Observations of 4–5 h oscillation periods in virtual height during the fall seasons of 2020–2023 across the eastern United States

Since the early 1960s, many studies have documented ionospheric virtual height oscillation with periods ranging from 30 min to 3 h. However, in this investigation, we find a consistent, much larger period in the oscillations of virtual height, as estimated using Coastal Ocean Dynamics Applications Radars (CODARs), whereas few previous studies have consistently observed periods of oscillation in this range. The purpose of this investigation is to investigate day-to-day bottomside variability, and we report a consistently observed 4–5-h period of virtual height oscillation, along with providing potential physical interpretations of the observations. We use bistatic high-frequency data from CODARs, collected nearly continuously between 2020 and 2023 at a 2-min time cadence. The frequency modulated continuous wave (FMCW) waveforms of these radars are suitable as a single-frequency oblique ionospheric sounder. We estimate the group delay from which the virtual height of an ionospheric layer can be estimated. We present statistical results from several virtual height oscillations with varying periods across September, October, November, and December during 2020–2023. For each year, we compare three to four CODAR transmitters located in the Eastern United States, operating at 4–5 MHz. The receiver is located near Clemson University. We frequently observe ( ≈ 25% of the time) the 4–5-h period of virtual height oscillation each September through December over the 4-year span, which has a longer period than the standard categorization of a Large Scale Traveling Ionospheric Disturbance (LSTID). To further understand the characteristics of these structures, we separate them into both propagating and non-propagating structures. We found that for all 4–5-h structures with southward propagation, the night prior was geomagnetically active. We also find significant similarities between the non-propagating structures and two ionospheric phenomena, the midnight collapse and post-midnight enhancement.

Day-to-day ionospheric variability remains an area of active investigation because it corresponds to the “weather” observed, while the “climatology” of the ionosphere is reasonably well understood in empirical models, such as the International Reference Ionosphere. The main goal of many day-to-day studies is to better understand the drivers of that variability and quantify deviations from the mean behavior. It is well documented that ionospheric variability is dependent on many factors, including local time, season, solar cycle, geographic location, solar photon flux, geomagnetic activity, and terrestrial weather (Forbes et al., 2000; Rishbeth and Mendillo, 2001). These types of drivers can cause variability on time scales ranging from minutes to years. For the purposes of this study, we will primarily focus on shorter-term variability, specifically day-to-day and hour-to-hour.

Since the early studies of ionospheric variability, it has been well documented that a large contributor to variability on day-to-day and shorter timescales is caused by Traveling Ionospheric Disturbances (TIDs) (Hines, 1960; Hunsucker, 1982; Hocke and Schlegel, 1996). TIDs are wave-like fluctuations in the ionospheric electron density that can alter the reflection height of high-frequency (HF) radio waves propagating through the ionosphere (Davies, 1990). TIDs are commonly categorized into two groups: Large Scale TIDs (LSTIDs) and Medium Scale TIDs (MSTIDs). LSTIDs have periods of 30–120 min, horizontal velocities of 300–1,000 m/s, and horizontal wavelengths of 300–3,000 km. In comparison, MSTIDs have periods of 10–60 min, horizontal velocities of 50–300 m/s, and horizontal wavelengths of 100–300 km (Hocke and Schlegel, 1996). In general, TIDs are thought to be manifestations of atmospheric gravity waves (AGWs) propagating up to ionospheric heights or the consequence of instability processes. One of the main drivers of LSTIDs is thought to be geomagnetic activity (Richmond, 1978; Hunsucker, 1982; Borries et al., 2009; Crowley et al., 2016), while MSTIDs are thought to be caused by more localized AGWs (Hines, 1960; Yeh and Liu, 1974; Hunsucker, 1982; Hocke and Schlegel, 1996) or electrodynamics processes typically associated with the Perkins instability (Perkins, 1973; Kelley, 1989; Chou et al., 2017).

Although most LSTIDs propagating southward are thought to be from auroral sources, LSTIDs propagating westward and northwestward have been shown to be caused by the passing of the solar terminators (Galushko et al., 1998; MacDougall and Jayachandran, 2011; Song et al., 2013), with a majority of these effects happening due to the sunrise terminator rather than the sunset terminator (Segarra et al., 2025) predominantly between the hours of 0600–1,200 local time. Another possible driver of TIDs is terrestrial weather. Several studies have shown the generation of concentric TIDs from events such as earthquakes, volcanic eruptions, cyclones, and severe thunderstorms (Hung et al., 1978; Taylor and Hapgood, 1988; Vadas and Liu, 2009; 2013; Liu et al., 2011; Vadas et al., 2012; Nishioka et al., 2013; Lu et al., 2024). Recently, more work has been done on the generation of TIDs due to weather fronts, which are more common in the midlatitudes in comparison to most of the large-scale weather events listed above (Kogure et al., 2024). However, it remains unresolved how the coupling processes at the lower altitudes can drive the structure of TIDs at higher altitudes in the F-region ionosphere. In addition, understanding how each of these different source mechanisms contributes to day-to-day ionospheric variability requires further observational investigations and analysis of existing datasets.

The purpose of this study is to further expand upon the work done in Markowski et al., 2025; Kaeppler et al., 2022 by applying their methodology to a significantly larger dataset to quantify the day-to-day F-region ionospheric variability. Although the Coastal Ocean Dynamic and Application Radars (CODARs) have been used extensively in the ocean sensing community (Georges, 1985; Paduan and Rosenfeld, 1996), they are seldom used for ionospheric research. Previously, Kaeppler et al., 2022 showed that CODARs can be used as bistatic single-frequency oblique ionospheric sounders, providing a relatively low-cost solution to increase the spatial and temporal coverage of existing ionosonde networks (Reinisch et al., 2018). Troyer et al. (2025), also demonstrated the ability of the CODAR network to be used for ionospheric research and showed how it could be used to detect wave activity caused by both natural and man-made events.

As an expansion to Kaeppler et al., 2022, Markowski et al., 2025 demonstrated how the CODAR network could be used to track an LSTID moving across the Eastern United States and derive the structures’ horizontal velocity and propagation direction in a generalized sense using a potentially arbitrary number of transmitters and receivers. We present an analysis of CODAR observations from September, October, November, and December in 2020, 2021, 2022, and 2023. First, we present the results of an introductory day-to-day variability study of the oscillations of the nighttime ionospheric F-region virtual height. Next, we focus on a consistent period of oscillation (4–5-h) observed throughout the study, which, to the best of our knowledge, has not been previously reported. Finally, we end with a discussion of the possible drivers of these oscillations. Currently, disturbances with periods that are subharmonics of a solar day (24, 12, 8, and 6 h) are generally attributed to atmospheric tides, whereas disturbances with periods less than 2 h are commonly classified as TIDs. To the best of our knowledge, this is one of the first studies to document periods of virtual height oscillations between these two categories, 4–5-h, that are consistently observed over the Eastern United States. Similar to that done by Lynn et al. (2016), in this work we also explore the possibility that these oscillations are caused by either propagating or non-propagating disturbances, which, in both cases produce results similar to standard TID-induced oscillations. However, the drivers in both cases appear to be different.

Measurement techniques

For this investigation, we are using twelve CODAR transmitters as bottomside ionospheric sounders, from which we extract a time series of virtual heights. Details regarding the data acquisition, calibration methodology, basic signal processing, group delay extractions, and virtual height calculations can be found in Markowski et al., 2025; Kaeppler et al., 2022. The CODAR transmitters selected for this study were chosen based on data availability, location to the receiver, and transmission frequency. Note that for all stations used in this study, a transmission frequency between 4 and 5 MHz was selected which we assume probes similar isodensity contours. Based on geographic location and transmission frequency, the height at which our signal is reflected is similar for all stations used in each comparison. The twelve stations selected have the following call signs: NAUS, AMAG, ANWR, PINS, MARA, VENI, NAPL, BRIG, WILD, CORE, MRCH, and LISL. Due to our lack of control over the CODAR transmission frequencies and the stronger D-region absorption of lower frequencies during the day, we are limited in our study to only consider nighttime disturbances. Table 1 shows each station’s location, operating frequency, and distance from the CODAR transmitter to the receiver at the Clemson Atmospheric Research Laboratory, CARL. All ground distances shown in this paper were calculated using the Haversine formula to account for the curvature of the Earth. Due to data availability, stations NAUS, PINS, MARA, NAPL, and LISL were used in 2020 and 2021, while stations AMAG, PINS, ANWR, VENI, BRIG, WILD, MRCH, and CORE were used in 2022 and 2023. However, only three to four stations were selected per day. For simplicity, we group our transmitters into four geographic regions: Texas, Florida, Northeast, and Mid-Atlantic. Each region is represented by a circle in Figure 1. From this point forward, we will refer to transmitters by their geographic region rather than by their callsign. The operating frequencies listed in Table 1 are those used by the stations at the time of data collection and may not reflect their current operating frequencies. For more information regarding the current status of all CODAR stations please see https://hfrnet.ucsd.edu/sitediag/stationList.php.

For increased data coverage when necessary, we also use the following two ionosondes: Wallops Island, VA (WP937) and Austin, TX (AU930), both with operating frequencies of 1.0–30.0 MHz. For this study only frequencies of 2.0–3.0 MHz were used. It should be noted that the ionosonde at Eglin Air Force Base, FL (E.G., 931), was initially considered; however, due to poor data quality, it did not improve the temporal coverage of this study and was therefore not included.

Categorization of trends

To identify the 4–5-h oscillations, this study follows a similar process to that used in Markowski et al., 2025. This methodology is shown in Figure 2 below. First, for each transmitter, we use a time series of virtual heights only for days with data spanning the entire period of interest (panel A Figure 2) to avoid false positives in the Lomb-Scargle calculation. We only consider hours from 0100 to 0900 UT (2,100–0500 LT 2000 to 0400 LT after daylight savings) to avoid potential oscillations caused by the passage of the terminators and by daytime D-region absorption of lower-frequency transmissions. To remove any possible diurnal trends, we subtract a reconstructed 12-hour-period waveform from the data and then apply a Gaussian filter (dotted lines in panel B of Figure 2). Note that in previous work published by Markowski et al. (2025), an 8-hour-period waveform was used to remove possible diurnal trends rather than the 12-hour-period waveform used in this study. This updated methodology did not alter the periods of interest for this study. We then perform a Lomb-Scargle calculation to identify any remaining dominant periods of oscillation and flag the days with periods of 2–3-h and 4–5-h, both with false alarm probabilities below 0.1% (VanderPlas, 2018). The false alarm probability here refers to the probability of measuring a peak at some height assuming there is no periodic signal and the data consists of just Gaussian noise (Astropy et al., 2013; Astropy et al., 2018; Astropy et al., 2022).

To determine the dominant periods of oscillation smaller than 4–5-h, each day underwent recursive subtraction until the remaining signal was just noise. An example of this can be found in the Supplementary Material. To reduce noise in the calculations, we reconstruct the waveforms (panel C of Figure 2) using the dominant periods identified in the periodogram and perform a cross-correlation analysis to obtain the time delays (panel D of Figure 2). Since this study focuses on waveforms with periods of 4–5-h (and 2–3-h for comparison), only the period of interest is used in the reconstruction. Several days also had dominant periods of oscillations of 1–2-h, which indicate MSTID activity; however, those occurrences are outside the scope of this investigation.

Once we have identified the days with a 4–5-h oscillation, we classify them as propagating, non-propagating, or having insufficient information. In order for a day to be marked as having a propagating or non-propagating structure the following criteria must be met: the day must have data from 3 or more transmitters, have periods of oscillation between 4-5-h, and the time series must have correlation coefficients higher than 0.5. If all qualifications were met, a cross-correlation analysis was performed on the reconstructed waveforms. Days were classified as non-propagating if all traces had a time delay of 0 min, and as propagating if any trace had a nonzero time delay. An example of the propagating case is shown in Figure 2. An example of the non-propagating case is shown in Figure 3. In this figure, all panels follow the same setup as in Figure 2; however, they now show a case with no time delay. These panels show the virtual height time series from 3 stations, flagged with a 4–5-h period before (panel A) and after (panel B) the 12-h trend was removed and the reconstructed waveforms using the dominant period of oscillation between 4-5-h only (panel C). Panel D is the output from the cross-correlation analysis done for the Northeast and Mid-Atlantic regions (black) and the Northeast and Florida regions (red). This day was flagged as non-propagating because the time lag calculated from this cross-correlation analysis was 0 min.

All remaining days were marked as having insufficient information. The time delays of all days marked as propagating were used to calculate the horizontal phase speed (V h) and propagation direction (ϕ) of each disturbance. These values were calculated using the following equations described in Markowski et al., 2025: where V x and V y are defined as, In Equation 2, n is the number of stations used, with n ≥ 3, x 1 i and y 1 i is the distance in the x and y direction between the midpoints of transmitter 1 and transmitter i, and t 1 i is the time delay between transmitter 1 and transmitter i. Transmitter 1 is the station first affected by the disturbance.

Figures 4–7 show the dominant periods of virtual height oscillation for the fall and winter months of 2020, 2021, 2022, and 2023 over the Northeast, Mid-Atlantic, Florida, and Texas regions. Note that this includes data collected from the CODAR transmitters, and when necessary, the ionosondes located in the same regions. To create the summary figures, one transmitter per region was selected to represent the entire region, based on data quality and coverage. The process described in Section 2.2 was then applied to each day and transmitter. Therefore, each vertical bar shown represents the output from the normalized periodogram. The white space represents a lack of available data for that specific region. The lack of data could be due to little to no F-region return or technical issues with our instrumentation. It is clear from these figures that the dominant period of oscillation was 4–5-h, with a secondary peak at the 2–3-h mark across all regions of interest. The figures also highlight variability from day-to-day. For example, in September of 2020, in Figure 4 the most dominant daily period changes each day. Although the same oscillation periods are consistently observed, they do not occur every day, nor do they always occur simultaneously across regions.

To quantify the occurrence of this trend, we now only consider the days with a 4–5-h period of virtual height oscillation. In doing so, we flagged 535 of 1,348 possible days with a virtual height oscillation period of 4–5-h, corresponding to an average occurrence rate of 39% across the entire dataset. To get the total number of days flagged with 4-h periods of oscillation, all flagged days and total possible days with sufficient data were summed across transmitters, giving the final counts of 535 days flagged out of 1,348 possible days. For example, if we had perfect data coverage over the duration of our study, the total would be 1,952 possible days. Note that, for days that counted towards the total possible days, sufficient data was available we had a strong return signal. However, a 4–5-h period may or may not have been detected. For more information regarding the usable days in this study please see the Supplementary Material. To further consider how robust this 4–5-h trend is, Figure 8 categorizes this occurrence rate by transmitter location, month, and year. Here, the data are color-coded by transmitter region, as described in Section 2.1. To create Figure 8, the transmitter with the strongest return signal and most observable minutes for each region was selected to represent that specific month. To calculate these rates, each transmitter was counted separately. For example, in October 2020, the MARA (Florida) virtual height trace flagged 11 out of 20 days with a 4–5-h trend, while the LISL (Mid-Atlantic) virtual height trace flagged 8 out of 24 days with a 4–5-h trend. Therefore, the occurrence rate for the Florida station in October 2020 was 55% and the occurrence rate for the Mid-Atlantic station was 33%. The errors in Figure 8 are calculated using the Standard Error of Proportion. To help increase the monthly data coverage, we attempted to combine data from two transmitters if they had ground distances of approximately 50 km; however, this did not drastically improve our coverage. A summary figure showing which transmitter was used in each month is available in the Supplementary Material.

From Figure 8, it can be seen here that we observe this trend more than 25% of the time during this study. It should be noted that we found no correlation between the monthly occurrence rates and F10.7, suggesting that the monthly occurrence rates are independent of solar activity. For comparison, we also consider the occurrence rates of a 2–3-h period of virtual height oscillation in Figure 9, which is more commonly believed to be LSTID activity (Hunsucker, 1982). By doing so, we found 630 days out of 1,348 possible days with this trend, giving us an average occurrence rate of 48%. From this figure we can see that our 4–5-h period of virtual height oscillation happens nearly as often as our 2–3-h period of virtual height oscillation during our period of interest across all regions.

The data used to generate (Figure 8) are the initial count of days with a 4–5-h oscillation period. To further investigate these trends, we now consider days with propagating and non-propagating disturbances, using the same process as described in Section 2.2. Following that process, we identified 9 days that met the criteria for a propagating disturbance and 20 days that met the criteria for a non-propagating disturbance. A more detailed discussion of the high attrition rate can be found in Section 4.3. For all cases of westward/northwestward propagation, it was geomagnetically quiet, whereas for all cases of southward propagation, it was geomagnetically active. This is shown in Figure 10, which plots the propagation direction versus phase speed, calculated using Equation 1, for these structures. Each point is color-coded by the nightly-averaged Kp index. The nightly averaged Kp value here refers to the average Kp calculated between 1,400–0000 UT the day prior to our observed activity. For each propagating structure, the average wavelength was approximately 5,000 km. Now, if we consider our 20 non-propagating cases, we find that it was geomagnetically quiet (Kp ≤ 3.5 and SME ≤ 250 nT) in all 20 cases. This is shown in Figure 11, which plots the nightly-averaged Kp and SME values.

Possible drivers of propagating cases

In order to determine the possible drivers of our 9 propagating disturbances, we first consider their relationship with geomagnetic activity. As previous work has shown, one of the main drivers of LSTIDs is thought to be geomagnetic activity (Richmond, 1978; Crowley et al., 1987; Hocke and Schlegel, 1996; Hunsucker, 1982; Tsugawa et al., 2004; Ferreira, Arthur Amaral et al., 2020; Kishore and Kumar, 2023). Although these disturbances occur on a much larger timescale than the standard definition of LSTIDs, they are similar in nature, causing fluctuations in ionospheric virtual height, with several propagating southward. To make this comparison, Figure 10 shows the direction of propagation of the 4–5-h disturbance as a function of horizontal phase speed, with each point color-coded by nightly averaged Kp value (King and Papitashvili, 2005). Figure 10 also shows that all days with southward/southwestward propagation, which is the dominant propagation direction for geomagnetically induced LSTIDs, were nights with strong geomagnetic activity, Kp ≥ 3, in the 10 hours leading up to the disturbance. This indicates a correlation between geomagnetic activity and these propagating structures. All southward-propagating structures also had speeds between 300 and 800 m/s, which is within the standard range of accepted phase velocities for LSTIDs (Hocke and Schlegel, 1996). Each disturbance with a southward propagation direction also had a maximum SME value (Gjerloev, 2012) exceeding 750 nT sustained between 1,400 and 0000 UT on the night prior to the disturbance. This result is similar to the outcome of a statistical study performed by Segarra et al., 2025, who found that majority of the southward propagating LSTIDs detected between 2014 and 2023 were linked to auroral activity.

Due to the speed and propagation direction of our events, the Kp and SME values from the previous night, in UT, were selected, since a geomagnetic event could not have caused these oscillations so close to their onset. The remaining cases all had westward/northwestward directions of propagation and the 6 h leading up to the event were geomagnetically quiet. For all cases of westward propagation except one, there were eastward-moving thunderstorms covering our entire region of interest. Several previous studies have shown the generation of atmospheric gravity wave-induced TIDs due to tropospheric weather systems such as hurricanes and or thunderstorms (Hung et al., 1978; Kogure et al., 2024; Lu et al., 2024). Therefore, we believe these weather systems may have caused our large perturbations however, further modeling studies are needed to test these theories.

To the best of our knowledge only a few studies have found oscillations in virtual height within this period range. Two examples are from Lei et al. (2008) and more recently Gong et al. (2025b). Lei et al. (2008) observed ionospheric oscillations with periods of 4–5-h during the 15 December 2005 geomagnetic storm, which reached a maximum Kp of 8. In this work, they determined that the primary driver of these oscillations is neutral winds rather than ambipolar diffusion and photochemical processes. Gong et al. (2025b) detected the first 4.8-h tide in data collected from Arecibo in 2 March0,215. In this study, they suggest the generation of the 4–8-h tide is caused by a very strong geomagnetic storm. Although the environmental conditions were very different from this study, we wanted to bring attention to both studies, as the periods detected in both are very similar to our own. However, we observed them during both geomagnetically active and quiet times.

Another possible mechanism is the solar terminator, which shows clear day-to-day variability. While TIDs associated with the solar terminator have been observed in other studies, there may be additional secondary impacts generated by the terminator. In particular, Hedlin et al. (2018) made observations of surface pressure waves with periods of 2–6 h associated with the solar terminator. These wave periods are consistent with our day-to-day results, albeit in terms of electron density rather than surface pressure variations. Although observed at the surface, this investigation discusses how these surface pressure perturbations could be a source of gravity waves that could propagate into the thermosphere, thereby producing TIDs with periods similar to the 4–5-h period of our observations. Modeling is required to test whether these vertical coupling processes can account for the observed impact on electron density.

Possible drivers of non-propagating cases

To determine the possible drivers of these events, we consider the automatically scaled values for foF2 and hmF2 from the WP 937, E.G., 931, or AU 930 ionosondes obtained from the GIRO online database (Reinisch and Galkin, 2011), depending on data availability. Note that foF2 and hmF2 refer to the critical frequency of the F2 layer and the height at which the maximum electron density occurs in the F2 layer. In doing so, we can further categorize the data into three groups. In the first group, we consider two scenarios. For group 1A, the foF2 values decrease before 0100 UT as expected. The rate of decrease then slows, and around local midnight they undergo a final decline before leveling off. This pattern of the foF2 values is evident in the CODAR-derived virtual height time series, as the virtual heights inversely follow this pattern. Based on the timing of these decreases, we suggest that they may be responsible for the 4–5-h fluctuations observed in the CODAR-derived time series.

In category 1B, the foF2 values remain consistent between 0100 and 0900 UT. However, the hmF2 values oscillate, indicating a change in layer height rather than in electron density, which again is consistent with the pattern of the CODAR-derived virtual heights. Across all 20 cases, there was no day on which both the foF2 and hmF2 values remained consistent between 0100 and 0900 UT.

In category 2, the foF2 values decrease gradually before leveling off, while the hmF2 values exhibit a sudden decrease between 0300 and 0400 UT. In addition to the ionosonde data, the CODAR-derived time series of virtual heights also clearly shows a dip in virtual height beginning around local midnight (0400 UT), which lasts approximately one to 4 hours. After this decrease, the virtual height begins to rise again, reaching a peak around 0700 UT before descending again. Therefore, based on the 4–5-h dominant period observed in the periodogram, the patterns of foF2, hmF2, and CODAR-derived virtual heights in these cases could be examples of the midnight collapse. The midnight collapse is an ionospheric phenomenon occurring in the F-region ionosphere, categorized by the somewhat sudden decrease in hmF2 around local midnight while the electron density remains the same, decreases or in some cases increases (Nelson and Cogger, 1971; Behnke and Harper, 1973; Crary and Forbes, 1986; Harper, 1979; MacPherson et al., 1998; Vlasov et al., 2005; Seker et al., 2009; Kelley, 1989; Lynn et al., 2016). The early studies suggested that the reversal of the meridional winds primarily caused the midnight collapse. They concluded that this reversal induced a downward motion of ions and was also related to a sharp increase in the F-region temperature near midnight (Gong et al., 2025a). More recently, it has been found that although the meridional wind reversal is still a primary driver, the midnight collapse occurs due to a combination of different processes which all control the redistribution of plasma in the F-region, such as electric fields driving vertical ion drift, field-aligned ion motion due to neutral winds and diffusion, and tidal waves in the meridional winds (Gong et al., 2012; Dandenault and Richards, 2015; Gong et al., 2026). However, much work is still being done today to better understand the complex mechanisms driving these events. An example of the foF2, hmF2, and CODAR-derived virtual heights for this group can be seen in panels B and D of Figure 11. It should be noted that this ionospheric phenomenon is very common in the midlatitudes. Recently, Gong et al. (2025a) found the midnight collapse occurred in 94% of the measurements at Arecibo, a standard midlatitude station, taken from 1971 to 2019. A follow-up study performed by Gong et al. (2026) compared the occurrences and morphologies of this event during both geomagnetically active and quiet times. In doing so, they found that during disturbed nights, this phenomenon occurred 97.2% of the time, and during quiet times, it occurred 92.2% of the time.

In the third category, the hmF2 values remain relatively consistent. However, after the initial decrease, the foF2 values level off before increasing again between 0400 and 0600 UT. Using the relation N m F 2 = f o F 2 2 / 80.6 (Chen, 2015) where NmF2 here refers to the maximum electron density of the F2 layer, we believe this increase in the foF2 value and corresponding decrease in CODAR-derived virtual height may be an example of the post midnight enhancement. This ionospheric phenomenon is characterized by an increase in midlatitude electron density at local midnight (Davies et al., 1979; Zhao et al., 2022; Zhang et al., 2024; El-Desoky et al., 2024). This increase is also reflected in the CODAR-derived virtual height time series, with a reduction in height corresponding to the increase in NmF2. Similar to the midnight collapse, the mechanisms responsible for this are rather complicated and still under investigation today. One suggested cause of this phenomenon is the downward plasma flux from higher altitudes diffusing into the nighttime F layer along the geomagnetic field lines. This downward flux can compensate for the plasma recombination loss. Therefore, when the accumulation of plasma exceeds the recombination loss, the electron density increases (Zhao et al., 2022; Evans, 1965; Burns et al., 2008; Liu et al., 2013; Le et al., 2014; Chen et al., 2015). Two other possible causes are due to electric fields or equatorward neutral winds (Wang et al., 2022; Li et al., 2020). However, whether all of these processes independently cause this or work in tandem is still an open question. An example of the foF2, hmF2, and CODAR-derived virtual heights for this group is shown in panels A and C of Figure 11. Therefore, this 4–5-h trend is likely due to midlatitude phenomena, the midnight collapse or post-midnight enhancement, as over half of the cases fell into either category.

Overlapping days

One extremely important aspect of our study is the large drop off in data from the initial count of occurrences to the count of propagating vs. non-propagating structures. To demonstrate this decrease in data from Figures 8–10 and Figure 11, we show a comparison between the number of occurrences of the 4–5-h trend and the number of days with the 4–5-h trend simultaneously occurring at a minimum of three stations. In Table 2, the column labeled Total number of occurrences is found by summing together the days flagged with a 4–5-h trend from a transmitter in each available region as described in Section 3. The column labeled, Number of days with 3 stations overlapping, represents the number of days in each month that had 3 or more stations flagged for the 4–5-h period oscillation. The final column shows the number of days each month on which the 4–5-h trend occurred simultaneously at three or more stations. However, the time series of virtual heights at each station were uncorrelated across the region. Note that if only three regions had data and one of the three virtual height traces was uncorrelated, the entire day was marked as uncorrelated and removed. If all four regions had data on a given day, only three of the four virtual height traces needed to be correlated to study the day further. Not represented in this table is the number of days on which the 4–5-h trend occurring in 2 of the 4 regions. All days in that scenario had to be removed, as the technique used to derive propagation direction and phase speed requires at least 3 stations.

As discussed in Forsythe et al. (2020), the horizontal correlation length in the midlatitudes, although more stable than in both high and low latitudes, remains quite variable with respect to magnetic local time (MLT). The correlation length in both the zonal and meridional directions in the midlatitudes during winter ranges from approximately 5–15° (≈ 500–1,600 km in latitude and 400–1,300 km in longitude), depending on the time of day. Our observations suggest that, in general, the correlation length of the ionosphere for the 4–5 h period oscillations is smaller than this scale size. The significant attrition of data to 29 events out of over 1,000 further suggests that correlation over 1000s of km, i.e., across the Southeastern United States, is an unusual occurrence. We note that the correlation over large geographic distances and southward propagation is consistent with auroral-driven LSTIDs, which we expect to have scale sizes of the order of 1000s of km (the radius of curvature of the auroral oval). In fact, Frissell et al. (2022) has shown that during intervals of geomagnetic activity, LSTIDs can be continental scale in size, as deduced from amateur radio spotting data. Further investigation is warranted to determine the ionospheric correlation length using observations at different characteristic oscillation periods.

Potential sources of error and limitations

Although our technique has been shown to increase the spatial and temporal coverage of the Eastern United States, there are still some limitations. As discussed in (Markowski et al., 2025), the CODAR stations operate at fixed frequencies that we cannot adjust, which limits our analysis to nighttime effects only. Another limitation arises from the robust criteria set for our study, which significantly reduces the number of usable days. For example, a day could only be classified as having a propagating or non-propagating trend if data were available from at least three stations, with a minimum ground distance of 500 km between at least two of them. This requirement explains the discrepancy between the number of days shown in Figures 8–10. Finally, our last limitation occurs when no data are available from stations in Texas. The inferred propagation direction may be skewed as it is difficult to determine East or West propagation. For example, with data only from stations in the Northeast, Mid-Atlantic, and Florida, a westward-propagating structure might appear to propagate southward because of the limited station coverage.

The purpose of this investigation was to quantify day-to-day ionospheric variability using observations from Coastal Oceanographic Dynamics Applications Radars (CODARs) between 4 and 5 MHz. We presented a study covering September through December of 2020, 2021, 2022, and 2023, comparing CODAR- and ionosonde-derived virtual heights from four transmitter regions with large spatial distributions across the Eastern United States. Our observations are limited to local nighttime conditions. A time-series analysis was performed to extract the dominant oscillation periods in the virtual height. We find a prominent 4–5-h oscillation in ionospheric virtual height that, to the best of our knowledge, has not been previously reported. Using a TID propagation method developed by Markowski et al. (2025), we determined whether these 4–5-h oscillations were propagating and non-propagating disturbances based on the calculated time delay between virtual height traces. The observations were limited to nighttime conditions.

We summarize the key findings of our investigation: We find that 535 of 1,348 possible nights had a 4–5-h oscillation, corresponding to an occurrence rate of 39%.

We also found 630 of 1,348 possible nights had a 2–3-h oscillation, which could be attributed to a more traditional definition of LSTIDs based on period. This occurrence rate was 48%.

We also found 29 cases in which the 4–5-h virtual height oscillations were correlated across at least 3 stations, which we further analyzed. We found 9 cases of propagating disturbances and 20 cases of non-propagating disturbances.

For all southward-propagating cases, we found it was geomagnetically active the day prior, Kp ≥ 3 and SME ≥ 750 nT, indicating a correlation with these propagating structures and geomagnetic activity.

For most non-propagating cases we demonstrate strong similarities to the midlatitude ionospheric phenomena known as the midnight collapse and midnight enhancement.

Finally, we also observed a lack of consistent correlation between virtual height time series across regions, consistent with previous studies demonstrating that horizontal correlation lengths are highly variable.

We hypothesize that some of the 4–5-h period virtual height oscillations have signatures similar to the midnight collapse, midnight enhancement, or possibly secondary impacts from the solar terminator. However, testing these hypotheses requires modeling efforts beyond the scope of this investigation. We also encourage future studies to consider oscillation periods of this size rather than filtering out periods larger than 3-h.