Polar Mesoscale Cyclones (PMCs) are mesoscale cyclonic weather systems that frequently occur over open waters or sea-ice edges in regions poleward of the main polar front zones (Condron et al., 2006; Rasmussen and Turner, 2003). They are often identified on satellite imagery by comma-shaped or spiral cloud patterns. PMCs occur in all seasons but are most active in winter, with a lifespan of approximately one day and horizontal scales of less than 1000 km (Harold et al., 1999). The most intense subset of these cyclonic systems, termed Polar Lows (PLs), are major hazardous weather phenomena in polar regions, characterized by average maximum wind speeds exceeding 15 m s⁻¹ and extreme values surpassing 30 m s⁻¹. They can generate significant wave heights over 11 m (Rojo et al., 2019), posing severe threats to human activities and maritime safety in high-latitude regions (Harrold and Browning, 1969; Orimolade et al., 2016). Additionally, PLs induce rapid sea-ice changes and intensify ocean-deep convection through dynamic and thermodynamic effects, producing complex regional climatic impacts (Clancy et al., 2022; Condron and Renfrew, 2013; Parkinson and Comiso, 2013). The Nordic Seas (encompassing the Greenland, Norwegian, and Barents Seas) form a critical oceanic gateway between the Arctic and Atlantic Oceans. This region is a primary convergence zone for Arctic and Atlantic water masses and plays a key role in global ocean circulation and climate (Smedsrud et al., 2022). The complex meteorological and oceanographic conditions in this area make it the most frequent PLs occurrence region (Stoll, 2022). Consequently, research on mesoscale cyclones in the Nordic Seas is critical for improving Arctic maritime safety and understanding regional climate change impacts.

Cyclonic cloud morphology and surface wind fields derived from remote sensing data serve as the primary criteria for distinguishing and categorizing PMCs and PLs (Rasmussen and Turner, 2003). The former can be manually identified through visible or infrared imageries from passive radiometers (e.g., Fig. 1), while the latter can be estimated using scatterometer or microwave data. While PLs exhibit higher destructive potential and detection feasibility compared to broader PMCs, current dataset development efforts have predominantly targeted PLs, leaving PMCs relatively underrepresented in existing observational records. Blechschmidt (2008) combined Advanced Very High-Resolution Radiometer (AVHRR) infrared imagery (Kalluri et al., 2021) with wind speed data derived from the Hamburg Ocean Atmosphere Parameters and Fluxes from Satellite Data (HOAPS, Andersson et al., 2010) to manually identify 90 PLs occurring in the Nordic Seas between 2004 and 2005. Noer et al. (2011) utilized AVHRR infrared imagery, Advanced Scatterometer (ASCAT), and Quick Scatterometer (QUIKSCAT) wind data to detect 121 PLs in the Nordic Seas over a decade (2000–2009). Smirnova et al. (2015) identified 637 PLs between 1995 and 2009 using Special Sensor Microwave/Imager (SSM/I) data for atmospheric total water vapor (TWV) content fields, near-surface wind speed fields, and AVHRR infrared imagery. Golubkin et al. (2021) employed Moderate Resolution Imaging Spectroradiometer (MODIS) and ASCAT data to identify PLs over the North Atlantic, compiling a catalog of 131 PLs between 2015 and 2017. In all PL lists derived from remote sensing data, the Rojo list (Rojo et al., 2015, 2019) is currently the longest temporally spanning remote sensing-derived PLs track dataset, providing tracks of 420 PLs occurring in the Nordic Seas from 1999 to 2019. It includes basic information such as cyclone location, size, type, development stage, and maximum 10 m wind speed. The manually tracked datasets described above have provided valuable PLs information, contributing to ongoing research efforts. However, the unique high-latitude geography of polar regions creates significant observational challenges. Polar-orbiting satellites typically observe these regions at intervals ranging from tens of minutes to several hours, resulting in temporal gaps that make it difficult for manual tracking datasets to capture complete cyclone life cycles. Additionally, some PLs forming near sea-ice edges may exhibit distinct cyclonic cloud features exclusively during their transition over moisture-rich open waters (Bromwich, 1991), implying that remote sensing datasets could potentially miss capturing the initial developmental stages of such PLs. Consequently, while the Rojo list provides developmental pattern annotations for individual PLs, the objectivity and quantitative reliability of these annotations remain constrained by the inherent limitations of remote sensing in achieving comprehensive characterization of PL evolution throughout their complete lifecycle. Furthermore, the occurrence of polar night, coupled with low contrast between sea-ice/snow surfaces and overlying clouds, further limits the detection capabilities of remote sensing (particularly visible-band remote sensing) methods for PLs.

With the improved resolution of reanalysis datasets, their ability to characterize PLs has progressively advanced (Laffineur et al., 2014; Smirnova and Golubkin, 2017), making them an increasingly critical data source for constructing PLs track datasets. Researchers have employed various combinations of identification criteria to detect PLs. For instance, Zappa et al. (2014) utilized the difference between 500 hPa temperature and near-surface temperature to represent cold air outbreak characteristics during PLs formation, while utilizing maximum near-surface wind speed to indicate PLs intensity, and 850 hPa relative vorticity to capture their cyclonic properties. Subsequent studies adopted or adapted these criteria (Stoll et al., 2018; Terpstra et al., 2016; Yanase et al., 2016). Building on the fifth-generation European Centre for Medium-Range Weather Forecasts Reanalysis (ERA5, Hersbach et al., 2020), Stoll (2022) established a four-criteria linear-based combination defining PLs as intense mesoscale cyclones forming within polar oceanic air masses northward of the polar front. This approach successfully reproduced 60 %–80 % of PLs from five manual PL lists, validating ERA5's robust capability in PLs representation. However, ERA5 significantly underestimates near-surface wind speeds within PL-affected regions (Gurvich et al., 2022; Haakenstad et al., 2021), attributed in part to insufficient representation of transient wind variability, surface divergence, and unresolved mesoscale features (Belmonte Rivas and Stoffelen, 2019). This limits its ability to objectively capture PLs' high-wind characteristics, thereby introducing notable limitations.

In summary, remote sensing and reanalysis datasets provide complementary perspectives on PLs' characteristics, with the former capturing cloud morphology and the latter resolving meteorological field distributions, highlighting their respective advantages. This complementary nature motivates the integration of both data types to construct more comprehensive PL tracking datasets – a key objective of this study. Furthermore, existing datasets primarily focus on PLs, while weaker PMCs that share similar cyclonic cloud features and environmental conditions lack comprehensive publicly available track datasets. This disparity likely stems from the fact that PMCs generally have smaller average intensities, shorter lifespans, and smaller scales compared to PLs, making them more difficult to detect. Although some researchers have proposed PMC track datasets using either remote sensing (Verezemskaya et al., 2017) or reanalysis data (Michel et al., 2018; Pezza et al., 2016; Watanabe et al., 2016), these approaches face significant limitations. Remote sensing-based datasets often have inadequate temporal coverage or lack critical near-surface wind speed records (Condron et al., 2006), while reanalysis-based datasets encounter challenges in developing effective identification criteria without remote sensing validation. As a result, no universally accepted PMC identification standards currently exist (Michel et al., 2018). Notably, while PLs have been well-documented in relation to large-scale circulation patterns such as the North Atlantic Oscillation (Claud et al., 2007) and Scandinavian blocking (Mallet et al., 2013), the climatic impacts of PMCs remain insufficiently investigated (Michel et al., 2018). These knowledge gaps highlight the critical need to establish a more comprehensive tracking dataset capable of capturing PMCs throughout their lifecycle. Such a dataset would enable the complete characterization of these weaker polar mesoscale systems, representing another key motivation for this study.

Based on the above analysis, this study aims to comprehensively integrate the advantages of reanalysis datasets in characterizing the dynamical and thermodynamic structures of polar mesoscale weather systems and remote sensing data in capturing cloud morphology to establish a long-term PMCs (hereafter, “PMCs” when used alone include “PLs”) track dataset in the Nordic Seas encompassing the extended winter seasons (November–April) between 2001 and 2024. This dataset will contain the tracks of the PMCs in reanalysis fields and remote sensing imagery, as well as multi-dimensional attributes such as intensity, cloud morphology, and near-surface wind features. The objective is to provide a long-term, multi-attribute catalog of PMCs, offering reliable data support for atmospheric and oceanic research in the Nordic Seas.

To establish a more comprehensive cyclone track dataset in the Nordic Seas, we first utilize ERA5 reanalysis data which exploits the evolving global observing system to obtain all vortex tracks. In this process, a lower vorticity maxima criterion is applied to extract vorticity perturbations within the reanalysis data. Subsequently, vortex tracks and their merging and splitting processes are identified based on spatial and boundary changes of vortices across consecutive time steps. For each vortex with available AVHRR data, we generate Vortex-Centered Infrared (VCI, mentioned in the following text) images to identify corresponding cyclonic cloud features with a cyclone-detection deep-learning model. Finally, near-surface wind fields derived from QuikSCAT/ASCAT are matched to characterize cyclones' core wind speed. The algorithm workflow is outlined in Fig. 2, with methodological details provided in subsequent subsections.

Our analysis began by applying a vortex tracking algorithm to reanalysis data, which identified 59 975 vortex tracks. Validation against VCI imagery confirmed 1110 cyclone-related vortex tracks, encompassing 16 001 distinct cyclone cloud features. Subsequent analysis of surface wind speed characteristics revealed 4472 instances with measurable wind patterns, among which 794 tracks exhibited maximum wind speeds exceeding the 15 m s⁻¹ threshold. These validated 1110 vortex tracks, along with their corresponding remote-sensing images, form the IMPMCT track dataset. The accuracy of IMPMCT was rigorously evaluated through comprehensive comparisons with existing track datasets derived from manual identification and reanalysis products.

First, to validate the accuracy of the vortex track datasets obtained from the vortices tracking algorithm, they are compared with the manually identified PL lists published by Noer et al. (2011), Rojo et al. (2015, 2019), and the objectively derived PL track datasets from reanalysis data by Stoll (2022). All reference datasets are spatially and temporally co-located with our derived tracks, retaining only those persisting for ≥ 3 h. We applied the following matching criteria: a vortex track is considered matched with a PL track if more than 50 % of temporally coincident track points (within ±1 h) fall within a 150 km radius (applying an 80 % threshold for Stoll's dataset). To avoid spurious matches of short-lived spurious tracks, only vortex tracks with lifespans exceeding 60 % of the corresponding reference PL track's duration were included. A single vortex track was permitted to match multiple PL tracks from reference datasets, provided that these PL tracks did not overlap temporally and each was uniquely paired with its nearest vortex track. As presented in Table 2, the validation results demonstrate strong agreement with Stoll's dataset, confirming the robustness of our vortex tracking algorithm. Moreover we achieve higher matching rates with manual PL lists by using lower vortex identification thresholds, which further underscores the improved capability of ERA5 reanalysis data in representing PL characteristics. Additional validation using tracks from the sensitivity experiment (Sect. 3.1) revealed a critical insight: vortex tracks derived under lenient thresholds consistently produced higher matching rates when compared against established PL datasets (Table S2). This suggests that some PLs exhibit weaker vorticity signals in the lower atmosphere, highlighting intrinsic intensity diversity that stricter thresholds may fail to capture.

To further investigate the mismatches between reanalysis-derived tracks and existing PL datasets, we conducted a nearest-point matching analysis (Table 2). A match was considered successful when a PL center from any reference dataset had at least one temporally coincident vortex center within a 120 km radius (60 km for the Stoll dataset). Track-level mismatches were found to originate primarily from these point-level discrepancies. The variation across datasets can be largely attributed to methodological differences: while the Noer list derives from numerically modeled and AVHRR-assimilated hourly positions (typical of operational forecasting systems), the Rojo list relies on direct AVHRR identification at irregular temporal intervals, leading to greater deviation from ERA5 representations. Furthermore, the Rojo compilation includes numerous secondary PL centers, which are features inherently less resolved by reanalysis data (Stoll, 2022), whereas Noer focuses primarily on dominant PLs of operational significance. This distinction is clearly reflected in our results: major PL centers (n = 2527) showed an 80 % matching rate, compared to only 54 % for secondary centers (n = 1115), thereby lowering the overall match rate for the Rojo dataset.

For the Stoll dataset, we also computed a vortex matching rate (Table 2), defined as the proportion of Stoll centers falling within the spatial extent of the nearest co-temporal vortex. This measure helps account for positional discrepancies caused by misalignment of vorticity peaks, which appear to stem from differences in smoothing techniques (see Fig. S1). Our algorithm applies stronger uniform smoothing compared to Stoll's approach, explaining why more lenient identification thresholds improve track matching with Stoll's dataset. This finding offers valuable insight for algorithm application: although the algorithm is not highly sensitive to the specific input vorticity fields, provided their grid spacing is sufficient to capture mesoscale vortices, the choice of smoothing method significantly influences identification outcomes, alongside the threshold parameters examined in the sensitivity experiments (Sect. 3.1.1). The smoothing strategy should be tailored to the assimilation noise and effective resolution of the input vorticity field. For example, Gaussian smoothing may be better suited for model data with lower noise levels, as it better preserves the spatial coherence of vortex cores.

After excluding vortex tracks with over 60 % land coverage (resulting in an approximately 20 % reduction), 47 167 tracks remained eligible for AVHRR matching. The matching procedure required: (1) complete spatial coverage within a 200 km radius for individual vortex points, and (2) at least two temporally matched points within ±3 h of peak vorticity, along with a minimum of six matched points over the track's lifetime. Figure 13 presents the matching statistics for the winter months (November to April): on average, 43 % of points and 61 % of tracks were successfully matched. However, only about 3 % of the matched tracks were ultimately incorporated into the IMPMCT dataset. This low inclusion rate can be attributed to several factors: frequent cloud obstruction, limitations in cloud–ice contrast, temporal resolution constraints, and inherent detection methodology (e.g., the higher inclusion rate in 2001 reflects meticulous manual identification, whereas the lower rate in 2023 resulted from incidental post-publication discoveries). Importantly, the proportion of cyclones in IMPMCT likely underestimates the true prevalence of polar mesoscale cyclones (PMCs), as many systems with low cloud cover lack discernible vortex structures. In cases where AVHRR data are unavailable, an alternative approach using hourly wind field data calibrated with scatterometer measurements may provide a more robust method for validating ERA5-derived vortex tracks (Furevik et al., 2015).

We further assess the reliability of vortex properties in IMPMCT by comparing three key parameters (850 hPa relative vorticity, SLP minima, and vortex equivalent diameter), with the corresponding values from Stoll's dataset, in addition to evaluating the spatial distance between vortex centers. From this comparison, 638 matched tracks were identified between IMPMCT and Stoll's dataset. As shown in Fig. 14a, among the matched tracks, 90 % of vortex points remain within 50 km of each other at the same time step. The mean absolute differences of the three vortex properties at these proximate track points remain small: 1.11 × 10⁻⁵ s⁻¹ for relative vorticity, 0.43 hPa for sea-level pressure, and 22.79 km for vortex equivalent diameter. Furthermore, these property discrepancies exhibit a positive correlation with separation distance, suggesting that differences between IMPMCT and Stoll's tracks primarily arise from their respective identification thresholds.

To demonstrate that these discrepancies reflect divergent tracking methodologies rather than detection errors, we calculated the standard deviation of each vortex property over three consecutive time steps for every track and then averaged these values across each track. Low amplitude in these local variations implies consistent feature identification by a given method. Figure 14b–d present the track-averaged local standard deviations of the three properties for both datasets. Importantly, the magnitudes of these short-term variabilities are generally comparable between IMPMCT and Stoll's tracks. This consistency indicates that the increasing property differences at larger separations stem from intrinsic peak misalignments due to differing detection logics, rather than fundamental errors in either tracking approach. In fact, the IMPMCT dataset often exhibits slightly smoother variability, which is consistent with its specific algorithmic configuration.

IMPMCT uses hourly-resolution vortex tracks from reanalysis data as a basis for cyclone tracks. The correspondence between vortex and cyclone tracks is established exclusively via continuous spatiotemporal matching of their respective centers. To ensure the accuracy of this correspondence, we perform subjective validation to confirm that each cyclone track does not incorporate irrelevant cyclonic processes. Notably, while the average matching distance between vortex and cyclone tracks is constrained within 150 km, approximately 95 % of track pairs have average matching distances below 100 km (as shown in Fig. 15), demonstrating strong consistency between cyclone and vortex tracks.

The cyclone properties in IMPMCT include cyclone scales and maximum core near-surface wind speeds. These properties are validated through comparison with the Rojo list. For scale validation, we compare the diameter from the Rojo list with the approximate cyclone scale in the IMPMCT dataset (calculated as the average of cyclone width and length). We matched cyclone tracks between IMPMCT and Rojo list based on the following criteria: the nearest cyclone centers are matched if their distance is less than 120 km and their overpass times fell within 60 min of each other. A cyclone track pair was deemed matched if more than 50 % of the points in a Rojo track were matched. Using this approach, 1424 cyclone centers from the Rojo list (corresponding to 139 distinct tracks) were matched to tracks in IMPMCT. It is worth noting that although the maximum permitted matching distance was 120 km, the 90th percentile of all actual matching distances was only 56 km. This indicates that cyclone center identification remained consistent even when exact temporal alignment was not achieved.

Comparisons of cyclone cloud scale and maximum wind speeds between the matched time periods are shown in Fig. 16. When cyclone center identification errors are small, the discrepancies in diameter relative to the Rojo list arise not only from methodological differences in measurement, but also significantly from subjective interpretation. The frequent presence of frontal cloud bands associated with cyclones makes consistent measurement of the long axis highly subjective. Moreover, when a cyclone is adjacent to other cloud systems, its boundaries often become ambiguous, leading to variability in extent estimation. Therefore, a standard deviation of up to 120 km in diameter is still considered acceptable. Furthermore, as the dataset includes corresponding remote sensing images, users can readily examine the visual context of each cyclone and adjust the properties according to their specific research needs.

To statistically evaluate the agreement between IMPMCT and the reference datasets (Stoll, 2022, and Rojo et al. 2015, 2019), we applied Bland–Altman analysis (Bland and Altman, 1999). This method quantifies the agreement between two measurement techniques by estimating the mean difference (bias) and the 95 % limits of agreement (LoA), defined as the mean difference ±1.96 standard deviations of the differences. A summary of the Bland–Altman results for key vortex and cyclone properties is provided in Table 3, while the corresponding plots of differences versus averages are included in Supplement Fig. S2. As indicated in Table 3, vortex properties derived from ERA5 reanalysis show a small systematic bias relative to the other datasets, which is likely due to differences in computational algorithms or processing workflows. Importantly, the Bland–Altman results demonstrate strong agreement between the datasets: approximately 94 % of the differences for each property fall within the respective 95 % limits of agreement (final column of Table 3), supporting the overall consistency and reliability of IMPMCT

For most newly identified mesoscale cyclones not documented in existing PL databases, direct validation can be performed by applying objectively derived identification thresholds from previous studies to independently verify three essential characteristics: polar origin, mesoscale size, and cyclonic intensity:

Polar-front criterion. As PMCs are defined as mesoscale cyclones forming north of the polar front (Rasmussen and Turner, 2003), we employ two indicators to distinguish polar air masses from extratropical air masses: tropopause potential temperature (θtrop) and the maximum poleward value of 200 hPa wind speed (U200,p). For each cyclone, we compute the track-averaged θtrop averaged within a 250 km radius of the cyclone center and the track-averaged U200,p within a longitudinal band of ±1.0° great-circle distance. Following Stoll (2022), θtrop < 300.8 K is used to identify polar air mass origin. This threshold effectively distinguishes PLs from extratropical cyclones, retaining 76 % of systems across subjective archives while capturing 90 % of known PLs. Han and Ullrich (2025) employed U200,p < 25 m s⁻¹ to position PLs north of the polar jet, achieving an approximately 80 % hit rate for PL classification with a miss rate of only 11.9 %.

The comprehensiveness of the dataset is constrained by the cyclone representation capabilities of ERA5 reanalysis and the availability of remote sensing data. Since the number of in-orbit satellites carrying the AVHRR sensor peaked around 2013, the IMPMCT track dataset includes the highest number of tracks during this period. Additionally, due to the use of more lenient identification thresholds, IMPMCT tracks typically include longer life compared to the Stoll dataset. The extended portions of these tracks may include: weak vorticity periods during the early/late stages of cyclone development or the vortices pass over land/sea-ice, or redevelopment processes of vortices after interacting with blocked extropical cyclones or frontal zones. If users require only the core development phases of tracks, they should select segments based on vortex properties or cyclone images that represent the system's core development. The dataset also includes some tracks with high vorticity at their start/end points, which may arise from splitting/merging events or jumps of the vortex center position during tracking. It is noteworthy that while this study demonstrates ERA5 reanalysis data's enhanced capability in capturing PMCs and PLs, it does not reflect ERA5's predictive skill for such systems. This predictive capability should be evaluated by testing ERA5 background states in characterizing PLs/PMCs, thereby isolating the influence of real-time assimilated data – particularly scatterometer measurements (Furevik et al., 2015).

The dataset does not explicitly distinguish between PMCs and PLs due to the time-sparse wind speed data, particularly when the cyclone's wind speed at a given time step falls below the 15 m s⁻¹ threshold. In such cases, it is difficult to determine whether the cyclone is a PMC or merely in a weaker phase of a PL. A more reliable validation method may be provided by the hourly bias-corrected sea surface wind product (WIND_GLO_PHY_L4_MY_012_006) from the E.U. Copernicus Marine Service Information (Copernicus Marine Service, 2024). Such product systematically corrects ECMWF ERA5 model fields using scatterometer observations to reduce persistent biases and includes uncertainty estimates. Furthermore, the L3 scatterometer products available through CMEMS, which contain the spatial derivatives of the wind vector fields (vorticity and divergence), offer a more direct characterization of the dynamical core of mesoscale systems. These observed fields hold significant potential for refining objective identification criteria, moving beyond a reliance on wind speed thresholds alone. Due to the low resolution of AVHRR infrared images at scan edges, a significant portion of VCI images appear blurred. However, these images are retained as long as cyclonic features remain recognizable, prioritizing the preservation of high temporal resolution for cyclone track records. Additionally, while the YOLOv8-obb-pose model facilitates detection and feature extraction of cyclonic cloud characteristics in VCI images, the process still involves subjective steps to ensure continuity in cyclone features (e.g., size, type, and position). This implies that objective methods for constructing multi-parameter PMC track datasets remain underdeveloped. Consequently, cyclone-evolution-aware deep-learning tracking algorithms could further enhance the efficiency of track construction.

The IMPMCT dataset described in this paper is freely accessible on Zenodo via the following link: 10.5281/zenodo.17142448 (Fang and Ding, 2025), accompanied by comprehensive documentation. The YOLOv8-obb-pose model weights for detecting PMC cloud features, along with the test set, are open-sourced and available at: 10.5281/zenodo.15119534 (Fang, 2025b). All code is developed in Python and stored at: 10.5281/zenodo.17498230 (Fang, 2025a).

The Integrated Multi-source Polar Mesoscale Cyclone Track (IMPMCT) dataset represents a major advancement in the study of polar mesoscale cyclonic systems. By integrating ERA5 reanalysis, AVHRR infrared imagery, and QuikSCAT/ASCAT wind data, this dataset provides a comprehensive record of 1110 vortex tracks, 16 001 cyclonic cloud features, and 4472 wind speed observations across the Nordic Seas (2001–2024). This integrated approach overcomes key limitations of previous single-source datasets by enhancing detection sensitivity for weaker polar mesoscale cyclones (PMCs), capturing complete lifecycle evolution from genesis to dissipation, and providing simultaneous cloud morphology and wind fields observations. Rigorous validation against established datasets (Stoll, 2022 and Rojo et al., 2019) confirms IMPMCT's accuracy, demonstrating 90 % spatial consistency with track points cyclone centers alignments within 50 km (60 km for cyclone centers) and minimal parameter discrepancies including a 1.11 × 10⁻⁵ s⁻¹ mean absolute difference in relative vorticity and 0.43 hPa mean absolute difference in sea-level pressure.

The IMPMCT dataset serves as a critical benchmark for evaluating high-latitude numerical weather prediction model performance, while simultaneously functioning as a unique case library for comparative studies of PLs and PMCs concerning their formation mechanisms, intensity thresholds, and sea-ice interaction dynamics. Furthermore, it constitutes an essential resource for enhancing polar maritime hazard forecasting. The repository of cyclone cloud morphology facilitates automated identification of model-undetected systems. This is enabled by advanced deep learning frameworks, enabling systematic evaluation of model representation fidelity for PLs/PMCs.