AusSmoke: Dataset & Smoke Vortex Event

• AusSmoke is a dual-use term referring to both an Australian smoke segmentation dataset and the 2019–20 wildfire-induced stratospheric smoke vortex event.
• The segmentation dataset overcomes regional data scarcity by providing a fully-labelled and geographically diverse benchmark that enhances model generalization.
• The atmospheric study reveals that wildfire smoke can form long-lived, anticyclonic vortices that alter stratospheric dynamics and radiative heating.

AusSmoke denotes two distinct objects in recent arXiv literature: a smoke segmentation dataset collected from Australia, and, in atmospheric-dynamics work, the 2019–20 Australian stratospheric smoke-vortex episode. In computer vision, AusSmoke is introduced alongside MultiNatSmoke as a fully-labelled, geographically diverse benchmark intended to address data scarcity in Australia and to improve generalization across diverse geographical contexts (Li et al., 26 Apr 2026). In atmospheric science, the “AusSmoke” event refers to the smoke-charged vortices generated by the 2019–20 Australian wildfires, which self-organized in the stratosphere into compact, long-lived anticyclonic structures that rose to high altitude and circled the globe (Lestrelin et al., 2020).

1. Terminological scope

In the provided arXiv literature, “AusSmoke” is used in two distinct senses. One is a dataset name in the paper "AusSmoke meets MultiNatSmoke: a fully-labelled diverse smoke segmentation dataset" (Li et al., 26 Apr 2026). The other is an event label in the study "Smoke-charged vortices in the stratosphere generated by wildfires and their behaviour in both hemispheres : comparing Australia 2020 to Canada 2017" (Lestrelin et al., 2020).

Usage	Description	Source
AusSmoke	A new smoke segmentation dataset collected from Australia	(Li et al., 26 Apr 2026)
“AusSmoke” event	The austral summer 2019–20 Australian smoke-vortex episode	(Lestrelin et al., 2020)

This dual usage is important because the two literatures operate at different levels of analysis. The dataset work addresses AI-enabled camera-based smoke detection and smoke segmentation. The atmospheric work addresses pyroconvection, stratospheric transport, potential vorticity, and radiative-dynamical maintenance of smoke-charged vortices. A plausible implication is that the shared term reflects a common empirical anchor—Australian wildfire smoke—while serving distinct research programs.

2. AusSmoke as a smoke segmentation dataset

The dataset paper presents AusSmoke as “a new smoke segmentation dataset collected from Australia to address the data scarcity in this region” (Li et al., 26 Apr 2026). It further introduces “a MultiNational geographically diverse and substantially larger fully-labelled benchmark, called MultiNatSmoke, that consolidates publicly available international datasets with the newly collected Australian imagery, expanding the scale by an order of magnitude over previous collections.” The same abstract states that smoke segmentation models are benchmarked, “demonstrating improved performance and enhanced generalization across diverse geographical contexts,” and notes that “the project is available at Github” (Li et al., 26 Apr 2026).

The motivation is explicitly tied to limitations in prior resources. Existing wildfire smoke segmentation datasets are described as “limited in scale, geographically constrained, and often rely on synthetic imagery, which hinders effective training and generalization” (Li et al., 26 Apr 2026). Within that framing, AusSmoke functions as an Australia-specific data acquisition effort, and MultiNatSmoke functions as a consolidation benchmark spanning multiple countries.

The evidentiary scope of the available description is, however, sharply delimited. The provided material explicitly states that it does not include “the AusSmoke dataset’s actual collection details, annotation protocols, statistics, benchmarks, or preprocessing steps—only the paper’s bibliographic entry and unrelated template material” (Li et al., 26 Apr 2026). It also states that, without the sections describing “Dataset Scope and Collection,” “Annotation Protocol,” “Dataset Statistics,” “Comparison to Existing Datasets,” “Benchmarking and Metrics,” and “Data Augmentation and Preprocessing,” a detailed technical overview cannot be generated. Accordingly, any stronger claims about label taxonomy, class balance, image resolution, split construction, or benchmark metrics would exceed the supplied evidence.

3. Synthetic-smoke antecedents and domain adaptation context

The dataset paper’s critique of prior resources—especially their reliance on synthetic imagery—has a clear antecedent in earlier smoke-detection work. "Deep Domain Adaptation Based Video Smoke Detection using Synthetic Smoke Images" uses Blender-Python to generate synthetic smoke images via a numerical fluid–dynamics solver, a density field, and a renderer, with randomized initial flow, wind direction and magnitude, lighting, and background image (Xu et al., 2017). That pipeline generated 30 000 synthetic smoke images spanning “a wide variation in the smoke shape, background and lighting conditions,” and each synthetic image carried two labels: $y_f \in \{\text{smoke}, \text{non-smoke}\}$ and $y_a \in \{\text{synthetic}, \text{real}\}$.

The model family in that work uses the convolutional backbone of AlexNet, a classification head, a domain head with a Gradient Reversal Layer, an optional adaptation layer, and a correlation alignment module. The overall loss is

$$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$

with

$$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$

$$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$

and

$L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$

In experiments, the best model, combining GRL, adaptation, and CORAL, achieved $CD=0.947$, $ED=0.045$, and $MD=0.062$ on a real-image test set of 1 000 images, with derived accuracy $94.7\%$, precision $y_a \in \{\text{synthetic}, \text{real}\}$0, recall $y_a \in \{\text{synthetic}, \text{real}\}$1, and $y_a \in \{\text{synthetic}, \text{real}\}$2 (Xu et al., 2017).

This earlier line of work is directly relevant to AusSmoke because it formalizes the dataset-bias problem that AusSmoke is meant to alleviate. The synthetic-smoke paper states that the appearance gap between synthetic and real smoke images “degrades significantly the performance of the trained model on the test set composed fully of real images,” and that domain adaptation is required to confuse the distributions of features extracted from synthetic and real smoke images (Xu et al., 2017). This suggests that a real Australian smoke dataset is not merely additive in scale; it is also corrective with respect to geographical and domain bias.

4. AusSmoke as the 2019–20 Australian stratospheric smoke event

In atmospheric science, the “AusSmoke” event refers to the smoke injected into the stratosphere during the 2019–20 Australian “Black Summer” fires (Lestrelin et al., 2020). Record-breaking fires produced PyroCb towers that penetrated the tropopause through intense updrafts of buoyant, smoke-laden air. Satellite lidar and limb-sounder observations show that these PyroCb events injected a pulse of smoke and black carbon between roughly $y_a \in \{\text{synthetic}, \text{real}\}$3 and $y_a \in \{\text{synthetic}, \text{real}\}$4 on 30–31 December 2019, with injections continuing intermittently through early January 2020. Optical depths in the nascent stratospheric plume exceeded $y_a \in \{\text{synthetic}, \text{real}\}$5–$y_a \in \{\text{synthetic}, \text{real}\}$6 at $y_a \in \{\text{synthetic}, \text{real}\}$7, implying total smoke masses comparable to moderate volcanic injections, $y_a \in \{\text{synthetic}, \text{real}\}$8–$y_a \in \{\text{synthetic}, \text{real}\}$9.

Once in the stratosphere, the black-carbon-rich plume experienced secondary radiative heating. Absorption of solar radiation by smoke provided an internal heating rate that enhanced buoyancy and allowed portions of the plume to rise an additional $$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$0–$$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$1 over the next $$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$2–$$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$3 weeks. In ECMWF forecasts, black carbon was missing, but the study infers from assimilation-increment patterns “an effective net heating on the order of $$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$4–$$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$5 within the smoke-charged vortices” (Lestrelin et al., 2020).

The event is therefore not described as a passive aerosol veil. The supplied analysis instead frames it as an internally heated, dynamically coherent stratospheric structure produced by pyroconvection and sustained, in analysis fields, by data assimilation acting against radiative damping and background descent.

5. Vortex organization and dynamical description

Within days of injection, a compact anticyclonic “bubble” formed in the stratospheric plume (Lestrelin et al., 2020). These bubbles are described as bubbles of low absolute potential vorticity and low ozone that trace coherent cores of smoke-laden air remaining internally buoyant. Their typical length scales are $$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$6–$$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$7 horizontally and $$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$8–$$L = \alpha_{\text{label}}\,L_s + \beta_{\text{domain}}\,(\lambda\,L_d + \gamma_{\text{coral}}\,L_{\text{coral}}),$$9 vertically, giving the pancake aspect ratio

$$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$0

The same study places them on the border of quasi-geostrophic motion, with

$$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$1

and

$$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$2

When plotted in stretched coordinates $$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$3, the vortices are nearly spherical.

The paper gives the absolute Ertel PV in hybrid coordinates as

$$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$4

and the Lait PV as

$$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$5

with $$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$6 and $$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$7 in the Southern Hemisphere or $$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$8 in the Northern Hemisphere. Radiative damping of the vortex temperature anomaly is represented as

$$L_s = -\frac{1}{N}\sum_{i=1}^{n}\log \mathrm{softmax}(a_i),$$9

with $$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$0–$$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$1 days.

The longest-lived Australian vortex, “Koobor,” persisted from early January to late February, completed nearly two full circumnavigations of the Southern Hemisphere, and climbed from $$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$2 up to $$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$3 (Lestrelin et al., 2020). Smaller sister vortices rose more slowly to $$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$4–$$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$5 over a similar lifespan. The paper’s interpretation is that these vortices constitute a new dynamical regime in which wildfire smoke, once lofted above the tropopause, self-organizes into long-lived anticyclonic eddies that rise against the mean stratospheric circulation through internal radiative heating.

6. Tracking, maintenance, comparison, and implications

The Australian vortices are tracked using both observations and reanalysis (Lestrelin et al., 2020). CALIOP lidar at $$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$6 measures total attenuated backscatter, from which the scattering ratio is obtained by dividing by molecular backscatter; aerosol-only layers are separated via the Level 2 aerosol product. Typical along-track resolution is $$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$7 horizontally by $$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$8–$$L_d = \frac{1}{2N}\sum \max(0,\,1-o(l_i=k)\,t_{ik})^2,$$9 vertically. In ERA5 $L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$0 reanalysis at full model levels and 3-hourly resolution, the analysis scans for local minima of $L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$1 and collocated negative ozone anomalies within $L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$2 latitude, $L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$3 longitude, and $L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$4 in $L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$5, then constructs trajectories at 6 h intervals by following these extrema.

A central result is that the vortices are maintained in analysis but not in free forecast. Once identified in the analysis, each vortex is present in the free forecast but decays in $L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$6 unless it is continually re-enforced by data assimilation. The 12 h “analysis minus first-guess” increments of temperature, vorticity, and ozone show a dipolar or tripolar heating signature that counteracts longwave cooling, rebuilds the anticyclonic PV/ozone hole, and supplies ascent against the Brewer–Dobson descent. The study writes the assimilation-induced translation tendencies in a background shear $L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$7 as

$L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$8

with

$L_{\text{coral}} = \frac{1}{4d^2}\|C_s-C_t\|_F^2.$9

and tilt ratio

$CD=0.947$0

The paper compares Australia 2020 with Canada 2017. “Koobor” reached $CD=0.947$1 ($CD=0.947$2), whereas Canada’s 2017 vortices peaked at $CD=0.947$3 ($CD=0.947$4). The Southern Hemisphere Australian vortices circled the globe two to three times at mid-latitudes; the Northern Hemisphere Canadian plume formed a parent vortex on 12 August 2017 near $CD=0.947$5, rose to $CD=0.947$6 by 25 August, elongated under zonal-jet shear, and split into three offsprings over Europe by 30 August. In both hemispheres, the free-running model alone fails to sustain ascent or vorticity.

The broader implications are stated in terms of circulation, climate, and ozone. These vortices act as “bubbles” of tropospheric air carrying low PV and enhanced black carbon, oppose the descent branch of the Brewer–Dobson circulation, and inject smoke deep into the mid-stratosphere at $CD=0.947$7–$CD=0.947$8, extending aerosol lifetimes from months to a year or more (Lestrelin et al., 2020). The compact plumes achieve a higher mean altitude than a dilute layer, boosting their warming effect, with radiative forcing estimates comparable to moderate volcanic eruptions. Each vortex encloses a core of reduced ozone; localized “mini-holes” form, but the net chemical impact remains to be quantified. A common misconception would be to treat the phenomenon as merely advected smoke. The supplied evidence instead identifies coherent mesoscale vortices, internal radiative heating, and continual re-enforcement by data assimilation as essential to their observed evolution.