Introduction
Ionosondes for studying the ionosphere were invented in the late 1920s, and
their worldwide implementation started in the 1940s as military shortwave
communication became important. An ionosonde is a radio echo instrument
transmitting radio waves of alternating frequency, receiving reflections from
the ionosphere, and measuring the travel times of the waves at different
frequencies. The return power vs. frequency and travel time is presented as
graphs called ionograms, from which information on the structure of the
ionosphere can be derived. The travel times Δt are presented as
virtual heights h′=cΔt2; the true reflection heights are
different due to the significant radio-wave refraction close to the critical
ionospheric plasma frequencies. Up to four distinct layers, called the D, E,
F1 and F2 regions, respectively, in increasing altitude order, can be
identified in ionograms, depending on location, season and time of day.
Historically, the shapes of ionograms are described by a number of
standard parameters . This is useful in order to characterise the main features
of the ionosphere, such as what frequencies are usable for long
distance communication. These parameters have been read out manually
at several observatories since the 1950s, a procedure referred to
as ionogram scaling. Thus, time series of continuous ionospheric
observations span several solar cycles, and in order to monitor
long-term environmental changes these observations must be continued
in a consistent way without significant methodological changes.
However, manual scaling of ionograms requires substantial work efforts
and the results are subjective, so development of automatic scaling
routines has started and progressed together with recent developments
in computer performance.
One such routine is Autoscala , which has
been developed at the Italian National Institute of Geophysics and
Volcanology (INGV) in Rome. Autoscala is a program able to perform an
automatic scaling of vertical soundings, giving as output the main
ionospheric characteristics
and an estimation of the electron density profile
. It is based on an image recognition technique
and can run without polarisation information, which allows the
algorithm to be applied to any kind of ionosonde
.
Autoscala works by defining a set S of N pairs of empirical curves
S≡Ti[o],Ti[x],
i=1,2,…,N, fitting the typical shape of the F2 ordinary and
extraordinary traces. For each pair of curves Ti[o],
Ti[x] the local contrast C with the recorded ionogram is
calculated, making allowance for both the number of matched points and their
amplitude. The pair of curves Ti[o], Ti[x] having
the maximum value of C is then selected. If this value of C is greater
than a fixed threshold Ct, the selected curves are considered as
representative of the traces. The values of the critical frequencies are thus
obtained from the selected curves. If C does not exceed Ct, the
routine assumes that the F2 trace is not present on the ionogram. With
a similar procedure, the F1 and Es traces are also detected.
The accuracy of the Autoscala output has been tested against manually
validated data at low and middle latitudes. Specifically, the
low-latitude stations considered to test Autoscala were those of
Tucumán (26.9∘ S, 294.6∘ E), Argentina
, and São João do Cariri
(7.5∘ S, 323.8∘ E), Brazil
. The corresponding results showed that the
abilities of Autoscala are fairly good and reliable, except for the
cases of ionograms characterised by the presence of F1.5 or F3 layers,
as was illustrated by .
The mid-latitude stations considered to test Autoscala were those of
Rome (41.8∘ N, 12.5∘ E) and Gibilmanna
(37.9∘ N, 14.0∘ E), Italy
,
Moscow (55.5∘ N, 37.5∘ E), Russia ,
Warsaw (52.2∘ N, 21.1∘ E), Poland ,
Delaware (43.0∘ N, 278.8∘ E), Canada
and Boulder (40.0∘ N, 105.3∘ W), Colorado, USA .
The corresponding results showed that the output given by Autoscala is
reliable and accurate and, at times, better than that given by the
Automatic Real-Time Ionogram Scaling with True-height (ARTIST) system
.
Simplified schematic diagram of the SGO ionosonde system and
data flow. Note that the transmitter and receiver station are
separated by almost 1 km and have no direct electronic
connection. Transmission and reception of the FM CW chirps both
start by GPS synchronisation.
Left column: schematics of transmitter station with exciter (direct
digital synthesizer) and power amplifier. Filters are not shown. The
transmitter feeds a wide-band rhombic antenna.
Centre column: schematics of one receiver channel. Signals from all
receiver antennas are combined into two channels, one per
polarisation. The local oscillator is identical to the exciter. Raw
data are saved in binary files containing four real (or two complex)
vectors: two linear polarisations, samples of in-phase and
quadrature-mixed filtered signals.
Right upper: the real-time
processing computer reads the raw data, combines the polarisations
into O-mode circular polarisation (applying phase corrections), and
produces ionograms. O-mode ionograms are stored and those from each
full hour are interpreted manually.
Right lower: during the described test phase, O- and X-mode ionograms
were uploaded to INGV in the raw data format (RDF) required by
Autoscala.
Experience from high-latitude stations, including the two different
automatically scaling ionosondes at the EISCAT transmitter site at
Ramfjordmoen, Norway (69∘ N latitude),
a Dynasonde and a Digisonde see
e.g., shows that in the auroral zone automatically
scaled parameters may differ significantly. These two instruments,
however, use quite different echo detection and selection techniques
and the results depend much on how noise and interference are
rejected.
The present comparison is based on ionograms from the ionosonde at the
Sodankylä Geophysical Observatory (SGO, 67∘ N,
26∘ E, 64.1∘ CGMLAT). In contrast to the
two instruments above, the SGO ionosonde is a so-called chirp
sounder. This technique is based on a continuous wave (CW) transmission
of increasing frequency instead of radar pulses.
Example of SGO ionogram scaling.
Upper: manual scaling interface with a typical Alpha Wolf O-mode
ionogram (5 June 2013, 11:00 UT). Colours in the ionogram represent
reflected power from 20 (blue) to 110 dB (red). This
colour scale is not used directly by the scaler but is adjustable for best contrast.
Lower: graphical output of Autoscala, input based on the same raw
data but with separated, filtered O-mode (red) and X-mode (green)
traces. The boxes in the right panels show the scaled parameters as described in the text.
The Sodankylä ionosonde
Ionosondes have been running at SGO, URSI station SO166, since the
International Geophysical Year 1957. The present ionosonde, called
Alpha Wolf, is the third instrument in order and was installed in
2005. Details of the instrument have also been described in
. The Alpha Wolf is a frequency modulated
continuous wave (FM CW) chirp sounder developed at SGO. The FM CW
chirp technique implies that transmission and reception must be
simultaneous, so the transmitting and receiving antennas are separated
by approximately 1 km. The transmitter antenna is a rhombic
wide-band loop. At the receiver site an array of crossed magnetic loop
antennas is installed in order to provide imaging
capabilities. However, during these tests the signals from all
antennas were combined into two channels, one per linear
polarisation. The transmitter and receiver are completely separate
systems, starting on the same full second by GPS synchronisation. The
exciter of the transmitter and the local oscillator of the receiver
are basically identical, producing identical frequency sweeps at a
constant rate (chirps). The name Alpha Wolf derives from the distinct
“howling” sound produced when the down-converted received signal is
fed to a loudspeaker. After mixing and filtering, the base-band
converted signal is digitised into data streams of in-phase and
quadrature channels, in total four real vectors or one complex signal per
polarisation. The conversion from complex sampled data to ionograms is
simple: because of the linear frequency sweep, the frequency spread
around the centre frequency at each instant in time corresponds to an
interval of range (virtual height). The ionograms are thus obtained by
applying a windowed fast Fourier transform (FFT) to the digital signal
after combining the linearly polarised data into the O- or X-mode
circular polarisation. The centre time of the FFT window gives the
centre frequency and the length of the window determines the range
resolution, both determined by the frequency sweep rate. The selection
of length and overlap of the FFT windows thus implies a tradeoff
between frequency and range resolution. The ionograms in this paper
were produced with an FFT window length of 4096 points and thus the
length of the alias-free virtual height scale is up to 2048 points.
Subintervals limited to a maximum of 760 or 1500 km are shown
in the figures. Interference from known shortwave transmitters may be
filtered out in the process.
Figure shows a simplified block diagram of the whole
system including data processing. The receiver computer reads the
digital signals from the A/D converter of the receiver and stores the
data temporarily as binary files. These data are then made available
to the processing computers through the local network. Most important
of those computers is the operational real-time processing
computer. Its analysis software, a single Matlab script, produces O-mode
ionograms, which are archived and also available online in real time.
Automatic scaling
Autoscala works using input ionograms in a specific binary RDF
, in which O- and X-mode traces are saved
separately. Hence, in order to apply Autoscala to the ionograms recorded by
the SGO ionosonde, a change of file format was required. During the Autoscala
test phase presented here, a second set of real-time ionogram processing
software was therefore installed on a separate computer in order to produce
the required O- and X-mode traces. A modified and improved version of the real-time
analysis software was used, running under the open source
Matlab-compatible language Octave . In order to improve the
contrast of the O- and X-mode traces, an improved filtering to mask out weak
echoes was applied when converting the data to ionograms. This filtering may
be necessary in order for the automatic scaling to find the normal E- and F-layer
ionogram traces, but as will be seen it may also mask out real features,
such as spread F and sporadic E layers. The filtered O- and X-mode traces were
interpolated to matrices of fixed frequency and height resolution as required
by Autoscala. Subsequently the data were saved as RDF files and automatically
copied to INGV in Rome for processing. Matlab scripts for writing RDF files
are available on request from the corresponding author.
Figure shows an example that makes the difference
between the standard O-mode real-time ionograms and the filtered RDF
ionograms clear. The standard ionogram in the upper panel does show
some X-mode leakage, which is disregarded in the manual scaling
process. The panel below is the output from Autoscala, showing both
the filtered traces and retrieved parameters. The ionospheric traces
in the filtered RDF ionogram are more distinct as compared with the
standard ones, but there are many additional spurious apparent
echoes that look like noise. These may at least partly be due to the
interpolation to the fixed range–height grid of the RDF format. An
additional filtering (e.g. median filtering) could be applied to
remove them, but it seems that these spurious echoes do not affect the
Autoscala scaling.
Results of comparison
The Autoscala scaling was applied to SGO ionograms from the period
June–December 2013: 1 June–1 July (summer solstice, ionosphere constantly
sunlit), 29 July–13 October (late summer and equinox) and
1–19 December 2013 (midwinter); in total 117 days. Out of the data from
these days, 2610 ionograms were analysed both manually and automatically.
Comparison of scaled parameters
Figure presents a comparison of manually and
automatically scaled parameters for 5 June 2013 (selected as an instructive
case with well-defined E- and F-layer traces). The five panels in the plot
represent (from top to bottom):
M(3000)F2, shown by red asterisks (manual) and black circles
(Autoscala);
F-layer O-mode critical frequencies: foF2 (red asterisks for manual and
black circles for Autoscala), and foF1 (blue stars for manual and black
squares for Autoscala);
E-layer O-mode critical frequency, foE (red asterisks for manual and
black circles for Autoscala);
critical frequency of sporadic E layers, foEs (red asterisks
for manual and black circles for Autoscala); and
virtual height of sporadic E layers, h′Es
(red asterisks for manual and black circles for Autoscala), manually scaled
virtual height of E layer, h′E (blue diamonds), lower edge of E
layer, always set to 90 km in Autoscala (dashed line).
Comparison of manually and automatically scaled parameters for
4 June 2013. Panel 1: M(3000)F2, panel 2: F-layer O-mode critical
frequencies, panel 3: foE, panel 4: foEs and panel 5: E-layer
virtual heights. Autoscala always assumes the lower edge of the E layer at
90 km (dashed line). The yellow dot on the time axis indicates a time
(18:00 UT) when the ionosonde was not operating.
Figure shows a good agreement between
automatically and manually scaled foF2 and foF1. However, scaling of sporadic
layers appears to be more problematic. Similar daily summaries were made for
all 117 days when at least one of the 24 ionograms was scaled by Autoscala.
The results are presented in Tables ,
and and shown as histograms
in Figs. and .
Distributions (histograms) of the differences between manually and
automatically scaled parameters. Numbers (N) on top of the panels
indicate the numbers of ionograms for which both manual and
Autoscala parameter values were obtained, and, hence, the
differences were calculated. Vertical lines show medians and
quartiles.
Number of identifications of F- and E-layer parameters, % out of in total
2610 ionograms. Auroral activity is indicated by the mean AE index for the
cases of both manual and automatic identification and only manual
identification, respectively.
Parameter
Manual
Auto
Man and
Man only
Auto only
Mean AE,
Mean AE,
Auto
man and auto
man only
foF2 and M(3000)F2
90.4
87.4
85.9
4.5
1.4
133
360
foF1
35.6
17.6
16.3
19.3
1.3
170
191
foE
58.9
84.9
54.9
4.0
29.9
133
318
foEs and h′Es
63.0
20.0
19.4
43.5
0.6
257
130
Differences between manually scaled and Autoscala values (manual - auto).
Parameter
Mean ± SD
Median
Q1
Q3
ΔM(3000)F2
-0.02±0.16
-0.04
-0.11
0.06
ΔfoF2 (MHz)
0.0±0.4
0.1
-0.1
0.1
ΔfoF1 (MHz)
0.1±0.1
0.1
0.0
0.2
ΔfoE (MHz)
0.0±0.4
0.0
-0.2
0.1
ΔfoEs (MHz)
0.5±1.2
0.6
0.1
1.0
Δh′Es (km)
-6.0±8.3
-3.7
-9.3
-0.1
Manual h′E (km)
96±11
92
89
98
SD = standard deviation, Q1 = first quartile, Q3 = third quartile.
Scaled parameters of sporadic E layers.
Type Es
N man
% auto
ΔfoEs (MHz), man - auto
Δh′Es (MHz), man - auto
Mean ± SD
Median
Q1
Q3
Mean ± SD
Median
Q1
Q3
C
328
35
0.7±1.0
0.7
0.1
1.0
-0.5±4.5
0.1
-1.3
1.6
H
244
7
0.6±0.6
0.4
0.1
0.8
0.2±2.6
0.7
-2.6
1.8
L
373
25
0.4±1.2
0.4
0.0
1.0
-5.4±4.8
-6.2
-9.3
-1.6
F
128
41
0.5±1.3
0.5
0.1
1.1
-4.2±6.7
-3.2
-6.2
-0.1
A
105
19
0.2±1.4
0.5
0.1
0.8
-14.5±11
-13.0
-25.7
-4.7
R
389
43
0.6±1.3
0.7
0.3
1.2
-9.5±9.1
-6.9
-13.7
-3.2
K
79
53
0.0±1.1
0.3
-0.3
0.6
-9.5±8.8
-8.9
-15.4
-1.6
SD = standard deviation, Q1 = first
quartile, Q3 = third quartile.
Distribution of h′E scaled manually. The red line at
90 km indicates the fixed height assumed in Autoscala. The mode of
the manually scaled h′E distribution is close to 90 km as
well.
Table presents the relative number (%) of
ionograms in which a certain parameter was identified manually,
automatically, by both methods, and by one method only. The F2 layer
is in general well identified by Autoscala, whereas for the other
layers (F1, E and Es) the coincidence of visual and
automatic identifications is rather low. In particular, the automatic
routine identifies an E layer in many cases when the manual scaler
does not. This happens mainly because the manual routine assumes that no E
layer exists during nighttime, so for night hours (intervals depending
on time of year) foE is not scaled manually at all. In such night
cases the Autoscala output frequently provides a model value of
foE = 0.5 MHz. Moreover, in many cases it seems that the
Autoscala foE is close to prior model values for daytime as well. On
the other hand, F1 and especially Es layers are
frequently not identified at all by Autoscala.
The two rightmost columns in Table present
averaged values of the AE index (which characterises geomagnetic
activity at auroral latitudes) for the cases when layers were
identified both manually and automatically, and only manually,
respectively. One may expect that under more disturbed conditions the
automatic scaling works less satisfactorily. Indeed, the data in
Table indicate such a tendency for the F2
and E layers, namely that cases of only manual identifications occur
under noticeably larger average AE conditions (360–318 nT
vs. 133 nT). The relative number of such identifications
apparently affected by auroral activity was of the order of
5–7 %. For F1 the AE dependency of automatic identification
is not obvious.
Table also shows that automatic
identification of sporadic layers works better under disturbed
conditions. However, this is not surprising. Indeed, many sporadic
layers at auroral latitudes are produced by auroral precipitation, so
that the most dense and distinctive Es (which are most
easily detected by Autoscala) occur during disturbed conditions. On the other
hand, an experienced scaler may notice less prominent, weaker layers
during more quiet conditions, which are not identified by the
automatic scaling.
Some examples of erroneous identifications and possible reasons will
be discussed further in Sect. .
Table presents statistics of the
differences between the manually and automatically scaled values of
all parameters. In the bottom row we present only manually scaled h′E,
since Autoscala always assumes the height to be 90 km. These
results are also illustrated by histograms in
Figs. and , where
Fig. shows the distributions of the
differences between manually and automatically scaled parameter values
with medians and quartiles, and Fig. shows
the distribution of manually scaled h′E values and the fixed height of
90 km assumed by Autoscala. The distribution of manually scaled
values has its most probable value close to 90 km as well, so
the assumption is reasonable, but there is a long tail of h′E
observations well above 100 km. It will be seen that the
occurrence of such E layers presents a problem.
The results can be summarised as follows:
Generally, there is a very good agreement between the
manually and automatically scaled F2 parameters, foF2 and
M(3000)F2. They were identified by Autoscala in 86 % of the cases
with an accuracy of 0.4 MHz and 0.16 units, respectively.
The F1 layer was detected by Autoscala in only about half of those
cases when it was identified manually. In the cases of detection, however,
the values of foF1 were reliable within an accuracy of 0.1 MHz.
For E, the foE values from Autoscala were within 0.4 MHz of those
obtained manually when scaled. Autoscala identifies foE in many
cases (30 % of all ionograms) when no E layer was
identified by the manual scaling. See also an example in
Sect. below.
Sporadic E layers (Es) were identified by Autoscala in
relatively few cases (about one-third of the manual detections). The difference
between manually and automatically scaled parameters may be significant.
It thus appears that sporadic E layers are the most difficult to scale
automatically. In the next section this is considered in more detail.
Scaling of sporadic E layers
Es layers e.g. are typically
caused by metallic ions. In the auroral oval, however, there are
additional specific sporadic layers caused by auroral
ionisation. Since auroral precipitation varies rapidly both in
intensity and horizontal localisation within the timescale of typical
ionosoundings (even the 1 min resolution of Alpha Wolf), and is
also often followed by strong D-layer absorption preventing the
observation of higher layers, the scaling of sporadic E layers at
auroral latitudes is often a difficult task.
Sporadic E layers are classified by assigning them to one or more of
the following types according to the shape and strength (blanketing of
higher layers) of the trace. More than one type can be observed in a
single ionogram and in these cases foEs and
h′Es are given for the layer with maximal
foEs. The rules for classification are described
in detail in the URSI handbook of ionogram interpretation
.
Mid-latitude types
The following types of sporadic E layers are those which can be
observed at subauroral latitudes. They can occur everywhere,
however. The list does not include the equatorial type (q) which is
only seen close to the magnetic equator.
C (cusp) is an Es trace showing a relatively symmetrical
cusp at or below the critical frequency of the normal E or particle E layer.
H (high) is an Es trace showing a discontinuity in height
with the normal E or particle E-layer trace at or above the critical
frequency. The cusp is not symmetrical.
L (low) is a flat Es trace below the normal E or particle E
minimum virtual height.
F (flat) is a clean Es trace which shows no
appreciable increase of height with frequency. Only applicable when
no values of foE are obtainable (i.e. during nighttime). At other
hours, similar Es traces are classified as L, C or H.
Differences between manual and Autoscala
scaling arising from high latitude sporadic E phenomena. Black
lines show manually scaled parameters, and red colour (lines
and/or letter) indicates Autoscala results. Example 1: foEs
identified as foE, and the slant Es (type S)
identified as an F layer.
Differences between manual and Autoscala
scaling arising from high latitude sporadic E phenomena. Black
lines show manually scaled parameters, and red colour (lines
and/or letter) indicates Autoscala results. Example 2: overestimated foEs
(13.7 MHz vs. manual 4.9 MHz).
Differences between manual and Autoscala
scaling arising from high latitude sporadic E phenomena. Black
lines show manually scaled parameters, and red colour (lines
and/or letter) indicates Autoscala results. Example 3: Underestimated
foEs (3.0 MHz vs. manual 7.0 MHz)
Differences between manual and Autoscala
scaling arising from high latitude sporadic E phenomena. Black
lines show manually scaled parameters, and red colour (lines
and/or letter) indicates Autoscala results. Example 4:
Error in h′Es scaling (117 km vs. manual
91 km).
High-latitude types
The following types of sporadic E layers occur as a result of particle
precipitation.
R (retardation) is an Es trace showing an increase in virtual
height at the high frequency end but which becomes partially transparent
below foEs.
K (particle) denotes the presence of a particle E layer, similar in
appearance to normal E, which obscures higher layers up to its critical
frequency.
A (auroral) denotes all types of very spread Es
traces typical of oblique reflections from structures that are both
small in spatial scales and varying rapidly in time
as compared with the sounding time. The typical pattern shows
a well-defined flat or gradually rising lower edge with stratified
or diffuse traces present above it.
Indicated but not scaled
S (slant) is a diffuse Es trace whose virtual height rises
steadily with frequency.
D (D layer) is not strictly a sporadic E layer but a weak diffuse trace at or below 95 km associated
with high absorption and consequently high fmin.
In Table , we present a comparison of manual and
Autoscala parameters separately for each type of Es. The first three
columns in the left present types of Es, numbers of manual
identifications and percentage of Autoscala identifications. Then,
averaged differences between manually and automatically scaled
values are presented as mean values with standard deviations and
median values with upper and lower quartiles.
From Table it is evident that K, R and F types were better
recognised by Autoscala, whereas the H type was seldom
identified. Autoscala typically underestimates foEs by about
0.5 MHz, and the virtual height is typically overestimated,
especially for the high-latitude type Es (up to the order of
10 km).
Examples of problems
Scaling of high-latitude ionograms is often difficult even for an
experienced scaler, due to phenomena such as particle precipitation
and oblique reflections. A comprehensive consideration of particular
cases is beyond the scope of the present paper, but to illustrate
possible problems we present a few examples in
Figs. ,
, and
that point out differences between
manual and Autoscala scaling. In these ionograms black lines show
manually scaled parameters, and red colour (lines and/or letters)
indicates Autoscala results.
In Fig. , foEs (scaled
manually at 3.5 MHz) was identified as foE. The
foEs value was automatically detected at 4.8 MHz
and h′Es was detected near 126 km. The slant
Es (type S) was identified as an F layer with
foF2 = 9.9 MHz. In Fig. , the
manually scaled foEs is 4.9 MHz, whereas
Autoscala finds foEs = 13.7 MHz. One more example
of foEs difficulties is given in
Fig. , where the Autoscala value is
3.0 MHz, whereas the manually scaled value is 7.0 MHz.
Figure shows an example of
h′Es, identified manually at 91 km and by
Autoscala at 117 km.
However, with regard to this issue it is worth underlining the two following points:
The Autoscala routine for autoscaling the Es layer is designed
mostly for mid-latitude ionograms (Scotto and Pezzopane, 2007).
The filtering process applied when generating RDF ionograms from
raw SGO Alpha Wolf ionosonde data often causes a deletion of
significant parts of the ionogram trace, especially those related to
spread F and Es features. Although improving the
contrast of normal E and F traces, this clearly affects the ability
of the Autoscala Es routine.
Conclusions
In summary, this comparison between
manual and Autoscala scaling of ionograms at the high-latitude
Sodankylä Geophysical Observatory site has shown that:
The F2 parameters (foF2 and M(3000)F2) were identified by Autoscala in 86 % of the manually identified cases, within 0.4 MHz and 0.16 units, respectively.
F1 is identified by Autoscala in significantly fewer cases
(about 50 %) than manually identified, but when identified the
values of foF1 were reliable within 0.1 MHz.
E-layer parameters found by Autoscala are close to the manually
scaled ones when those are scaled, foE agreeing within an accuracy
of 0.4 MHz. However, Autoscala detects E layers in many cases
when the manual scaling process either does not identify one or
assumes that none exist during nighttime.
The identification and classification of sporadic E layers are in
many cases very different from those of the manual scaling.
Scaling of ionograms at auroral latitudes is, in many cases, a demanding
task. Often there are multiple oblique echoes from different ranges
which show up as spread traces, and another difficulty is presented by
the frequent sporadic E layers caused by auroral
precipitation. Energetic particle precipitation also sometimes causes
D-layer absorption that blankets E and F echoes. This makes automatic
scaling less straightforward as compared to scaling of ionograms from
mid latitudes. More studies, including scaling of ionograms processed
from the raw recorded data with different FFT and filter parameters,
will be required in order to find optimal settings for the contrast of
normal E and F traces, which is a tradeoff with the detectability of
sporadic E and spread F.