Automatic recognition of type III solar radio bursts: Automated Radio Burst Identification System method and first observations

Transcription

1 SPACE WEATHER, VOL. 7,, doi: /2008sw000425, 2009 Automatic recognition of type III solar radio bursts: Automated Radio Burst Identification System method and first observations Vasili V. Lobzin, 1 Iver H. Cairns, 1 Peter A. Robinson, 1 Graham Steward, 2 and Garth Patterson 2 Received 13 July 2008; revised 16 December 2008; accepted 22 December 2008; published 9 April [1] Because of the rapidly increasing role of technology, including complicated electronic systems, spacecraft, etc., modern society has become more vulnerable to a set of extraterrestrial influences (space weather) and requires continuous observation and forecasts of space weather. The major space weather events like solar flares and coronal mass ejections are usually accompanied by solar radio bursts, which can be used for a real-time space weather forecast. Coronal type III radio bursts are produced near the local electron plasma frequency and near its harmonic by fast electrons ejected from the solar active regions and moving through the corona and solar wind. These bursts have dynamic spectra with frequency rapidly falling with time, the typical duration of the coronal burst being about s. This paper presents a new method developed to detect coronal type III bursts automatically and its implementation in a new Automated Radio Burst Identification System. The central idea of the implementation is to use the Radon transform for more objective detection of the bursts as approximately straight lines in dynamic spectra. Preliminary tests of the method with the use of the spectra obtained during 13 days show that the performance of the current implementation is quite high, 84%, while no false positives are observed and 23 events not listed previously are found. Prospects for improvements are discussed. The first automatically detected coronal type III radio bursts are presented. Citation: Lobzin, V. V., I. H. Cairns, P. A. Robinson, G. Steward, and G. Patterson (2009), Automatic recognition of type III solar radio bursts: Automated Radio Burst Identification System method and first observations, Space Weather, 7,, doi: /2008sw School of Physics, University of Sydney, Sydney, New South Wales, Australia. 2 IPS Radio and Space Services, Sydney, New South Wales, Australia. 1. Introduction [2] The rapidly increasing role of technology, including complicated electronic systems, spacecraft etc. makes modern society more vulnerable to a set of extraterrestrial influences that are now referred to as space weather. Today it is impossible to imagine the life and the activity of the society without continuous meteorological observation and weather forecasts. In the near future, observation and prediction of space weather may start to play a similar role. New centers of observation and studies of space weather are already organized and functioning. One of the goals of these centers is to provide a forecast and quick notification to registered users about forthcoming geomagnetic storms, increases in the intensity of radiation, fade-out of radio waves etc. [3] The major drivers of space weather are closely related to complicated explosion-like events on the Sun, i.e., solar flares and coronal mass ejections (CME) [e.g., Warmuth and Mann, 2005]. They are usually accompanied by solar radio bursts, in particular type II and III bursts [e.g., Dulk et al., 1985; Warmuth and Mann, 2005; White, 2007]. Two different systems for automatic recognition and classification of CMEs were already developed [Robbrecht and Berghmans, 2004; Qu et al., 2006]. [4] Both type II and III solar radio bursts are assumed to be generated by fast electrons, the emission being at the plasma frequency and/or its second harmonic. These events can be also used for a real-time space weather forecast. Indeed, a large fraction of flares exhibit a group of type III bursts in their rise phase [Cane and Reames, 1988; White, 2007; Pick and Vilmer, 2008]. Then, upon the Copyright 2009 by the American Geophysical Union 1of12

2 flare onset type II bursts are frequently produced. Thus such groups of type III bursts can be used as potential precursors of space weather events. In addition, observations of the frequency drift of type III events can provide information about the coronal density profile along open magnetic field lines originating in the chromosphere. Similar observations of type II bursts could be useful for analyzing the dynamics of shock waves that reflect and accelerate electrons while moving from the Sun [Warmuth and Mann, 2005; Lobzin et al., 2008]. [5] Typically solar type III radio bursts are produced by relatively low-energy electron beams at a speed of about c/3, where c is the speed of light, with the kinetic energy being 30 kev. Neither the electron beams nor the radio emission is usually expected to produce significant space weather consequences. Rather, these bursts are often associated with solar events that may lead to space weather events at the Earth. Type III bursts are commonly observed whenever there is a bright active region on the visible side of the Sun [e.g., Suzuki and Dulk, 1985] and most frequently the bursts occur in groups of 10 or more. During solar maximum, many thousands of type III bursts are observed. Indeed, the list of such events detected by the solar radio spectrograph at Learmonth contains 1181 records for year 2002, many of the records correspond to large storm-like groups lasting several hours. Thus type III events can be considered as an indicator of increased solar activity. However, for forecasting space weather at the Earth not all of them are of equal importance. First of all, it is necessary to detect type III burst groups at the start of major solar events. Storm type III bursts, which could appear later together with type II bursts and type I storms, can also be useful. [6] It is worth noting that there exist relatively rare solar radio events that do have significant direct space weather consequences. In particular, Terkildsen [2007] and Cerruti et al. [2008] reported that the intense solar radio bursts associated with X class solar flares during December 2006 had extremely high intensity in the frequency range of GHz where the Global Positioning System (GPS) transmits its signals. The bursts decreased considerably the signal-to-noise ratio for receivers tracking GPS satellites. Near the peak of the bursts some GPS receivers lost tracking of the satellites and failed to produce a navigation solution. These unusual events were observed during solar minimum and the occurrence frequency of similar events will probably increase with solar activity. [7] For future development of real-time automated prediction of space weather with the use of solar radio data, it is necessary to develop objective, accurate, and efficient numerical methods for classifying solar radio bursts and estimating their parameters (start/end times, intensity, frequency range, etc.). However, at present, meter wavelength solar radio bursts cannot be identified and classified automatically using real-time data. Instead, they are usually analyzed by eye, so different observers sometimes provide different characteristics for the same event. [8] This study presents a new method for recognition of type III radio bursts, the first version of an Automated Radio Burst Identification System (ARBIS) for finding them in real-time data, and the first detection of type III bursts in real time. This work can be considered as the first step toward an automated system for space weather forecasting based on radio spectrograph data. The technique described here and ARBIS itself should be also useful for radio observatories engaged in the routine detection and documentation of solar radio events. [9] The paper is organized as follows. The data used in the study are described in section 2. The procedure of data processing is outlined in section 3. Section 4 presents the results obtained, including the system performance and the first real-time observations of type III radio bursts. Section 5 summarizes the results and gives the conclusions. 2. Data [10] In the present study we use the solar radio spectral data provided by Learmonth Solar Radio Observatory (Western Australia, E, S), Culgoora Solar Observatory (Eastern Australia, E, S), and Palehua Solar Radio Observatory (Hawaii, W, N). [11] The Learmonth and Palehua solar radio spectrographs belong to the Radio Solar Telescope Network (RSTN) operated by the U.S. Air Force. They cover a frequency range of 25 to 180 MHz, and complete a frequency sweep every 3 s. The total frequency range is split into two bands. The low band covers from 25 to 75 MHz and the high band from 75 to 180 MHz. Both low and high bands are split into 401 subbands with linearly spaced central frequencies. A relative logarithmic scale is employed for the signal intensity. Because of many factors, it is not presently possible to provide an absolute flux calibration. The logarithm of the relative intensity takes integer values in the range of 0 to 255. For further details, see the documentation by Kennewell and Steward [2003]. Every minute the data acquisition system writes 20 subsequent spectra into a separate file. Both archived data files stored by NOAA s National Geophysical Data Center (NGDC) and real-time data provided by the Learmonth observatory are used in this study to elaborate suitable techniques of data processing and to illustrate the results obtained. [12] The Culgoora Solar Radio Spectrograph observes the radio emission of the Sun from 18 to 1800 MHz and performs a frequency sweep every 3 s. The frequency range is split into four bands: , , , and MHz. Each band is split into 501 subbands with linearly spaced central frequencies. The daily files provided by IPS Radio and Space Services (anonymous ftp to ftp://ftp.ips.gov.au/wdc-data/spec/data/culgoora/raw/) are used for comparison with the Learmonth observations. 2of12

3 [13] Using the NGDC event listing (anonymous ftp to ftp://ftp.ngdc.noaa.gov/stp/solar_data/solar_ RADIO/SPECTRAL/SPEC_NEW.02), we examined the occurrence of type III events observed at the Learmonth observatory near the solar maximum in 2002 and found that during this year the monthly number of events varies in the range of 64 to 151. For further analysis we selected the period July 2002, when 48 type III events were listed. The extract of the NGDC event list is presented in Table 1. Relatively long time intervals with intermittent activity, which sometimes last several hours, are considered in Table 1 as single events. In order to explain the procedure of data processing, a simple isolated type III burst observed on 20 July 2002 is used. The possibilities of the method are then illustrated using three different examples: a very weak singular burst detected at Learmonth, a storm-like group of type III bursts which is associated with a type II burst on 18 July 2002 and observed at Learmonth, and a remarkable outburst, which was detected by the Palehua spectrograph on 29 October The last event is associated with an X10 solar flare, one of the biggest flares observed during the intense space weather storms of the period 19 October to 7 November 2003, and with an extremely fast CME. 3. Data Processing [14] The processing of the spectra has three stages: preprocessing, recognition, and classification Preprocessing [15] The preprocessing module finds missing or bad data and creates a two-dimensional array G(i t, i f ) of dynamic spectrum, where the indices i t and i f correspond to time and frequency, respectively, and i f =1,..., 802. The array G does not contain gaps, i.e., the time interval between successive spectra is always equal to 3 s, with bad and missing data replaced by random synthetic samples. Of course, the occasional type-iii-like events that can be found in the synthetic data are not taken into account. [16] When the system is working in real time, this module is also responsible for accumulation of data into longer dynamic spectra by appending new data. When the desired length of the spectrum is reached, the preprocessing module begins removing older data from the array, thereby keeping the same array size. The main criterion for choosing an appropriate array size is that the spectrum should be longer than any event under consideration. On the other hand, the time required for data processing is approximately proportional to the spectrum length. Thus when the system is working in real time, it is undesirable to process the entire accumulated daily spectrum. Rather, it is worthwhile to cut it into shorter overlapping fragments and to process only one fragment with the most recent data. For short events like type III bursts and the present time resolution it is easy to meet these criteria. [17] Overlapping of the spectra fragments is essential. Indeed, some procedures of data processing like filtering etc. can produce boundary effects. Thus the affected parts of the spectra should be discarded. Without overlapping, this will result in missed events that could occasionally occur not far from the boundaries between fragments Recognition [18] Type III events are fast frequency drift bursts, which can occur singly, in groups, or in storms. The typical duration of a single burst is 1--3 s, while a group could last min, and storms last from minutes to hours. With the given time resolution, which is equal to 3 s per spectrum, the bursts usually look like straight segments parallel or almost parallel to the frequency axis. Quite often they occupy almost the entire frequency range of the RSTN spectrographs. The signal intensity for a type III burst can vary with frequency. The observed variations can be nonmonotonic; while there is a tendency for increasing intensity with decreasing frequency, more intense parts seem to last longer than weaker ones. [19] The main property of type III bursts used to find them in spectra is their resemblance to straight segments, often with gaps. The main parameter that is used when making a decision is closely related to the number of frequency channels in which the burst signal is visible. The definition of signal visibility is discussed in the following. If a burst is observed either in the low- or highfrequency band only and there are no gaps, then the number of channels is proportional to the frequency range occupied by the burst. However, the proportionality coefficients are different for the two bands. Hence, when the burst is visible in both bands, there is no one-to-one correspondence between the parameter chosen and the physical characteristics of the burst. Nevertheless, if we consider a dynamic spectrum as an image, this approach is quite reasonable. [20] Since only the number of channels is chosen to be important for recognition, while the signal intensity is not used, it is convenient to transform the gray-scale image of the observed spectrum G(i t, i f ) into a binary image B(i t, i f ), with each pixel having one of only two discrete values: 1 or 0. The pixel value B(i t, i f ) is equal to 1 if there is a significant enhancement in the signal intensity in the corresponding channel and at the corresponding time. To decide whether the observed enhancement is significant, it is necessary to choose a threshold, which can depend both on frequency and time. In the available data files, the signal intensity is measured in relative units, and the logarithms of intensity can take integer values from 0 to 255 [Kennewell and Steward, 2003]. In the present study the threshold value for variations of the logarithm of signal intensity is chosen to be equal to 1, i.e., B(i t, i f ) = 1 if the pixel corresponds to a local maximum with respect to time, i.e., G(i t -- 1, i f ) G(i t, i f ) G(i t +1, i f ). Otherwise B (i t, i f ) = 0. At the array boundaries corresponding to 3of12

4 Table 1. Type III Radio Bursts Observed at Learmonth Observatory on July 2002 and a Sample Single Burst Found on 20 July 2002 a Date Start Time (UT) End Time (UT) Intensity Start Frequency (MHz) End Frequency (MHz) Remarks b 1 Jul S 1 Jul S 2 Jul Jul Jul c Jul G 3 Jul c Jul S 4 Jul Jul G 4 Jul c Jul G 4 Jul S 5 Jul S 5 Jul G 6 Jul c Jul c Jul G 6 Jul G 7 Jul G 7 Jul G 7 Jul c Jul Jul G 8 Jul S 8 Jul G 8 Jul Jul Jul S 9 Jul S 9 Jul G 9 Jul S 10 Jul Jul S 11 Jul c Jul c Jul S 11 Jul c Jul c Jul Jul c Jul Jul c Jul G 12 Jul G 13 Jul c Jul G 13 Jul G 20 Jul S a Data shown are extracted from NGDC event listing, except Remarks which contains the results of the present study. b Here 0 stands for a false negative, while S and G denote a single and group burst, respectively, and events not analyzed are marked by dashes. c Periods of intermittent activity. i t = 1 and i t = i t max, the definition of the local maximum should be modified, i.e., B(1, i f )=1ifG(1, i f ) G(2, i f ) and B(i t max, i f )=1ifG(i t max, i f ) G(i t max -- 1, i f ). However, this modification is of no importance because these boundaries are excluded from further analysis. [21] Then the Radon transform of the binary image is used to find the features corresponding to type III bursts. The Radon transform is used in many applications where it is necessary to detect straight lines within an image [e.g., Deans, 1983]. For a function of two variables, G (x, y), the 4of12

5 Radon transform is the integral transform consisting of the integrals of the function along straight lines G R ða 1 ; a 2 Þ ¼ Z þ1 1 Gxs ð ðþ; ys ðþþds; ð1þ where a 1 and a 2 are two parameters that uniquely determine a straight line and ds is the differential distance along the line. The choice of the parameters a 1 and a 2 is a matter of convenience. In particular, one can use the angle between the line and x axis and the minimum distance from the line to the origin. Another popular choice is the pair of parameters a and b from the slope-intercept representation of a straight line, y = ax + b. [22] For the problem at hand it seems convenient to choose a slope-intercept-like pair of parameters that can be interpreted with some reservations as the burst s start time t s and duration t. We assume that type III events are observed either as simultaneous enhancements (within the time resolution of 3 s) for all frequencies or else that high-frequency part of the burst comes first and the lowfrequency part is observed in the next few spectra. In accordance with our definition, the start time t s is the time when the enhancement should appear in the highestfrequency channel. The end time t e is the time when the enhancement reaches the lowest-frequency channel. Then the duration of the event is defined as t = t e -- t s. These definitions of times may seem oversimplified. Indeed, a typical event does not occupy the total frequency range of the instrument, hence, from simple geometrical considerations it follows that t s can be less than the actual start time, while t e can be greater than the corresponding end time, thereby the duration of the event can be overestimated. On the other hand, the time resolution of the Learmonth spectrograph data is insufficient to determine times of events with an accuracy better than 3 s. In addition, the timing accuracy of the lists provided by NGDC is about 1 min: this is an order of magnitude greater than the expected difference between our simplified estimates and the more accurate ones that could be obtained from instruments with higher time resolution. [23] To use the Radon transform for analyses of spectra, it is necessary to choose an appropriate approximation of (1). Instead of infinitely many integrals along straight lines, for each value of time index i t, we consider a finite number of sums of the matrix elements B(i t, i f ) along given segments. The segments have the same origin at (i t, 802), while their ends are at the low-frequency row of the matrix, i f = 1, and spaced 1 pixel apart, the corresponding event durations are t = 0 s, 3 s, 6 s etc. To calculate a pixel value on the segment, the nearest-neighbor interpolation is used. Thereby the matrix B(i t, i f ) is Radon-transformed into a matrix B R (i t, i t ), where the second index enumerates the durations t. [24] A maximum value of B R (i t,i t ) is calculated for each column, thereby obtaining a time series S(i t )=max B R (i t, i t ). 0i t i tmax Then a background subtraction is performed for this time series, i.e., S III (i t )=S(i t )-- S0(i t ), where the background value S0(i t ) is calculated as the minimum of S in the vicinity of i t. Hence, S III (i t ) is a nonnegative time series. [25] To make a decision on whether a type III event is observed, the values of S III (i t ) are compared with a threshold. When the threshold is exceeded, then this event is considered to be a potential type III burst. The final decision is made by a classification module Classification [26] In the current implementation the classification module differentiates two classes of events, i.e., single bursts and groups, and produces two lists. The bursts are combined into a group if they are less than 1 min apart. It is worth recalling that 1 min is equal to the time resolution of the NGDC listings. The first list contains the dates and times of the start and end of each single burst or group with an accuracy of 1 min, and a remark as to whether the event is a single burst or a group of bursts. Times with higher precision (3 s) are included in the second list. In addition to this information related to single bursts, the second list contains information about the fine structure of the groups found, i.e., dates and times corresponding to each burst in the group with 3 s relative accuracy. Thus the first list resembles the NGDC listings (see the caption of Table 1), while the additional information, which could be rather extensive, especially for long storms, is included in the second list An Illustrative Example and Choice of Optimal Threshold for Recognition [27] To illustrate the data processing procedure outlined above, consider a single type III burst observed on 20 July 2002 at 0023 UT. [28] Figure 1 shows the 6-min dynamic spectrum with the burst in the middle, the corresponding binary image, and the profile of S III versus time. The type III radio burst is clearly seen in the dynamic spectrum as an almost simultaneous increase in signal intensity in almost all channels. The corresponding vertical line in the spectrum is crossed by several horizontal lines resulting from interference related to communications systems. The lowfrequency part of the burst is also crossed by an oblique segment caused by interference from a digital ionosonde. It is easily seen that the signal intensity does not vary smoothly with frequency. Rather, there is an apparent jump at the boundary of the two frequency bands. In addition, the effective gains of different channels in the low-frequency band seem to differ considerably. This makes it difficult to estimate the intensity of events but does not influence the performance of the method developed. Indeed, the transformation of gray-scale spectral images into binary images includes the comparison of pixel values corresponding to the same frequency channel. This approach results in a binary image of event with a 5of12

6 [30] To decide whether the observed peak in the S III (t) profile corresponds to a type III radio burst, it is necessary to choose a threshold value. This value is of crucial importance, because it determines the error rate of the system. As in any statistical test, there are two kinds of errors, which are often referred to as false positives and false negatives [see, e.g., Neyman and Pearson, 1967a, 1967b]. The false positive, or false alarm, is an error when the system makes a decision that an event was found while it is actually absent. The false negatives correspond to the situations when the actual events are missed. There is a trade-off between the acceptable level of false positives and false negatives. Increasing the threshold makes the test more restrictive and results in an increased probability of false negatives. On the other hand, as the threshold decreases, the number of false positives becomes bigger. [31] The optimal threshold value can be determined empirically. If we assume that the acceptable probability of finding a false type III burst in one spectrum is 10 5, then the probability of a false alarm in one day s data will be about 0.1, because one day s data contain 10 4 spectra. By analyzing recent data obtained at Learmonth in April and May 2008 when there were no solar radio bursts, we found that the threshold value of 110 matches this criterion under current conditions. However, the optimal value of the threshold could depend on time through the mediation of different parameters essential for propagation of radio waves in the given frequency range, in particular ionospheric conditions, which in turn strongly depend on the levels of solar and geomagnetic activity. Figure 1. (top) Dynamic spectrum G(i t, i f ) of a single type III radio burst observed on 20 July 2002 at 0023 UT. (middle) The corresponding binary image B(i t, i f ). Pixels are black where B = 1 and white for B = 0. (bottom) Profile of S III (solid line) versus time and the threshold level (dashed line). relatively homogeneous background (see, e.g., Figure 1 (middle)). Interference is still visible in the binary image, but the total number of affected channels is usually relatively small and so have no significant effect on further image processing and event recognition. Indeed, the recognition, as is described in the following, is essentially statistical because it takes into account a relatively large number of pixels when making a decision as to whether a type III is observed. [29] Although the event shown in Figure 1 has low intensity (class I), it occupies almost the entire frequency range of the instrument. As a result, the corresponding peak in the S III profile is relatively high as compared with the background fluctuations and can be found easily. 4. Results and Discussion 4.1. Performance and Its Improvement [32] Using the NGDC event listing, a contiguous time interval July 2002 was chosen to perform the performance test of the method. For this interval, 48 events were previously listed, including relatively long time intervals with intermittent activity. The corresponding extract from the NGDC event listing is presented in Table 1. Upon removal of intervals with intermittent activity (such intervals are indicated by a footnote in Table 1), the total set of listed events contains 35 type III bursts or their groups. Our procedure of automated detection with the chosen threshold value reveals 26 of them without any false positives. The events that were not found are marked by zeros in Remarks in Table 1. With the results listed in Table 1 alone, the performance rate is 74%. However, we found 23 additional type III events that were not previously listed. These events are included in Table 2. These were confirmed by visual examination of dynamic spectra. Thus, the NGDC event listings are not complete. In fact, the NGDC listing rate is only /( ) 57%. [33] Obviously, the nonlisted events should be taken into account when estimating the performance of the method. Therefore the success rate of the current system 6of12

7 Table 2. Type III Radio Bursts That Were Found in This Study and Are Absent in the NGDC Event Listing for July 2002 Date Start Time (UT) End Time (UT) Remarks a 1 Jul S 3 Jul G 3 Jul S 3 Jul S 4 Jul S 4 Jul S 5 Jul S 6 Jul G 7 Jul S 7 Jul G 7 Jul S 8 Jul S 9 Jul S 10 Jul G 10 Jul G 11 Jul G 11 Jul G 11 Jul S 13 Jul G 13 Jul G 13 Jul S 13 Jul G 13 Jul S a Here S and G denote a single burst and group, respectively. can be estimated as 100 ( )/( ) 84%. This is better than the rate of NDGC lists by a factor of 1.5. [34] An examination of the false negatives (the missed events in Table 1) shows that all of them are associated with the bursts occupying relatively narrow frequency ranges. If these ranges do not contain interference signals, then these events are easily seen in the spectra, but the value of S III does not exceed the chosen threshold. On the other hand, as was already mentioned, the optimal value of the threshold could depend on time through the mediation of different parameters, in particular, levels of solar and geomagnetic activities. Using such a varying threshold could improve the performance of the system. [35] It is worth also noting that the optimal threshold values are actually dictated by the goal. Indeed, an automated system, which is expected to send an alarm signal or message about a forthcoming space weather event, should not produce too many false alarms. On the other hand, if the same procedure is used for a manual processing of relatively short time intervals with the aim of finding an exhaustive list of events, the false positives are not so important because the final decision will be made by an operator. In this case it seems reasonable to make the threshold lower in order to decrease the number of false negatives. [36] Both for automated recognition and manual data processing, one can complement this method by additional techniques and use them together. In particular, we have found that smoothing the measured spectra makes it easier to find relatively weak bursts that occupy a limited frequency range that is not wide enough to use the Figure 2. Smoothing of dynamic spectra for detection of weak events. The (top) dynamic spectrum of a single burst observed on 10 July 2002 at 0734 UT, (top middle) corresponding binary image, (bottom middle) binary image for the smoothed spectrum, and (bottom) profiles of S III versus time for unsmoothed and smoothed spectrum (red and blue solid lines, respectively) with the corresponding thresholds (red and blue dashed lines). In the binary images the pixels are black where B = 1 and white for B =0. 7of12

8 Figure 3. An example of storm-like type III events observed simultaneously with a type II burst on 18 July (top) Dynamic spectrum G(i t, i f ) observed at UT. (middle) The corresponding binary image B(i t, i f ). Pixels are black where B = 1 and white for B =0. (bottom) Profile of S III (solid line) versus time and the threshold level (dashed line). There are 29 significant peaks that are associated with type III bursts. method described above. The main contribution to this improvement most probably results from reducing the noise in the spectra and the subsequent decrease of fluctuations in S III. Such smoothing decreases the optimum threshold value and thereby allows one to detect events occupying narrower frequency ranges. [37] Figure 2 shows an example where this procedure reveals a burst that was not found by the method without smoothing. Figure 2 (top) shows the dynamic spectrum of the event observed on 10 July 2002 at 0734 UT, Figure 2 (top middle and bottom middle) show the corresponding binary images for the spectrum without and with smoothing, and Figure 2 (bottom) contains the plots of S III for these two techniques and the corresponding thresholds estimated with the recent data. The peak of S III for the unsmoothed spectrum is easily seen but it is below the threshold and not very high as compared with the amplitude of fluctuations. Smoothing the spectrum does not affect the peak considerably, but it increases the signal-to-noise ratio, and the peak of S III is above the revised threshold and can be automatically detected. [38] This method with smoothing seems to be better for finding single bursts. However, for a group of bursts, which are not far from each other in time and frequency, such smoothing can smear out the fine structure of the group Large Groups and Storms [39] As another example, consider a fragment of a long time interval that was mentioned in the NGDC list of type III bursts as an event with intermittent activity in the entire frequency range of the instrument. The Learmonth observations on 18 July 2002 show intermittent activity for about 4 hours, from 0548 to 0940 UT. [40] Figure 3 shows a fragment of this event from 0738 to 0757 UT. A classic type III/II event is clearly seen on the spectrum. In contrast to short-lived type III bursts, typical type II radio bursts feature relatively slow frequency drift, less than 1 MHz s 1, and last several minutes [Nelson and Melrose, 1985; Pick and Vilmer, 2008]. In this particular case the type II burst lasts from 0747 to 0753 UT, in accordance with the NGDC event listing, and is preceded by a large group of type III bursts. Another large group of type III bursts follows the first one and overlaps with the type II burst. Indeed, the binary image of the processed spectrum shows a lot of approximately vertical lines throughout the entire time interval (see Figure 3 (middle)). Analysis of S III profile (Figure 3 (bottom)) reveals these two groups with a single burst between them. The total number of significant peaks with the threshold chosen is equal to 29, while the duration of the interval is only 19 min. Thus, on average the occurrence rate is about 1.5 bursts per minute. Obviously, the manual data processing of similar events, which could be rather long, is difficult, time consuming, and subject to errors. The automatic procedure will be useful if it is necessary to describe in detail a fine structure of large groups and storms. Shorter events could be processed by an operator, but even in this case the technique illustrated here with different thresholds seems to be quite helpful. [41] The second example of event containing large groups of type III bursts was chosen from observations during a remarkable period of space weather storms during 3 weeks from 19 October to 7 November 2003 when 17 major solar flares, 6 radiation storms, and 4 severe geomagnetic storms were detected [see, e.g., Balch et al., 2004]. There were three of the most significant solar flares that had themostimpactonspaceweatherattheearth.they occurred on 28 October at 1110 UT, on 29 October at 2049 UT, and on 30 October at 1600 UT. The first two flares were 8of12

9 2002 was relatively weak while the bursts on 30 October 2003 were followed by a severe geomagnetic storm, which affected power grids and plants. A lot of other space weather effects, including difficulties with HF communications and disruptions to the operation of satellites, were also observed on October 2003 and reported [see, e.g., Balch et al., 2004]. [44] These two examples of storm-like bursts show clearly that information contained in solar radio spectra is not sufficient to predict space weather events at the Earth. To make a reliable space weather forecast, it is necessary to combine different observations (optical observations of the solar disk, measurements of X-ray emissions Figure 4. Dst variations for the period of intense space weather storms at the end of October Short vertical segments above the plot show the times of major solar flares (segments with circle endings) and the arrival of CMEs at the Earth (segments with triangle endings). accompanied by Earth-directed CMEs. These two CMEs arrived at the Earth on 29 October at 0613 UT and on 30 October at 1600 UT, respectively, and caused severe geomagnetic storms [Balch et al., 2004]. Figure 4 shows the Dst variations at the end of October 2003, together with the times of the flare onsets and arrival of CMEs at the Earth. It is easily seen that severe geomagnetic storms developed after the CME arrivals. The minimum value of the geomagnetic index Dst during this period was 401 nt and maximum value of Kp was 9. It is worth noting that, in accordance with the classification by Sugiura and Chapman [1960], a storm is considered to be intense if the minimum Dst is less than 100 nt. These two geomagnetic storms are classified as severe in accordance with the NOAA space weather scale based on the geomagnetic index Kp. [42] The solar radio bursts that accompany the second solar flare were detected at the Palehua Solar Radio Observatory. The dynamic spectrum for this event is shown in Figure 5. It is easily seen that this spectrum is much noisier than spectra measured at Learmonth. Nevertheless, despite the facts that a lot of channels are affected by strong interference from communications and that there are also band stop filters incorporated into the hardware, our technique is still able to detect type III events even without adjusting the threshold that was estimated from Learmonth data (see Figure 5 (middle and bottom)). Qualitatively this dynamic spectrum is similar to that shown in Figure 3. Both type II and III radio bursts are clearly seen and the type III bursts are grouped into two bunches, with the second bunch overlapping the type II burst. [43] Although the spectra shown in Figures 3 and 5 are similar, the space weather consequences of the solar events associated with these bursts are considerably different. Indeed, the geomagnetic activity on July Figure 5. An example of the major outburst observed on 29 October (top) Dynamic spectrum G(i t, i f ) observed at UT. (middle) The corresponding binary image B(i t, i f ). Pixels are black where B = 1 and white for B = 0. (bottom) Profile of S III (solid line) versus time and the threshold level (dashed line). 9of12

10 Figure 6. (top) The Learmonth dynamic spectrum for a group of type III radio bursts observed in real time at UT on 6 June (middle) Profile of S III (solid line) versus time. The threshold level (dashed line) corresponds to the probability of false alarms of 10 5 or 0.1 per day. (bottom) The dynamic spectrum measured at the same time by Culgoora spectrograph. Clearly seen at UT is the pair of type III radio bursts that were found in the real-time Learmonth data. and energetic particles etc.) and perform numerical modeling with the use of observational data. Nevertheless, solar radio spectra alone are still useful for space weather purposes. Indeed, a bunch of type III bursts occurs in the early rise of the solar flare, while the X-ray, EUV and optical emissions attain their maximum values later, typically within min, while energetic particles arrive at the Earth after about 1 hour. Both solar X-rays and energetic particles make a considerable contribution to the ionization of the Earth s atmosphere up to stratospheric altitudes [e.g., Schlegel, 2005]. The increased ionization of the lower ionosphere results in increased damping of electromagnetic waves and thereby affects radio communication [Schlegel, 2005, and references therein]. Thus a bunch of type III bursts can potentially be used as a precursor of the main phase of solar flares and its ionospheric effects, as well as a precursor of type II solar radio bursts and all other space weather effects that flares cause, including CMEs, shock wave formation in the corona, arrival of the shocks and CMEs at the Earth, geomagnetic storms etc. Consequently such radio spectra features can be useful to call an alert that a significant space weather event may occur. The probability of false alarms can be reduced by further improving the technique for radio spectra processing and combining the results with other observations and/or numerical modeling First Observation of Type III Bursts in Real-Time Data [45] The first real-time detection of coronal type III radio bursts occurred on 5 June 2008, near 2318 UT, when the level of solar activity was very low. Figure 6 shows the fragment of a dynamic spectrum registered by the Learmonth radio spectrograph and the corresponding profile of S III (t). Clearly seen in the center of the spectrum is a group of three bursts, which occur at 2318:00, 2318:15, and 2318:24 UT. However, there are only two significant peaks in the S III (t) profile. These peaks correspond to the first and third burst, while the second one is only visible in a relatively narrow frequency range and was not detected by the system. Interestingly, there were no active regions on the Sun during this period. [46] Although the enhancements in the spectrum look like typical type III events, being observed almost in the entire frequency range of the instrument, these features could be spurious and/or related to interference. It is worth noting that with the chosen threshold value a spurious weak event could appear in the spectrum, but the probability of finding a group of two or three bursts within a short time interval is negligible. [47] On the other hand, wide-band interference of technological or natural origin could produce such features, particularly at low frequencies. Indeed, the Learmonth spectra sometimes contain such enhancements that last a few minutes. However, such interference has never been observed at high frequencies and with the chosen threshold it cannot produce false positives, as was mentioned above. The signals generated by lightning discharges are very short and can occupy a rather wide frequency range, but there were no lightning flashes observed at the observatory on 5 June [48] Irrespective of the nature of the potential interference, it should be local because radio waves in the given frequency range penetrate into the ionosphere and propagate into space, they are not observable on Earth s surface very far from their source. Hence, an independent 10 of 12

11 observation of the bursts at another observatory can be considered as definitive evidence in favor of a solar origin for the bursts. The Culgoora radio spectrograph, which is located about 3700 km from Learmonth, measured the spectrum shown in Figure 6 (bottom). Clearly seen is the same pair of bursts that were detected by the Learmonth radio spectrograph. Between the bursts there is no burstlike feature similar to that visible in the Learmonth spectrum but not detected by the automatic recognition system because of narrow frequency range it occupies. Thus we can conclude that the two features detected in real time do correspond to coronal type III bursts, while the third feature in between is probably a local interference effect. 5. Summary and Conclusions [49] This paper presents a new objective method developed to detect automatically coronal type III bursts, both in real time and in archived radio spectrograph data, and the corresponding Automated Radio Burst Identification System (ARBIS). The central idea of the implementation is to use the Radon transform to detect bursts more objectively as approximately straight lines in dynamic spectra. Preliminary tests of the method show that the performance of the current implementation is quite high, 84%, while no false positives were found for a 13-day test period. In addition to events included in the NGDC event listings, we found 23 type III events that were not listed. Visual examination of dynamic spectra confirmed that these nonlisted events are quite typical coronal type III radio bursts. In fact, the accuracy of the NGDC listing was only 57% for the time interval considered and the performance of the automatic system is higher by a factor of 1.5. [50] The main parameter used when making a decision about any event is closely related to the number of frequency channels in which the burst signal is visible. Thus bursts occupying narrower frequency ranges are more difficult to find. The performance can be improved by decreasing the thresholds used in the recognition module. However, this will make the criteria less restrictive and will increase the number of false positives. On the other hand, the same procedure with a lower threshold can be successfully used for a manual processing of relatively short time intervals with the aim of finding an exhaustive list of events. Another technique, which increases performance and can be useful both for automatic and manual data processing, includes smoothing of the measured dynamic spectra. This approach works well for detecting weak single bursts, but the fine structure of the groups can be smeared out. [51] An additional advantage of the presented automated system ARBIS and complementary techniques of data processing is that they allow the fine structure of type III groups and storms to be automatically described for the first time. This is important because such events can last several days and comprise hundreds or thousands of bursts. [52] The first type III bursts detected in real time by ARBIS were presented and shown to be of solar origin. Moreover, the algorithm was found to be resistant to narrowband interference. [53] It is shown that groups of type III bursts, which are often observed at the rising stage of solar flares, can be easily detected even if a lot of channels are affected by intense narrow-band interference. These groups can potentially be used for space weather forecasting as a precursor of space weather effects caused by flares. These include CMEs, shock wave formation in the corona, arrival of the shocks and CME at the Earth, geomagnetic storms etc. However, in order to discriminate such groups of bursts from similar groups with negligible space weather consequences and thereby to decrease the number of false alarms, it is necessary to combine this technique with other observations and/or numerical modeling. In particular, further examination of the radio spectrum in order to find type II radio bursts and estimation of the speed of the shock wave related to this burst may be very useful. On the other hand, with an accurate automatic recognition system available, more quantitative investigations of the timing and correlation of type III radio bursts with solar space weather events will become possible. [54] Acknowledgments. The authors are thankful to Phil Wilkinson, David Neudegg, John Kennewell, Alan Brockman, Owen Giersch, and Nigel Prestage for useful discussions and to USAF staff at Learmonth for cooperation. The authors thank the Australian Research Council for funding. References Balch, C., B. Murtagh, L. T. D. Zezula, L. Combs, G. Nelson, K. Tegnell, M. Crown, and B. McGehan (2004), Intense space weather storms October 19--November 7, 2003, serv. assess., U. S. Dep. of Commer, Silver Spring, Md. (Available at SWstorms_assessment.pdf) Cane, H. V., and D. V. Reames (1988), Soft X-ray emissions, meterwavelength radio bursts, and particle acceleration in solar flares, Astrophys. J., 325, , doi: / Cerruti, A. P., P. M. Kintner, D. E. Gary, A. J. Mannucci, R. F. Meyer, P. Doherty, and A. J. Coster (2008), Effect of intense December 2006 solar radio bursts on GPS receivers, Space Weather, 6, S10D07, doi: /2007sw Deans, S. R. (1983), The Radon Transform and Some of Its Applications, John Wiley, New York. Dulk, G. A., D. J. McLean, and G. J. Nelson (1985), Solar flares, in Solar Radiophysics, edited by D. J. McLean and N. R. Labrum, pp , Cambridge Univ. Press, Cambridge, U. K. Kennewell, J., and G. Steward (2003), Solar Radio Spectrograph [SRS] data viewer [Srsdisplay], technical report, IPS Radio and Space Serv., Sydney. (Available at ftp://ftp.ngdc.noaa.gov/stp/solar_ DATA/SOLAR_RADIO/SPECTRAL_RSTN/Srsdispl.doc.) Lobzin, V. V., I. H. Cairns, and P. A. Robinson (2008), Evidence for wind-like regions, acceleration of shocks in the deep corona, and relevance of 1/f dynamic spectra to coronal type II bursts, Astrophys. J., 677, L129-- L132, doi: / Nelson, G. J., and D. B. Melrose (1985), Type II bursts, in Solar Radiophysics, edited by D. J. McLean and N. R. Labrum, pp , Cambridge Univ. Press, Cambridge, U. K. Neyman, J., and E. S. Pearson (1967a), On the use and interpretation of certain test criteria for purposes of statistical inference, part I, 11 of 12

12 in Joint Statistical Papers, edited by J. Neyman and E. S. Pearson, pp , Cambridge Univ. Press, Cambridge, U. K. Neyman, J., and E. S. Pearson (1967b), The testing of statistical hypotheses in relation to probabilities a priori, in Joint Statistical Papers, edited by J. Neyman and E. S. Pearson, pp , Cambridge Univ. Press, Cambridge, U. K. Pick, M., and N. Vilmer (2008), Sixty-five years of solar radioastronomy: Flares, coronal mass ejections and Sun-Earth connection, Astron. Astrophys. Rev., 16, , doi: /s x. Qu, M., F. Y. Shih, J. Jing, and H. Wang (2006), Automatic detection and classification of coronal mass ejections, Sol. Phys., 237, , doi: /s Robbrecht, E., and D. Berghmans (2004), Automated recognition of coronal mass ejections (CMEs) in near-real-time data, Astron. Astrophys., 425, , doi: / : Schlegel, K. (2005), Space weather effects in the upper atmosphere: High altitudes, in Space Weather: The Physics Behind a Slogan, edited by K. Scherer et al., pp , Springer, Berlin. Sugiura, M., and S. Chapman (1960), The average morphology of geomagnetic storms with sudden commencement, Abh. Akad. Wiss. Gottingen Math Phys., 4 (sorderh.), Suzuki, S., and G. A. Dulk (1985), Bursts of type III and type V, in Solar Radiophysics, edited by D. J. McLean and N. R. Labrum, pp , Cambridge Univ. Press, Cambridge, U. K. Terkildsen, M. (2007), Solar radio burst effects on GPS: A study of the effects of the December 2006 solar activity on GPS in the Australian region, paper presented at 7th Australian Space Science Conference, Natl. Comm. for Space Sci., Sydney, N. S. W. Warmuth, A., and G. Mann (2005), The application of radio diagnostics to the study of the solar drivers of space weather, in Space Weather: The Physics Behind a Slogan, edited by K. Scherer et al., pp , Springer, Berlin. White, S. M. (2007), Solar radio bursts and space weather, Asian J. Phys., 16, I. H. Cairns, V. V. Lobzin, and P. A. Robinson, School of Physics, University of Sydney, Sydney, NSW 2006, Australia. (v.lobzin@ physics.usyd.edu.au) G. Patterson and G. Steward, IPS Radio and Space Services, Level 6, 477 Pitt Street, Sydney, NSW 2000, Australia. 12 of 12