoke, 76.8% within 3 minutes, 86.9% within 5 minutes, and 100% after 5+ minutes. The model was reported to detect smoke on average 2 minutes and 57 seconds faster than a human operator [2211.01629].

The paper was explicit about confounders and failure modes. Clouds, fog, and industrial smoke from chimneys could resemble wildfire smoke and generate false positives; false negatives also occurred, although some events were detected later as the smoke plume grew. This places SmokeSeer in the class of operational assistive systems rather than fully autonomous decision-makers. Its function was to continuously scan camera feeds, accelerate human awareness, and provide actionable early warnings for firefighting and infrastructure protection [2211.01629].

## 4. Smoke removal and 3D scene reconstruction

A third major use of the name is the 2025 system “SmokeSeer: 3D Gaussian Splatting for Smoke Removal and Scene Reconstruction,” which shifted the problem from smoke detection to multimodal inverse rendering [2509.17329]. Here the objective was simultaneous smoke removal and 3D scene reconstruction from multi-view video under dense, spatially heterogeneous, and temporally varying smoke. The key observation was that visible RGB light is strongly scattered by smoke, while long-wave infrared thermal radiation is much less affected; thermal imagery could therefore preserve geometry even when RGB was heavily obscured.

The method used a three-stage pipeline. First, RGB and thermal images were aligned and smoke regions in RGB were segmented so they could be excluded from feature matching. Second, a vanilla 3DGS reconstruction on the thermal sequence yielded a coarse but smoke-free estimate of scene geometry. Third, the scene was jointly optimized under RGB and thermal supervision using two Gaussian sets: \(\mathcal{G}\) for surface Gaussians and \(\mathcal{S}\) for smoke Gaussians. Surface and smoke components had modality-specific appearance parameters, and smoke Gaussians were time-varying [2509.17329].

The decomposition was framed volumetrically. For a ray \(\mathbf{r}(t)=\mathbf{o}+t\mathbf{d}\), transmittance was written as
$$
T_{\sigma}(t) \coloneqq \exp\left(-\int_{t_n}^{t} \sigma(s)\, ds\right),
$$
and the clear-scene rendering was the surface-only term
$$
C_{\text{clear}}(\mathbf{r}, \mathbf{d}) = \int_{t_n}^{t_0} c(\mathbf{r}(t), \mathbf{d}) \sigma(t) T_{\sigma}(t)\, dt.
$$
Operationally, rendering only \(\mathcal{G}\) produced a smoke-free view, while rendering \(\mathcal{G}\cup\mathcal{S}\) reconstructed the observed smoky scene. SmokeSeer therefore cast smoke removal as inverse rendering rather than frame-wise image restoration [2509.17329].

Optimization combined rendering loss with priors on smoke opacity, smoke color, monochromaticity, thermal-guided depth consistency, and smoke-mask alignment:
$$
L_{\mathrm{total}} = \lambda_{\mathrm{render}} L_{\mathrm{render}} + \lambda_{\mathrm{smoke\_alpha}} L_{\mathrm{smoke\_alpha}} + \lambda_{\mathrm{smoke\_color}} L_{\mathrm{smoke\_color}} + \lambda_{\mathrm{mono}} L_{\mathrm{mono}} + \lambda_{\mathrm{depth}} L_{\mathrm{depth}} + \lambda_{\mathrm{mask}} L_{\mathrm{mask}}.
$$
Dynamic smoke motion was modeled by a deformable 3DGS-style MLP deformation field that predicted offsets in position, scale, and rotation from canonical space as a function of Gaussian position and time. Opacity was additionally modeled with a bimodal Gaussian distribution following explicit-dynamic Gaussian splatting literature [2509.17329].

The system was evaluated on a synthetic Blender/Mantaflow dataset of 10 scenes with 150 frames per scene and on a real-world RGB-thermal multi-view dataset collected with a Spirit drone in collaboration with a county fire department. On synthetic data, SmokeSeer reported PSNR, SSIM, and LPIPS. The full multimodal method achieved 19.92 PSNR, 0.76 SSIM, and 0.247 LPIPS, compared with 15.08/0.38/0.326 for an RGB-only version and 14.42/0.37/0.318 for ImgDehaze + 3DGS. The paper stated that the improvement could be as large as 4.8 dB PSNR over RGB-only approaches in heavy smoke. Ablations showed that removing smoke priors degraded decomposition quality, that the depth consistency loss was especially impactful, and that omitting monochromaticity could produce unrealistically colored smoke [2509.17329].

Implementation used PyTorch, the official 3D Gaussian Splatting repository, and a deformable component adapted from Deformable 3D Gaussians. Stage 2 and Stage 3 each ran for 15,000 iterations. For a typical real-world scene, Stage 1 took about 1 hour, Stage 2 about 10 minutes, and Stage 3 about 30 minutes on an RTX 4090 system with 128 GB RAM and an Intel i9-13900K. The method explicitly could not handle very dense smoke that completely occluded the scene, and the authors suggested Navier–Stokes or other physics-informed priors, as well as diffusion-model priors, as future directions [2509.17329].

## 5. Architectural motifs and adjacent smoke-analysis research

Although the three explicit SmokeSeer systems are heterogeneous, they define a progression in methodology. The 2018 system relied on static-camera heuristics, hand-tuned thresholds, texture segmentation, and interactive visualization [1809.06263]. The 2022 system used discriminative deep detection with a one-stage detector, large-scale annotated data, and a human confirmation loop [2211.01629]. The 2025 system moved into multimodal inverse rendering with explicit scene decomposition, Gaussian primitives, and physically motivated regularization [2509.17329]. The name therefore spans detection, segmentation, alerting, and reconstruction, rather than any single representational paradigm.

Related smoke-segmentation work sharpened several of these themes. FoSp, designed for early smoke segmentation, decomposed the task into coverage and boundary purity through Focus, Separation, and Domain Fusion modules, introducing the real-image SmokeSeg dataset and reporting \(72.05\%\) \(F_\beta\) on SmokeSeg and a 7.71-point \(F_\beta\) gain over SegFormer on the small-smoke subset [2306.04474]. TGBGM addressed overconfidence and ambiguity by combining a Bayesian latent-variable segmentation model with a transmission-guided local coherence loss, reporting \(F_\beta=0.791\) and mMse \(=0.002\) on the total SMOKE5K test set and improving ECE from 0.0159 to 0.0039 after calibration [2303.00900]. A later uncertainty-aware monocular framework fused optical-flow-derived motion maps with RGB appearance and reported perfect validation-set Accuracy, Precision, Recall, and F1-score of 1.00 with near-zero ECE, although its contribution was not explicitly named SmokeSeer in the manuscript [2508.14597].

The sensing frontier also expanded beyond RGB. Hyperspectral smoke segmentation via Mixture of Prototypes introduced HSSDataset and MSSDataset, using band split, prototype-based spectral representation, and a dual-level router for adaptive band weighting. In the accuracy-oriented setting, MoP reported 73.63 F1 and 59.21 mIoU on HSSDataset, and 88.11 F1 and 79.55 mIoU on MSSDataset [2602.10858]. At a larger geophysical scale, a wildfire-to-smoke map based on multilinear operators learned a fast surrogate from time-since-ignition to smoke concentration fields, with CPU training taking less than 30 seconds and each forward call less than 1 ms; on holdout data it reported IoU \(=65\%\) and AUC \(=0.95\) for smoke detection [2605.04164]. These neighboring systems indicate that the underlying field has moved from heuristic detection toward calibrated prediction, spectral sensing, and reduced-order environmental surrogates.

A separate but adjacent branch concerns 3D restoration. SmokeGS-R, for example, pursued a geometry-first strategy for real-world multi-view smoke restoration by generating physics-guided pseudo-clean supervision, training a clean-only 3DGS source model, and harmonizing appearance post hoc in LAB space. Its reported re-evaluation on released challenge scenes yielded PSNR \(=15.209\), SSIM \(=0.644\), and LPIPS \(=0.551\) [2604.05301]. Compared with the 2025 SmokeSeer formulation, this neighboring line suggests an emerging distinction between explicit smoke-scene decomposition and geometry-first restoration pipelines.

## 6. Evaluation regimes, limitations, and broader significance

A recurring misconception is that smoke analysis is a single task. The SmokeSeer literature shows otherwise. Industrial timelapse monitoring targets event discovery and documentation under static viewpoints [1809.06263]; wildfire early warning emphasizes detection speed, false alarms, and alert routing [2211.01629]; multimodal 3D SmokeSeer addresses visibility recovery and scene reconstruction under dense, dynamic smoke [2509.17329]. Metrics accordingly differ: precision, recall, and F-score for event detection; accuracy, true detection rate, false detection rate, and time-to-detection for wildfire alerts; PSNR, SSIM, and LPIPS for restoration and novel-view synthesis. Direct comparison across these problem classes is therefore usually ill-posed.

Another misconception is that modern smoke systems are fully autonomous. The record is more restrained. The 2018 SmokeSeer was explicitly a workflow accelerator for citizen scientists, not a general autonomous detector [1809.06263]. The 2022 SmokeSeer inserted a human operator between model output and downstream alert propagation as a mitigation for false positives [2211.01629]. The 2025 SmokeSeer was an optimization-heavy offline pipeline whose first stage alone took about 1 hour for a typical real-world scene [2509.17329]. Even high-performing related systems retain strong assumptions: daytime imagery, static cameras, curated datasets, approximate labels, or failure under extreme smoke density.

Confusion between smoke and visually related phenomena remains a central challenge. Industrial systems reported false positives from steam and fast-moving shadows [1809.06263]. Wildfire early-warning systems identified clouds, fog, and industrial smoke from chimneys as confounders [2211.01629]. Segmentation work emphasized non-rigid shape, transparency, weak boundaries, and label ambiguity [2303.00900], while hyperspectral work explicitly targeted cloud interference and semi-transparent smoke regions [2602.10858]. This pattern indicates that smoke perception is as much an uncertainty-management problem as a classification problem.

The broader significance of SmokeSeer-type systems lies in their embedding within larger socio-technical infrastructures. In industrial monitoring, the citizen-science orientation aligns with the advocacy and environmental-justice perspective later formalized in Project RISE, which built a large-scale industrial smoke video dataset in collaboration with community members [2005.06111]. In wildfire applications, smoke detection is tied to response latency, utility protection, and exposure mitigation, while public-health work using the Smoke Sense app demonstrates that smoke is simultaneously a sensing target and a health-risk communication problem [2005.12017]. A plausible implication is that SmokeSeer is most consequential not when treated as an isolated algorithm, but when treated as an evidentiary or decision-support layer linking sensing, interpretation, and action.

Across its different incarnations, SmokeSeer therefore marks a research trajectory rather than a single method: from heuristic event detection and participatory documentation, through real-time wildfire alerting, to explicit multimodal decomposition of smoky scenes for reconstruction. The continuity is not architectural. It is operational: each system seeks to make smoke legible quickly enough, reliably enough, and in a form useful enough for intervention.

Source: https://www.emergentmind.com/topics/smokeseer