1. Introduction
Ice fog forms in the Arctic when cold stagnant air pools in low-lying regions. Further radiative cooling leads to the relative humidity increasing and possibly reaching saturation. The American Meteorological Society (AMS) and the World Meteorological Organization (WMO) definitions of ice fog both state that ice fog most frequently occurs at temperatures colder than −30°C and that anthropogenic sources of water vapor contribute to ice fog formation (American Meteorological Society 2022; WMO 2017). Especially in urban environments, pollution particles and water vapor from anthropogenic activities can substantially enhance ice fog (Oliver and Oliver 1949; Appleman 1953; Benson 1965). Once ice fog has formed, already low levels of incoming solar radiation at the surface will be further reduced by reflection, thus hampering vertical mixing (e.g., Oliphant et al. 2021). The stratified air is extremely stable, which can lead to temperature inversions of tens of degrees (Malingowski et al. 2014). Ice fog typically dissipates when vertical mixing resumes due to changes in the weather situation. While ice fog has been shown to occur at warmer temperatures (Gultepe et al. 2015, 2017), the majority of cases in Arctic noncoastal locations occur in cold (≤−30°C) conditions as stated in the AMS and WMO definitions. Our study is based on the AMS and WMO definitions of low-temperature ice fog, as these are well aligned with observed ice fog in the Fairbanks, Alaska, region.
Several locations along the Chena River, which flows through downtown Fairbanks, are ice free throughout the winter because of human activities, thus adding vapor to the system and contributing to ice fog conditions. Vehicular traffic, home heating, aircraft emissions, and industrial processes cause substantial near-surface pollution and vapor emissions that can be trapped below the strong temperature inversions and serve as nuclei for ice particles, further contributing to ice fog formation and persistence. In addition, several large coal-fired power plants generate electricity and heat through boilers in the Fairbanks region. Power plant plumes produce large clouds visible on satellite images during cold conditions. Plume particles can slowly settle toward the surface adding nuclei and moisture to lower levels. Due to the small sizes of ice particles in ice fog (American Meteorological Society 2022; WMO 2017), their fall velocities are very slow. Direct and indirect anthropogenic emissions of vapor and pollution can increase ice fog longevity. With very little insolation, ice fog often persists until the synoptic conditions change in the region.
Ice fog can be very dangerous to air and vehicular traffic. Air traffic is hampered by ice fog conditions due to very low visibility as well as due to potential turbulence and icing conditions at the temperature inversion (Gultepe et al. 2014, 2015; Leung et al. 2020). Vehicular traffic is also negatively affected by low visibility, and accidents can be particularly dangerous at ice fog temperatures due to cold exposure. Ice fog is linked to high pollution levels in urban environments due to the aforementioned trapping of emissions under the inversion and hence is relevant to air quality assessments (Kim et al. 2014).
During ice fog conditions in the Fairbanks region, temperature inversions with surface temperatures more than 20°C colder than air aloft are possible (Malingowski et al. 2014; Vas et al. 2021). Vertical mixing is severely reduced during these conditions, and horizontal winds are typically very low at the surface. Winds above the inversion level are often still low and typically vary from northerly to easterly (Oliver and Oliver 1949) (see Fig. S1 in the online supplemental material). Above-inversion winds often dictate the direction of movement for the power plant plumes.
Ice fog in Fairbanks has long been a subject of scientific research, with foundational studies from the late 1940s to the 1960s on the processes governing ice fog genesis and the linkage between ice fog and local sources of moisture and air pollution (e.g., Oliver and Oliver 1949; Appleman 1953; Robinson et al. 1957; Benson 1965; Kumai 1964; Ohtake 1967; Ohtake and Huffman 1969; Benson 1970). Gotaas and Benson (1965) and Bowling et al. (1968) describe synoptic conditions during ice fog events in Fairbanks. Bowling et al. (1968) highlight that upper-level cold-air advection may be a prerequisite for ice fog formation, as radiative cooling alone does not generally produce temperatures cold enough for ice fog in the Fairbanks region. They describe “normal” ice fog events as having three phases: a preliminary phase in which cold air moves in aloft and temperatures drop, a main phase during which temperatures remain consistently low, and a recovery phase during which temperatures rise and ice fog dissipates. Bowling et al. (1968) identify ridging over the Bering Sea as the main weather pattern leading to “normal” ice fog events. In this pattern, cold air of northern origin is advected toward Fairbanks on the downstream flank of the ridge. As the ridge moves gradually eastward over Alaska, cloud cover clears, often leading to further cooling at the surface and ice fog onset. Renewed cloud cover and/or changes in upper-level flow direction lead to increasing temperatures and the end of the given ice fog event. In their detailed numerical modeling study of ice fog in Fairbanks, Kim et al. (2014) describe a synoptic setting similar to the “normal” pattern found by Bowling et al. (1968) in which “migratory high pressure moves from Siberia across Alaska” for an ice fog event in 2012. Schmitt et al. (2013) detail the microphysical properties of ice fog in Fairbanks during an extended time period in 2012 and highlight the calm conditions at and near the surface, as well as low pressure gradients over the Fairbanks area during ice fog conditions.
In part because of the limited availability of long-term data, there are few studies on temporal trends in ice fog occurrence on climatic time scales. Leung et al. (2020) notably found a reduction in ice fog occurrence since the 1950s at most but not all stations in their study area in the Hudson Bay Region (Canada) and attribute this primarily to rising temperatures due to climate change. In addition to climatic changes and rising temperatures in the Arctic, the frequency of ice fog occurrence may change with changing air quality or longer-term shifts in the occurrence of certain weather patterns. The exact nature of the correlations and causal connections between trends in ice fog occurrence, climate change, and other meteorological and anthropogenic forcings—such as air pollution—are as yet incompletely understood.
To aid research related to the interconnected processes driving ice fog formation, we aim to provide an overview of long-term changes in ice fog occurrence in the Fairbanks region. Using observational meteorological data and reanalysis, we address the following research questions:
-
When does ice fog typically occur in Fairbanks and has the frequency of occurrence changed over the period of record (1948/49–2021/22)? To do this, we investigate an extensive dataset of surface observations from two regional airports to identify ice fog periods since 1948.
-
What kind of synoptic-scale upper-level weather patterns are associated with ice fog in Fairbanks? Using self-organizing maps (SOM), which is an unsupervised machine learning algorithm, we analyze ERA5 500-hPa-level data and identify weather patterns most often associated with ice fog conditions.
Results are presented after a description of the data and methods. We then discuss uncertainties and limitations of the analyses and offer an interpretation of the observed changes and their possible causes.
2. Data and methods
a. Surface observations
We use aviation routine weather reports (METARs) from Fairbanks International Airport (hereinafter PAFA) and Eileson Air Force Base (hereinafter PAEI) to identify ice fog conditions based on visibility and temperature as recorded in the METARs. METARs are surface weather observations predominantly used for aviation purposes and provide the longest available record of visibility at PAFA and PAEI, starting in 1948.
PAFA is located approximately 7 km southwest of downtown Fairbanks (Fig. 1). The power plant at the University of Alaska, Fairbanks, is the nearest power plant to the airport (4 km north-northwest), and the plume rarely directly impacts airport operations. PAEI is 37 km southeast of Fairbanks. The airstrip at PAEI is near the power plant that supplies power and heat for the Air Force base. The power plant plume can directly impact airport operations during ice fog conditions. PAFA, Fairbanks, the town of North Pole, and PAEI form an arc north of the Tanana River. Farther north, low hills serve as a boundary for air motion. To the south, the terrain is mostly flat until the foothills of the Alaska Range. This shallow basin can serve as a trap for cold air during periods with little atmospheric motion.
Location of PAFA and PAEI airports (blue markers) and nearby power plants (red markers). The background image is from Google maps terrain view.
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
The frequency of METAR observations is typically hourly but has varied over time at both airports. In some years, observations were recorded at less than hourly frequency at both PAFA and PAEI (see Fig. S2 in the online supplemental material). More recently, observation frequency is often more than hourly, from approximately half-hourly to every 5 min in rare cases. To obtain a homogenized dataset of hourly values for statistical trend analysis, higher-frequency data were downsampled to hourly mean values. During periods in which METARs were recorded at greater than hourly intervals, linear interpolation was used to upsample the data to an hourly resolution. The upsampling is most relevant for the period 1965–72 at PAFA. METARs were recorded every 3 h during this time period.
Visibility observations were recorded in stratified bins with smaller (sub half mile; 1 mi = 1.6 km) bins for lower visibilities. The unit of observation was miles. A small number of obvious outlier values for visibility (>200 or <0 mi) was excluded from the analysis. Visibility measurements changed from human observations to automated in 1997 at PAFA and in 2007 at PAEI.
Data were analyzed per water year (i.e., per hydrological year, 1 October–30 September), focusing on the winter season (1 October–30 April). Because the first water year (1 October 1947–30 September 1948) is incomplete, our analysis begins with the water year 1949 (winter 1948/49) and extends to 2022 (winter 2021/22). No METARs were recorded at PAEI in 1971 and 1972; hence, the winter seasons 1970/71, 1971/72, and 1972/73 are incomplete and were excluded from trend analyses.
In addition to the METARs, daily precipitation data from PAFA were used for an assessment of the relationship between temperature, visibility, and the occurrence of precipitation for the entire time series at a daily scale. For approximately the last 2 decades of the time series, precipitation data are also available at an hourly resolution from the PAFA station, and a similar assessment was carried out at an hourly scale for this time period.
b. ERA5 upper-level reanalysis
We assess synoptic-scale upper-level weather patterns during ice fog conditions in Fairbanks using ERA5 geopotential height (GPH) at the 500-hPa level. The standard ERA5 pressure level dataset is available from 1959 onward (Hersbach et al. 2020; Climate Data Store 2017). For the period 1950–58, we use the preliminary ERA5 back extension (Bell et al. 2021). Using one global gridded 500-hPa GPH dataset per day, a daily time series of spatial anomalies of GPH was generated. Further analysis of relative high and low pressure patterns over Alaska was carried out using a smaller spatial subset spanning from the western Bering Sea to the Canadian Yukon in a west–east direction and from the Beaufort Sea to southeast Alaska in a north–south direction (165°–239°E, 50°–80°N).
c. Terminology and definitions
We use the term “ice fog” to refer to fog consisting only of suspended ice crystals that occurs at low temperatures, following the terminology of Gultepe et al. (2017) and the definitions of the WMO and AMS mentioned previously (American Meteorological Society 2022; WMO 2017). We only consider low-visibility conditions (observed visibility ≤ 1 mi), thereby excluding “diamond dust” from the analysis. Diamond dust also consists of suspended ice particles and occurs during similar conditions as ice fog but does not impede visibility to the same extent (e.g., Gultepe et al. 2017). METARs include a subjective determination of fog, but the consistency of these data over time is hard to assess, and ice fog may be misclassified as mist or drizzle at temperatures too cold for the presence of liquid droplets. Our analysis is based on identifying ice fog conditions via visibility and temperature thresholds that adhere to AMS and WMO definitions in order to avoid the potential issues associated with the subjective classification and to allow for better comparability between stations.
For the purposes of this study, we define an ice fog observation as any observation in which the temperature and visibility thresholds are met (≤−30°C; ≤ 1 mi). In the hourly dataset, an ice fog hour is any hour in which the thresholds are met. For downsampled cases, an ice fog hour is an hour in which the mean temperature and visibility meet the respective thresholds. Ice fog days are days in which at least one ice fog observation was recorded. Ice fog events are defined as periods of consecutive ice fog days. We also consider “cold” and “very cold” hours and days and define these analogously to the ice fog hours and days but without a visibility threshold (i.e., cold and very cold hours and days meet temperature thresholds of ≤−30° and ≤−40°C, respectively).
Previous studies of ice fog in Fairbanks report ice fog occurrence at temperatures ≤−30°C (Thuman and Robinson 1954; Ohtake and Huffman 1969). We follow this approach with the aim of 1) consistency with previous work and 2) excluding edge cases where supercooled liquid droplets may be present. The temperature–visibility relationship at PAFA and PAEI shows a “window” of frequently relatively high visibility (>1 mi) between approximately −25° and −30°C, with a sharp rise in occurrence of low visibility (≤1 mi) below this range (Fig. 2). This is in line with Oliver and Oliver (1949), who show a strong increase in ice fog below −30°F in Fairbanks, and with a comprehensive early study by the Canadian Meteorological Service (Rae 1954), which states: “The probability of fog decreases as the temperature drops from 32° to about −20°F. Below −20°F, there is a gradual increase in probability, which becomes more rapid as the temperature drops below −30°F. For temperatures below −40°F, ice crystal fog is almost certain to occur, although it may not be of sufficient density to affect air operations.” (For convenience, we note the following Fahrenheit-to-Celsius conversions: 32°F = 0°C, −20°F = −28.9°C, −30°F = −34.4°C, and −40°F = −40°C). Figure 2 in Gultepe et al. (2007) also shows a “window” around −30°C where fog does not typically occur and a strong increase in fog occurrences at lower temperatures. These observations match the pattern found at PAFA and PAEI, further confirming that a threshold temperature for ice fog of −30°C captures a shift in the causes of low visibility at cold temperatures (Fig. 2). At PAFA, the majority of low-visibility days above about −25°C are associated with precipitation, while this is rarely the case below about −30°C. While it is possible to have ice fog conditions at temperatures warmer than −30°C (Gultepe et al. 2015), Fig. 2 suggests that warmer-temperature cases are very rare in our dataset. They may occur more frequently in other regions and climates.
(top) Temperature histogram (daily minimum) of all days with minimum visibility of ≤1 mi at PAEI (black outlines) and PAFA (black bars). Gray bars indicate days with visibility of ≤1 mi and recorded precipitation at PAFA. (bottom) Temperature and visibility histogram as above, for ice fog hours (precipitation is not available at hourly resolution for most of the time series and is not shown).
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
Using a visibility threshold of 1 mi follows from the nature of the METARs: visibility is recorded in miles, and values are stratified rather than continuous (Fig. S3 in the online supplemental material). Hence, a conversion to kilometers would not allow for a direct comparison with the more commonly used fog visibility threshold of 1 km.
Considering Fig. 2, it is apparent that there are rare occurrences in the daily dataset when the ice fog visibility and temperature thresholds are met and precipitation was recorded at PAFA. This suggests that low visibility due to precipitation may be misclassified as ice fog in certain cases. However, data from PAFA indicate that precipitation is very rare when both the daily high and the daily low temperatures are ≤−30°C. For days with lows below −30°C and recorded precipitation, the daily high temperature was typically well above −30°C. Thus, we argue that while misclassifications cannot be ruled out entirely, the ice fog days with precipitation as seen in Fig. 2 are most likely due to cases when, e.g., a frontal system brought precipitation in the morning and ice fog formed later in the day as cloud cover cleared and temperatures dropped, or ice fog was present in the morning but cleared during the day with a change in the weather and the onset of precipitation. We discuss this in more detail with specific examples in section 4b.
Using the above threshold criteria, the PAFA dataset contains 8893 ice fog observations. When high-frequency values are downsampled, 8100 ice fog hours remain (i.e., about 10% of ice fog hours are downsampled averages of multiple observations per hour). When low-frequency values are upsampled using linear interpolation, the resulting dataset contains 9441 ice fog hours (i.e., about 16% of the hourly dataset used for trend analyses consists of upsampled values). The corresponding numbers for PAEI are: 9765 ice fog observations, 8451 ice fog hours after downsampling (about 15% of ice fog hours are downsampled averages of multiple observations per hour), and 8565 ice fog hours after upsampling. These values correspond to 882 and 957 fog days for PAFA and PAEI, respectively. As with any averaging and resampling, some detail is lost in the process. We also acknowledge that the temperature and visibility thresholds may not capture 100% of ice fog occurrences, e.g., due to precipitation misclassifications mentioned above or to rare cases of ice fog at warmer temperatures. We also note that changes in observing methods at the stations over time (e.g., changing frequency of observations, shift from human to instrument-based visibility records) may cause some uncertainties in the time series that cannot be quantified in detail. Nonetheless, we believe the long, largely continuous, and high-quality METAR time series of visibility and temperature provide the best available data basis for a comprehensive assessment of trends in ice fog occurrence in the Fairbanks region.
d. Change analysis of ice fog time series
The frequency of occurrence of ice fog during the study period was investigated using several techniques. As a first-order assessment of change in ice fog occurrence over time, the number of ice fog hours and days for each 30-yr climate normal period covered by the time series was computed, along with additional statistics related to ice fog hours and days per normal period. The climate normal periods considered are: 1951–80, 1961–90, 1971–2000, 1981–2010, and 1991–2020.
The interannual variability of the number of ice fog days per year is high. A bootstrap analysis (Efron and Tibshirani 1994) was employed to determine whether the noise in the data could be the source of an artificial trend. For this analysis, a least squares fit line was calculated for the number of ice fog days per winter season. The slope of the best fit line is the rate of change in ice fog. The correlation coefficients between the best fit line and the raw data are low due to the variability in the dataset. Each dataset was randomly reshuffled 100 000 times. For each reshuffle, a least squares fit was calculated, and the number of times that the absolute value of the slope of the shuffled line fit exceeded the slope from the original data was tabulated. Few exceedances indicate confidence in the slope calculated for the actual data.
While “change per decade” derived from least squares fit lines can provide a useful and intuitive trend metric, linear regression per definition fails to show nonlinear characteristics of time series. Empirical mode decomposition (EMD) is more suitable for trend detection in nonlinear and nonstationary time series (e.g., Huang et al. 1998; Wu et al. 2007). EMD decomposes a signal into intrinsic mode functions (IMFs) and a residual component, where each IMF has one zero crossing between each pair of consecutive minima and maxima and a local mean of zero. The residual can be interpreted as the trend associated with the time series, while the IMFs represent noise and/or cyclical variability. We apply EMD as implemented for Python by Quinn et al. (2021) to assess trends in the frequency of occurrence of ice fog hours and related time series data and refer to the works cited above for further details on the EMD method. An increasing number of use cases shows the potential of EMD and variations thereof for trend and variability analysis of climatological data on multiple time scales (Franzke 2009; Lee and Ouarda 2011; Franzke 2012; Ezer and Corlett 2012; Franzke and Woollings 2011). The IMFs and residuals can reveal intrinsic, nonlinear trends as well as natural variability (Wu et al. 2007). Water years with large data gaps (1971–73 at PAEI) were removed before trend calculations.
e. Upper-level weather patterns during ice fog conditions
We assess the synoptic-scale weather patterns during ice fog conditions in Fairbanks using ERA5 GPH data on the 500-hPa level. To generate an overview of common wintertime patterns of troughing and ridging over Alaska and the surrounding areas, we applied a SOM algorithm to daily GPH anomalies over the region of interest. All days between 1 October and 30 April for every winter season in the 1950–2022 ERA5 time series were used as input data. The SOM acts as an unsupervised clustering mechanism that groups similar patterns into discrete nodes (i.e., clusters). Each “snapshot” of GPH over the region of interest acts as one input vector for the SOM, which places nodes throughout the input data space. During the training process, the nodes and neighboring nodes are updated iteratively based on a neighborhood function and learning rate to better match the input data. The size of the SOM, i.e., the number of nodes, as well as the initial values for learning rate and neighborhood function, are set before training and influence the outcome. In particular, a higher number of nodes (larger SOM size) produces a more detailed picture of different patterns, while a lower number of nodes produces broader clusters and resolves less detail. Nodes are vectors of the same shape as the input vectors and can accordingly be visualized as maps of GPH anomalies in our application. The result of the SOM analysis is a clustering of upper-level weather conditions into distinct patterns, the occurrence of which can be tracked over time.
Detailed background on the principles of SOM can be found in the publications of Kohonen (1990, 1991, 2012). SOMs are used for pattern analysis, clustering, and classification applications in many fields of research. Reviews of climatological and meteorological SOM applications are given in Liu and Weisberg (2011) and Sheridan and Lee (2011). Example use cases for SOM in climatological studies in the Arctic and Alaska are, e.g., Cassano et al. (2006), Cassano and Cassano (2010), Hartl et al. (2020), Litzow et al. (2020). We used the Python package MiniSom to implement the algorithm (Vettigli 2018). Mean or individual quantization errors (QE), that is, the Euclidean distance between input vectors and their assigned node, can serve as a performance measure for the SOM. However, suitable values for SOM size and initial values depend strongly on the nature of the data and the SOM application, and a low mean QE does not necessarily correspond to the most useful result. Accordingly, some level of subjectivity is associated with the selection of SOM size and other initial settings. In our case, the goal of the SOM analysis was to obtain a clustering of large-scale, upper-level weather patterns during ice fog conditions as well as general wintertime weather patterns. The clusters were then linked with METAR observations to identify patterns that occur during ice fog conditions. We used the PAFA data to identify ice fog conditions for the SOM analysis because PAFA has fewer data gaps than PAEI and local conditions are less affected by the nearby power plants. Based on trial and error and QE assessments, we chose a SOM size of 3 × 2, which results in 6 distinct patterns. Larger SOMs resolve differences in weather patterns at a greater level of detail, but the additional detail represents a trade-off in terms of interpretability for our applications. The initial values for the neighborhood function and learning rate were 0.5 and 0.5, respectively.
3. Results
a. Changes over time
The maximum number of fog days per winter is 36 at both stations and occurred in 1969 at PAFA and in 1956 at PAEI (Fig. 3). The maximum number of fog hours was recorded in the same years (618 at PAFA and 352 at PAEI). The PAFA time series contains 7 yr without any recorded ice fog: 1977, 1988, 2010, 2011, 2016, 2018, and 2021. At PAEI, 1977 and 2016 are the only years without any recorded ice fog. Ice fog occurs between late October and about mid-March at both stations. The latest dates on which ice fog was recorded are 23 March 1996, at PAFA and 17 March 1965, at PAEI. The earliest dates with recorded ice fog are 2 November 1976, and 28 October 1997, respectively. The latest ice fog onset date was 28 February at PAFA and 30 January at PAEI, considering only years in which ice fog occurred.
Daily minimum temperature by water year (x axis) and day of water year (days since 1 Oct; y axis) at (top) PAFA and (bottom) PAEI. Black crosses are ice fog days. The color scheme changes below −30°C to highlight the temperature threshold for ice fog. Data gaps are not colored. The associated scatterplots show the number of ice fog days per water year and the linear regression line.
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
The number of ice fog hours and days per season decreased at both stations during the study period. A linear regression shows that ice fog hours decreased by 28 and 33 h per decade at PAFA and PAEI, respectively. Similarly, the trend in ice fog days shows 2.6 fewer yearly ice fog days per decade at PAFA and 3.0 at PAEI. For both hours and days, the bootstrap analysis showed that it was extremely unlikely for the slope of the shuffled data to exceed the slope of the original data. Typically, less than 0.01% of the reshuffled trials led to a stronger slope, indicating that the trends are unlikely to be the result of random variability in the data.
The trend residuals of the EMD analysis for the PAFA data indicate that the number of fog hours began to decrease in the early 1970s (Fig. 4; Fig. S4 in the online supplemental material). The maximum decrease rates were reached around 1990 with 7–8 h per year. Since about 1990, the rate of change has become less negative, i.e., the decreasing trend continued but at a lesser rate than in previous decades. In the most recent years, the rate of change approached 0. This is in line with the evolution of fog hour and day statistics throughout the 30-yr climate normal periods. The median number of ice fog days per winter at PAFA declined from 16.5 in the period 1951–80 to 6 in the period 1991–2020. The median number of ice fog hours per winter at PAFA was highest in the 1961–90 period (171.5 h) and lowest in the 1991–2020 period (46.5 h). The greatest reduction between consecutive climate normal periods took place between the period 1971–2000 (15 days, 164.5 h) and the period 1981–2010 (9 days, 77 h). At PAEI, the EMD residual also shows a decreasing trend with varying rates of change. Relative to PAFA, the decrease in fog hours at PAEI began earlier. The rate of change was most negative in the mid-1970s and reached another minimum around 2000. Like at PAFA, the decreasing trend at PAEI has leveled off in the most recent years of the time series with rates of change near 0.
(top) Ice fog hours and cold hours per winter season with EMD trend residuals and (bottom) rates of change (derivatives of the trend residual) at (left) PAFA and (right) PAEI. For PAEI, the incomplete water years (1971–73) were removed before trend computation.
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
The residual trend line in the number of cold hours (≤−30°C) per season at PAFA tracks the trend in fog days relatively closely, with the decrease leveling off toward the end of the time series (Fig. 4). This is also reflected in the median number of cold days for the 30-yr climate normals: the number of cold days decreased for the earlier climate normal periods but showed a slight increase between the period 1981–2010 and the period 1991–2020 at both stations (Table 1). The most negative rate of change in cold hours at PAFA was reached in the mid-1980s, slightly preceding the period of the most negative rate of change in fog hours. At PAEI, trends in fog hours and cold hours appear less well correlated, with larger overall rates of change and greater fluctuation of rate of change for cold hours than for fog hours. Since the mid-1990s, the trend in the number of cold hours has stagnated or been slightly positive at PAEI, while the decrease of fog hours continued into the 2010s.
Climate normal values related to fog days and cold days at PAFA and PAEI for the 30-yr normal periods 1951–80, 1961–90, 1971–2000, 1981–2010, and 1991–2020.
The median first day of ice fog per season was 1 December (61st day of water year) for the climate normal periods 1951–80 and 1961–90 at PAFA (PAEI: 63rd day of water year) and shifted to later dates in the following periods (Table 1). Median ice fog onset in the period 1991–2020 occurred almost a month later in the period 1990–2020 than in the period 1961–90 [29 December (89th day of water year at PAFA; 81st at PAEI)]. The last day with ice fog in a winter season has been, on average, occurring earlier in recent years, but the changes in the normals are less pronounced than for ice fog onset. The median last day of ice fog occurred 20 days earlier in the period 1991–2020 than in the period 1961–90 at PAFA and about 6 days earlier at PAEI. Figure 5 shows the same general tendencies of shortening ice fog seasons. The EMD trend residuals indicate that the rate of change in first and last days is not constant throughout the time series.
(top) Ice fog hours and cold hours per winter season at PAFA and PAEI, with EMD trend residuals. (middle) First and last days with ice fog per season at both stations, with EMD trend residuals. (bottom) Percentage of cold (≤−30°C) and very cold (≤−40°C) hours that are ice fog hours at both stations, with EMD trend residuals.
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
The average duration of ice fog events (consecutive days with ice fog) also shows decreasing trends. During the 1951–80 climate normal period, the longest fog event per season was about 8–9 days, as compared with 3–4 days in the 1990–2020 period. Years with at least one fog event that lasted 1 week or more were relatively common until about 1980 and have since become more rare (Figs. 3 and 5). The most recent years with ice fog events lasting 1 week or longer are 2009 and 1991 at PAFA and PAEI, respectively.
Temperatures during ice fog hours show some variation over time and a slight overall increase during the period of record (Fig. S5 in the online supplemental material). Considering the ratio of fog days to cold days (≤−30°C) and very cold days (≤−40°C), a decreasing trend is apparent for both cold and very cold days at both stations. The trend residuals indicate that until about the mid-1970s, ice fog occurred in about 20%–30% of cold hours (Figs. 3 and 5). This value dropped to below 20% after the mid-1990s, with a stronger decrease at PAFA. Similarly, the trend line for very cold hours shows that ice fog occurred during 60%–80% of very cold hours in the early decades of the time series. This value dropped to below 60% after about 2000 and below 40% in the 2010s (Fig. 5).
b. Synoptic-scale weather patterns during ice fog
The main result of the SOM analysis is a general clustering of 500-hPa GPH anomalies during ice fog days at PAFA (Fig. 6). The 3 × 2 SOM shows that ice fog typically occurs during large-scale upper-level troughing patterns, when the trough is positioned in such a way that pressure gradients over Fairbanks are relatively low and the upper-level flow is broadly northerly. The SOM algorithm places more nodes in regions of the data space where there are more input vectors, i.e., generates more clusters for patterns more often represented in the input data. This is illustrated by clusters 0, 2, and 4 in Fig. 6, which all show variations of a strongly meridional pattern with ridging over the Bering Sea and Alaska’s west coast and troughing over interior Alaska and western Canada. Clusters 0 and 2 show a cutoff low embedded in the main, positively tilted trough, while cluster 4 is an open-wave trough with a negative tilt. Despite these differences, the clusters are relatively similar in terms of the overall pattern and together account for 48.6% of the ice fog days at PAFA, i.e., almost half of the ice fog days used to train the SOM are mapped to synoptic-scale weather patterns that have omega block–type ridging over the Bering Sea and long-wave troughing over Fairbanks downstream of the blocking high. The remaining ice fog days are distributed over clusters 1 (15.5%), 3 (12.9%), and 5 (22.9%). Cluster 1 is characterized by low GPH gradients and a broadly west-northwest upper-level flow over Fairbanks. In cluster 3, Fairbanks is located near the axis of a trough with a strongly positive tilt, which forms a high-over-low-type blocking pattern (Rex 1950; Sousa et al. 2021) with a high over the Chukchi Sea and low pressure in the southwest of the region of interest. Cluster 5 is the most zonal pattern and shows a short-wave trough over the northern and central parts of Alaska.
The 3 × 2 SOM trained on GPH anomalies (data: ERA5; 500-hPa level) during all ice fog days at PAFA since 1 Jan 1950. The unit is geopotential meters (gpm). The percentage value in the panel titles indicates the amount of input data mapped to each cluster. The black triangle marks the position of Fairbanks.
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
Grouping ice fog days by the SOM nodes (i.e., weather pattern clusters) they are mapped to shows that the average visibility is lower during ice fog associated with clusters 0, 2, and 4 than during ice fog associated with the other clusters, although all clusters span the full range of “allowed” visibility values (0–1 mi; Fig. 7). Median temperature during ice fog is below −40°C for all clusters except cluster 1. The median duration of ice fog events and the number of ice fog hours per ice fog day are higher for clusters 0, 2, and 4 than for the other clusters. In summary, persistent ice fog events with low visibilities and temperatures occur during all of the pattern clusters identified by the SOM, but the pattern associated with clusters 0, 2, and 4 (ridging over the Bering Sea, troughing over interior Alaska) is most common during long-lasting, low-visibility ice fog.
Boxplots of temperature, visibility, and ice fog duration grouped by the pattern clusters in Fig. 6.
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
For a direct comparison of typical ice fog upper-level patterns and other common winter season (October–April) weather patterns, Fig. 8 shows the winning nodes of a 3 × 2 SOM trained on all data using the same initial conditions as in Fig. 6. In Fig. 8, only cluster 0 resembles the broadly meridional pattern with ridging in the Bering Sea that appears three times in Fig. 6. Clusters 1 and 2 show variations of a less pronounced trough over the northern and central parts of Alaska. Clusters 3 and 4 represent high pressure over much of Alaska, while cluster 5 shows a low-gradient situation with low pressure in the Bering Sea extending northeast to Alaska’s Arctic coast and high pressure over parts of interior Alaska and Canada. Over 60% of ice fog days are mapped to clusters 0 and 1, which is to be expected given their similarity to the patterns in Fig. 6. Clusters 0 and 1 together make up about 21% of all wintertime patterns (i.e., 20.7% of all input data are mapped to clusters 0 or 1 and these account for about 60% of ice fog days, Figs. 8 and 9). Around 30% of the remaining fog days are mapped to clusters 2 and 5. Very few fog days are associated with the high pressure patterns of clusters 3 and 4. The frequency of occurrence of pattern clusters varies from year to year. Years with no or very few ice fog days in the METAR time series tend to show comparatively low fractions of weather patterns conducive to ice fog. No clear long-term trends over the length of the study period were detected (Fig. 9).
As in Fig. 6, but on all October–April days.
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
(top) Fraction of October–April days mapped to each of the pattern clusters shown in Fig. 8. (bottom) Percentage of fog days mapped to each of the clusters.
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
4. Discussion
a. Notes on trend analyses
Linear trends as produced by ordinary least squares regression are, arguably, an intuitive and useful way of communicating change over time but depend strongly on the time period assessed in the regression, particularly for noisy data, and have a number of other known shortcomings. Trend analysis based on bootstrapping ensures that identified trends have a low likelihood of being artifacts of the natural variability present in the data (e.g., Mudelsee 2019). Nonetheless, linear regression obscures potential nonlinear trends and signals, which may be important to understanding the physical processes underlying a given climatological time series (e.g., Wu et al. 2007; Mudelsee 2019). The computation of trend residuals via EMD provides an alternative visualization of trends in the ice fog time series that accounts for nonlinearities. As with any kind of trend analysis, results should be interpreted with care, for example, because of possible boundary effects related to how the algorithm extrapolates the signal beyond the edges of the time series to generate the IMFs (Stallone et al. 2020). However, the EMD residuals clearly show that changes in ice fog parameters and cold days did not occur at uniform rates over the length of the time series and are not necessarily monotonic (Figs. 4 and 5).
b. Possible misclassifications due to precipitation
Low visibility at the study sites occurs from either fog or precipitation. We chose the temperature threshold of −30°C to exclude cases of fog with liquid droplets, following previous work (Thuman and Robinson 1954; Ohtake and Huffman 1969; Oliver and Oliver 1949; Rae 1954). Because of the lower moisture content of cold air, precipitation below −30°C is not very common, but it does occur. Since we identify ice fog cases in the METAR data using temperature and visibility thresholds, this may lead to misclassifications when the temperature threshold is met and visibility is low because of precipitation rather than ice fog.
Small amounts of precipitation were recorded on 82 ice fog days during the period of record. This corresponds to about 9% of all ice fog days. However, this does not account for cases when ice fog is present during parts of a given day and precipitation was recorded before or after the ice fog occurrence on that same day, for example, when fog dissipates due to a change in the weather and precipitation sets in subsequently. For a more detailed assessment, we consider precipitation during ice fog hours: for the approximately 2 decades during which records of hourly precipitation at PAFA are available (November 2002–present), only 11 ice fog hours coincide with precipitation hours (1.5%), i.e., precipitation during ice fog hours is very rare and the values based on daily data likely overestimate the number of misclassifications resulting from precipitation.
The maximum amount of precipitation recorded on an ice fog day is 5.3 mm (5.6 cm of snowfall) on 14 January 2005. This was the last day of a multiday ice fog event, i.e., the fifth consecutive day classified as an ice fog day based on our criteria. Hourly precipitation from PAFA is available for this time period, so we can assess the timeline of this ice fog event (Fig. 10) in more detail.
Meteorological conditions during an ice fog event in January 2005 at PAFA: (top) Temperature and visibility, with horizontal dashed lines showing the temperature (black) and visibility (gray) thresholds for ice fog. Red shading indicates ice fog hours. (top middle) Hourly precipitation. (bottom middle) Wind speed and direction. (bottom) Sea level pressure.
Citation: Journal of Applied Meteorology and Climatology 62, 9; 10.1175/JAMC-D-22-0190.1
A small amount of precipitation was recorded on 9 January. The wind picked up toward the end of this snowfall. On 10 January, the wind died down, surface air pressure rose, and temperatures dropped more than 10°C to below the threshold of −30°C. A sharp drop in visibility was recorded shortly after the temperature fell below the threshold. The first ice fog hours were recorded at 2000 and 2100 UTC 10 January. Accordingly, 10 January was the first ice fog day of this event. Visibility then varied between 1 and 4 mi for the rest of the day and into 11 January, before falling to 1 mi and below in the afternoon of 11 January for a longer, continuous period. Temperatures hovered around −40°C during this phase of low visibility, which lasted until the evening of 13 January. Temperature and visibility then began rising again. The last ice fog hour was recorded in the early-morning hours of 14 January, making 14 January the last ice fog day of the event. Later in the day, precipitation set in. Visibility dropped again as precipitation intensified, but temperatures remained well above −30°C (Fig. 10). Accordingly, the low visibilities during the precipitation event on 14 January were not classified as ice fog hours. Low visibility due to ice fog and low visibility due to onsetting precipitation can be clearly distinguished. The data suggest that the general sequence of precipitation end and ice fog onset, or ice fog dissipation and precipitation onset, causes the majority of cases in which precipitation was recorded on ice fog days. An additional example of a shorter ice fog event at an hourly resolution is given in Fig. S6 in the online supplemental material. The very low amount of ice fog hours with precipitation in the PAFA dataset supports the assumption that true misclassifications due to precipitation are rare (about 1.5% oversampling of ice fog hours based on PAFA data).
c. Linking surface conditions with larger-scale weather patterns
Ice fog formation requires favorable conditions both locally in the near-surface atmosphere and at higher atmospheric levels at the synoptic scale. In Fairbanks, observational data are only available from PAFA, but visibility is—anecdotally—often considerably lower in the downtown area than at PAFA (Benson 1965). Similarly, local topography influences near-surface airflow, or lack thereof, and the formation of pools of cold air and inversions conducive to ice fog. Anthropogenic sources of moisture and the availability of ice nuclei can also have a strong impact on ice fog formation and vary strongly at small spatial scales.
Considering the larger-scale weather patterns relevant to ice fog as defined in this study, one might ask which kind of weather patterns allow for very low temperatures in Fairbanks (≤−30°C). Local radiative cooling and associated temperature inversions in near-surface layers undoubtedly play a central role. Generalizing broadly, low-gradient synoptic-scale weather patterns that produce calm, low-wind conditions on the ground would be favorable for the formation of strong inversions, as cold air can pool undisturbed. Furthermore, Bowling et al. (1968) suggest that cold-air advection on the synoptic scale may be needed in addition to radiative cooling to reach temperatures low enough for ice fog in Fairbanks based on their observations of very rapid, large temperature drops that precede the ice fog events in their study. Thus, the local phenomenon of ice fog is dependent on larger-scale weather patterns that allow for substantial radiative cooling and may also be impacted by cold-air advection at the synoptic scale.
Aiming for a generalized clustering scheme of common ice fog weather patterns, we use a SOM algorithm to group ERA5 GPH anomalies by recurring patterns based on the PAFA ice fog time series. We acknowledge that the reanalysis does not resolve equally important local factors, particularly local cold pools and inversions, but argue that it does provide valuable information at the synoptic scale.
The exact output of the SOM varies with the size of the SOM and other tuning parameters, although the overall characteristics of the resulting clusters remain similar. The presented winning nodes should not be considered a definitive classification of ice fog weather patterns. Different applications may well require different SOM setups. The SOM clusters also represent a form of averaging that leads to a loss of detail that is present in the input data (GPH anomaly maps for each ice fog and non ice fog day in the time series). For a general assessment of climatological trends in ice fog occurrence, we consider this kind of generalization useful.
We return to the ice fog event of January 2005 (Fig. 10) to briefly illustrate how the generalized clusters relate to the actual synoptic patterns. The first day of the ice fog event, 10 January, is mapped to cluster 5 (Fig. 6); 11 and 12 January are mapped to cluster 4; and 13 and 14 January are mapped to cluster 1. This means that in terms of the SOM clusters, the ice fog event progressed from a relatively zonal pattern (cluster 5) to a more meridional pattern (cluster 4), which was followed by an upper-level pattern characterized by low pressure gradients and a northwesterly upper-level flow (cluster 1, Fig. 6). In comparing this with the actual reanalysis data for the days in question (Fig. S7 in the online supplemental material), it is apparent the clusters broadly match the upper-level synoptic patterns, although details are lost. From 10 to 11 January, the relatively strong upper-level flow shifted from northwesterly to more northerly directions, leading to a more meridional pattern with ridging over southwest Alaska and the Bering Sea and troughing over Canada and Alaska’s northern and eastern interior, with relatively low pressure gradients in the Fairbanks area. In the following days, the upper-level flow over Fairbanks once more shifted to west-northwest directions between ridging in the south and an upper-level low over the Arctic Coast that likely brought more unsettled weather and precipitation.
d. General interpretation of results
Linear trends as well as a comparison of climate normal periods and the EMD analysis all clearly indicate an overall reduction in the number of ice fog hours and days at PAFA and PAEI. This is in line with a general decline in ice fog occurrence observed at airfields in Hudson Bay, Arctic Canada (Leung et al. 2020). Trends in ice fog hours and cold hours in Fairbanks show similarities during the early decades of the time series and tend to diverge in more recent years, with the number of cold hours recovering from the previous decline to a greater extent than the number of ice fog hours. The rate of change in ice fog hours is near 0 at PAFA and PAEI in the most recent measurement years. Leung et al.’s (2020) results also point toward a leveling off of ice fog decline in recent decades at some of their stations. More studies on long-term trends in ice fog occurrence are needed to investigate this in more detail.
Depending on the parameter and station, rates of change in Fairbanks peaked between the 1970s and approximately 1990. Trends are less pronounced in the more recent decades for ice fog hours as well as cold hours. A shift of the Pacific decadal oscillation (PDO) from a negative to a positive phase in 1976 has been tied to a shift toward a warmer climatological regime in Alaska, along with a deepening of the Aleutian low in the winter months and related advection of warmer air masses into interior Alaska (Hartmann and Wendler 2005). This PDO shift and associated changes may have played a role in the decline in ice fog, but the complex nature of teleconnections across different spatial and temporal scales does not allow for a definitive assessment of cause and effect. Furthermore, recent studies suggest that the relationship between phases of the PDO index and climatological parameters is not constant over time and may be changing under “novel climate conditions” (Litzow et al. 2020).
The disconnect between the reduction of cold days and the reduction of ice fog days is intriguing. Our data suggest that if there were a constant, linear relationship between ice fog and cold days, ice fog would be more common now than observations show. This implies that the reduction of cold days is not the only factor contributing to the reduction of ice fog days over the time series. Although the population of the Fairbanks North Star Borough has more than doubled since 1970, vehicle emission standards have improved dramatically in this time. Reduced pollution could lead to reductions in ice fog formation and longevity. Kumai (1964) showed very high levels of combustion particles as ice fog nuclei based on electron microscopic studies. Repeating such work would be desirable in order to shed some light on the processes driving the change in the ratio of ice fog days to cold days. As air pollution sources in the Fairbanks area have likely changed since the 1960s, the types of ice fog–relevant nuclei could be different, thus affecting ice fog formation, even if the nuclei concentration may not have changed significantly.
The output of the SOM when trained on only ice fog days as compared with all days in the time series (Figs. 6 and 8) shows that a northerly flow aloft is common during ice fog in Fairbanks, which supports Bowling et al.’s (1968) finding that synoptic-scale cold-air advection is relevant for ice fog formation in Fairbanks. Bowling et al. (1968) distinguish between “normal” and “exceptional” ice fog events based on upper-air and surface conditions during 15 ice fog events. Normal events are preceded by ridging over Siberia and a northerly flow over interior Alaska. A migratory high forms and moves eastward to the Bering Sea, into Alaska (skies clear, surface temperatures drop, and ice fog forms), and eventually farther on to Canada (cloud cover returns, surface temperatures rise, and ice fog dissipates). In contrast, exceptional events are longer lasting and tend to end with a retrogressive shift of large ridging and troughing patterns over the Bering Sea and central Alaska, respectively. In a general sense, versions of both the normal and the exceptional patterns described by Bowling et al. (1968) appear in the SOM-based clustering of ice fog weather patterns. However, it should be kept in mind that the SOM analysis is based on ice fog days rather than ice fog events, whereas the classification of Bowling et al. (1968) focuses on the evolution of synoptic patterns during ice fog events, i.e., during a series of consecutive ice fog days. Our analysis considers only the 500-hPa level to determine common large-scale GPH patterns, while Bowling et al. (1968) include a detailed assessment of the evolution of conditions at lower levels in their work. Although beyond the scope of this study, we note that there is much potential for applying machine learning techniques for further data analysis related to questions of ice fog classification based on meteorological conditions at various scales and along the vertical and horizontal spatial dimensions. An approach similar to that of Bowling et al. (1968) might be recreated with the much larger datasets that have since become available, to distinguish more comprehensively between different types of ice fog weather patterns and their occurrence over time.
5. Conclusions
Based on temperature and visibility data from Fairbanks International Airport and Eielson Airforce Base, we conclude that ice fog conditions are occurring less often than they have in the past at the study sites. METAR data show that there have been systematic reductions in ice fog occurrence throughout the over 70-yr record, with a reduction of 60%–70% between the climatological reference periods 1950/51–1979/80 and 1990/91–2019/20. While there is much variability in the data, including years with no ice fog cases, the observed trends are very robust. These results are consistent with expectations given past and ongoing warming in the region and a reduction in the number of days with temperatures cold enough for ice fog occurrence.
Trends in ice fog occurrence and sufficiently cold temperatures have diverged in recent decades. This suggests that factors other than the reduction of cold days and hours, such as changes in the amount or characteristics of local air pollution and sources of moisture, may also substantially impact trends in ice fog occurrence.
Synoptic-scale, upper-level weather patterns during ice fog conditions in Fairbanks are most often characterized by a northerly flow aloft and relatively low pressure gradients. This confirms earlier studies suggesting that cold-air advection aloft in addition to radiative cooling is conducive to ice fog formation in Fairbanks (Bowling et al. 1968). There is considerable variability in how often particular weather patterns associated with ice fog occur, without a clear signal of long-term trends.
The extreme variability in year-to-year cases suggests that ice fog is not likely to disappear altogether in the near future. However, years without any ice fog in the Fairbanks region will likely become more common given current trends. While the overall reduction in ice fog days in Fairbanks over the period of record is clear, more study is needed to fully understand the change in the ratio of ice fog occurrences to occurrences of cold temperatures, as well as the interconnected, larger-scale drivers of change and their relative importance, i.e., overall warming due to climate change, potential influences of oscillatory climate regimes such as the PDO, and shifts in synoptic-scale weather patterns that may be associated with a changing climate system. Addressing these issues requires research spanning from microphysical investigations of ice fog particles to boundary layer meteorology and high-resolution modeling thereof, and to synoptic-scale meteorology and climatology.
The physical processes leading to ice fog formation are clearly key to understanding change and variability in ice fog occurrence. However, ice fog is a weather phenomenon that not only affects human activities, at times presenting a considerable hazard to aviation and other traffic, but is also directly and causally influenced by human activities that add vapor and nuclei for ice fog particles to the near-surface atmosphere. Ice fog is not independent of factors such as local and regional economic development and adoption of new technologies that add to or reduce sources of air pollution (Benson 1965, 1970). Accordingly, these issues should be incorporated into research aiming to understand past changes and make predictions about future ice fog occurrence.
Acknowledgments.
This material is based upon work supported by the Broad Agency Announcement Program from the U.S. Army Cold Regions Research and Engineering Laboratory (ERDC-CRREL) under Contract W913E521C0017 from the U.S. Army Basic Research Program (Program Element 0603119A, Ground Advanced Technology). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Broad Agency Announcement Program or ERDC-CRREL.
Data availability statement.
METARs for PAFA and PAEI were downloaded from the Iowa State University ASOS Network data platform, where they are freely accessible (https://mesonet.agron.iastate.edu/request/download.phtml). Daily precipitation data from Fairbanks International Airport were obtained through the API of the RCC ACIS web services and are freely available there (https://www.rcc-acis.org/docs_webservices.html). Hourly data from PAFA were obtained online through the synopticdata API (https://developers.synopticdata.com/mesonet/). ERA5 data were accessed through the Copernicus Climate Data Store API and are also freely available (https://cds.climate.copernicus.eu/).
REFERENCES
American Meteorological Society, 2022: Ice fog. Glossary of Meteorology, https://glossary.ametsoc.org/wiki/Ice_fog.
Appleman, H., 1953: The cause and forecasting of ice fogs. Bull. Amer. Meteor. Soc., 34, 397–400, https://doi.org/10.1175/1520-0477-34.9.397.
Bell, B., and Coauthors, 2021: The ERA5 global reanalysis: Preliminary extension to 1950. Quart. J. Roy. Meteor. Soc., 147, 4186–4227, https://doi.org/10.1002/qj.4174.
Benson, C. S., 1965: Ice fog: Low temperature air pollution; defined with Fairbanks, Alaska as type locality. Geophysical Institute at the University of Alaska Tech. Rep., 194 pp., https://scholarworks.alaska.edu/handle/11122/3646.
Benson, C. S., 1970: Ice fog: Low temperature air pollution. Corps of Engineers, U.S. Army, Cold Regions Research and Engineering Laboratory Research Rep. 121, 118 pp., https://books.google.co.in/books?id=uoqV1JLgKoQC&source=gbs_navlinks_s.
Bowling, S. A., T. Ohtake, and C. S. Benson, 1968: Winter pressure systems and ice fog in Fairbanks, Alaska. J. Appl. Meteor., 7, 961–968, https://doi.org/10.1175/1520-0450(1968)007<0961:WPSAIF>2.0.CO;2.
Cassano, E. N., and J. J. Cassano, 2010: Synoptic forcing of precipitation in the Mackenzie and Yukon River basins. Int. J. Climatol., 30, 658–674, https://doi.org/10.1002/joc.1926.
Cassano, J. J., P. Uotila, and A. Lynch, 2006: Changes in synoptic weather patterns in the polar regions in the twentieth and twenty-first centuries, Part 1: Arctic. Int. J. Climatol., 26, 1027–1049, https://doi.org/10.1002/joc.1306.
Climate Data Store, 2017: ERA5: Fifth generation of ECMWF atmospheric reanalyses of the global climate. Climate Data Store, https://cds.climate.copernicus.eu/cdsapp#!/home.
Efron, B., and R. J. Tibshirani, 1994: An Introduction to the Bootstrap. CRC Press, 456 pp.
Ezer, T., and W. B. Corlett, 2012: Is sea level rise accelerating in the Chesapeake Bay? A demonstration of a novel new approach for analyzing sea level data. Geophys. Res. Lett., 39, L19605, https://doi.org/10.1029/2012GL053435.
Franzke, C., 2009: Multi-scale analysis of teleconnection indices: Climate noise and nonlinear trend analysis. Nonlinear Processes Geophys., 16, 65–76, https://doi.org/10.5194/npg-16-65-2009.
Franzke, C., 2012: Nonlinear trends, long-range dependence, and climate noise properties of surface temperature. J. Climate, 25, 4172–4183, https://doi.org/10.1175/JCLI-D-11-00293.1.
Franzke, C., and T. Woollings, 2011: On the persistence and predictability properties of North Atlantic climate variability. J. Climate, 24, 466–472, https://doi.org/10.1175/2010JCLI3739.1.
Gotaas, Y., and C. S. Benson, 1965: The effect of suspended ice crystals on radiative cooling. J. Appl. Meteor., 4, 446–453, https://doi.org/10.1175/1520-0450(1965)004%3C0446:TEOSIC%3E2.0.CO;2.
Gultepe, I., and Coauthors, 2007: Fog research: A review of past achievements and future perspectives. Pure Appl. Geophys., 164, 1121–1159, https://doi.org/10.1007/s00024-007-0211-x.
Gultepe, I., and Coauthors, 2014: Ice fog in Arctic during FRAM–Ice Fog project: Aviation and nowcasting applications. Bull. Amer. Meteor. Soc., 95, 211–226, https://doi.org/10.1175/BAMS-D-11-00071.1.
Gultepe, I., and Coauthors, 2015: A review on ice fog measurements and modeling. Atmos. Res., 151, 2–19, https://doi.org/10.1016/j.atmosres.2014.04.014.
Gultepe, I., A. J. Heymsfield, M. Gallagher, L. Ickes, and D. Baumgardner, 2017: Ice fog: The current state of knowledge and future challenges. Ice Formation and Evolution in Clouds and Precipitation: Measurement and Modeling Challenges, Meteor. Monogr., No. 58, Amer. Meteor. Soc., https://doi.org/10.1175/AMSMONOGRAPHS-D-17-0002.1.
Hartl, L., M. Stuefer, and T. Saito, 2020: The mountain weather and climate of Denali, Alaska—An overview. J. Appl. Meteor. Climatol., 59, 621–636, https://doi.org/10.1175/JAMC-D-19-0105.1.
Hartmann, B., and G. Wendler, 2005: The significance of the 1976 Pacific climate shift in the climatology of Alaska. J. Climate, 18, 4824–4839, https://doi.org/10.1175/JCLI3532.1.
Hersbach, H., and Coauthors, 2020: The ERA5 global reanalysis. Quart. J. Roy. Meteor. Soc., 146, 1999–2049, https://doi.org/10.1002/qj.3803.
Huang, N. E., and Coauthors, 1998: The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proc. Roy. Soc. London, 454A, 903–995, https://doi.org/10.1098/rspa.1998.0193.
Kim, C. K., M. Stuefer, C. G. Schmitt, A. Heymsfield, and G. Thompson, 2014: Numerical modeling of ice fog in interior Alaska using the Weather Research and Forecasting Model. Pure Appl. Geophys., 171, 1963–1982, https://doi.org/10.1007/s00024-013-0766-7.
Kohonen, T., 1990: The self-organizing map. Proc. IEEE, 78, 1464–1480, https://doi.org/10.1109/5.58325.
Kohonen, T., 1991: Self-organizing maps: Optimization approaches. Artificial Neural Networks, Elsevier, 981–990, https://doi.org/10.1016/B978-0-444-89178-5.50003-8.
Kohonen, T., 2012: Self-Organizing Maps. Vol. 30, Springer Science & Business Media, 502 pp.
Kumai, M., 1964: A study of ice fog and ice-fog nuclei at Fairbanks, Alaska, Part I. U.S. Army, Cold Regions Research and Engineering Laboratory Research Rep. 150, 32 pp.
Lee, T., and T. B. Ouarda, 2011: Prediction of climate nonstationary oscillation processes with empirical mode decomposition. J. Geophys. Res., 116, D06107, https://doi.org/10.1029/2010JD015142.
Leung, A. C. W., W. A. Gough, and K. A. Butler, 2020: Changes in fog, ice fog, and low visibility in the Hudson Bay region: Impacts on aviation. Atmosphere, 11, 186, https://doi.org/10.3390/atmos11020186.
Litzow, M. A., and Coauthors, 2020: The changing physical and ecological meanings of North Pacific Ocean climate indices. Proc. Natl. Acad. Sci. USA, 117, 7665–7671, https://doi.org/10.1073/pnas.1921266117.
Liu, Y., and R. H. Weisberg, 2011: A review of self-organizing map applications in meteorology and oceanography. Self-Organizing Maps: Applications and Novel Algorithm Design, Vol. 1, IntechOpen, 253–272.
Malingowski, J., D. Atkinson, J. Fochesatto, J. Cherry, and E. Stevens, 2014: An observational study of radiation temperature inversions in Fairbanks, Alaska. Polar Sci., 8, 24–39, https://doi.org/10.1016/j.polar.2014.01.002.
Mudelsee, M., 2019: Trend analysis of climate time series: A review of methods. Earth-Sci. Rev., 190, 310–322, https://doi.org/10.1016/j.earscirev.2018.12.005.
Ohtake, T., 1967: Alaskan ice fog I: A progress report of ice fog research. Phys. Snow Ice, 1, 105–118.
Ohtake, T., and P. J. Huffman, 1969: Visual range in ice fog. J. Appl. Meteor., 8, 499–501, https://doi.org/10.1175/1520-0450(1969)008%3C0499:VRIIF%3E2.0.CO;2.
Oliphant, A. J., S. A. Baguskas, and D. M. Fernandez, 2021: Impacts of low cloud and fog on surface radiation fluxes for ecosystems in coastal California. Theor. Appl. Climatol., 144, 239–252, https://doi.org/10.1007/s00704-021-03518-y.
Oliver, V. J., and M. B. Oliver, 1949: Ice fogs in the interior of Alaska. Bull. Amer. Meteor. Soc., 30, 23–26, https://doi.org/10.1175/1520-0477-30.1.23.
Quinn, A. J., V. Lopes-dos Santos, D. Dupret, A. C. Nobre, and M. W. Woolrich, 2021: EMD: Empirical mode decomposition and Hilbert-Huang spectral analyses in Python. J. Open Source Software, 6, 2977, https://doi.org/10.21105/joss.02977.
Rae, R., 1954: Meteorological activities in the Canadian Arctic. Arctic, 7, 119–128.
Rex, D. F., 1950: Blocking action in the middle troposphere and its effect upon regional climate. Tellus, 2, 275–301, https://doi.org/10.1111/j.2153-3490.1950.tb00339.x.
Robinson, E., W. C. Thuman, and E. J. Wiggins, 1957: Ice fog as a problem of air pollution in the Arctic. Arctic, 10, 88–104, https://doi.org/10.14430/arctic3756.
Schmitt, C. G., M. Stuefer, A. J. Heymsfield, and C. K. Kim, 2013: The microphysical properties of ice fog measured in urban environments of interior Alaska. J. Geophys. Res. Atmos., 118, 11 136–11 147, https://doi.org/10.1002/jgrd.50822.
Sheridan, S. C., and C. C. Lee, 2011: The self-organizing map in synoptic climatological research. Prog. Phys. Geogr., 35, 109–119, https://doi.org/10.1177/0309133310397582.
Sousa, P. M., D. Barriopedro, R. García-Herrera, T. Woollings, and R. M. Trigo, 2021: A new combined detection algorithm for blocking and subtropical ridges. J. Climate, 34, 7735–7758, https://doi.org/10.1175/JCLI-D-20-0658.1.
Stallone, A., A. Cicone, and M. Materassi, 2020: New insights and best practices for the successful use of empirical mode decomposition, iterative filtering and derived algorithms. Sci. Rep., 10, 15161, https://doi.org/10.1038/s41598-020-72193-2.
Thuman, W. C., and E. Robinson, 1954: Studies of Alaskan ice-fog particles. J. Atmos. Sci., 11, 151–156, https://doi.org/10.1175/1520-0469(1954)011%3C0151:SOAIFP%3E2.0.CO;2.
Vas, D. A., S. E. Peckham, C. G. Schmitt, M. Stuefer, R. G. Burgener, and T. E. Wong, 2021: Ice fog monitoring near Fairbanks, AK. Engineer Research and Development Center (U.S.) Tech. Rep. ERDC/CRREL TR-21-2, 57 pp.
Vettigli, G., 2018: Minisom: Self organizing maps. GitHub, https://github.com/JustGlowing/minisom/.
WMO, 2017: International Cloud Atlas: Manual on the observation of clouds and other meteors. WMO-407, accessed 6 June 2023, https://cloudatlas.wmo.int/en/ice-fog.html.
Wu, Z., N. E. Huang, S. R. Long, and C.-K. Peng, 2007: On the trend, detrending, and variability of nonlinear and nonstationary time series. Proc. Natl. Acad. Sci. USA, 104, 14 889–14 894, https://doi.org/10.1073/pnas.0701020104.
