RE – EVALUATION OF GEOMAGNETIC FIELD OBSERVATION DATA AT SYOWA STATION, ANTARCTICA

The Japanese Antarctic Research Expedition has conducted geomagnetic observations at Syowa Station, Antarctica, since 1966. Geomagnetic variation data measured with a fluxgate magnetometer are not absolute but are relative to a baseline and show drift. To enhance the importance of the geomagnetic data at Syowa S tation, therefore, it is necessary to correct the continuous variation data by using absolute baseline values acquired by a magnetic theodolite and proton magnetometer. However, the database of baseline values contains outliers. We detected outliers in the database and then converted the geomagnetic variation data to absolute values by using the reliable baseline values.


INTRODUCTION
Since 1966, the Japanese Antarctic Research Expedition (JARE) has conducted absolute geomagnetic observations and geomagnetic variation measurements at Syowa Station, Antarctica (N69.006, E39.590; Figure 1). The absolute geomagnetic observations have been carried out basically once per month during geomagnetically quiet periods with a magnetic theodolite and proton magnetometer to obtain absolute baseline values. The geomagnetic variation measurements are performed with a three-axis fluxgate magnetometer to obtain continuous readings of variations relative to the baseline. The results are publicly available on the website of the National Institute of Polar Research, Japan (http://polaris.nipr.ac.jp/~aurora/syowa.magne/magne.main.html). The continuous geomagnetic variation data acquired with the fluxgate magnetometer in general showed a drift caused by changes in the sensor temperature and/or tilt.
To convert the continuous geomagnetic variation record to absolute values, it was necessary to correct such drift. Therefore, we used the following correction procedure ( Figure 2): (1) Calculate each baseline value (I B ) at each time of the absolute observations (t A ) from each absolute baseline value (I A ) and the continuous variation data (I C ) at t A as I B = I A -I C .
(2) Calculate the baseline values during other periods by linear interpolation between the successive baseline values at the times of the absolute observations.
(3) Add the interpolated baseline values to the continuous variation data to obtain absolute variation data.
However, the database of baseline values contains outliers. Therefore, we developed a statistical procedure for objective detection of outliers in the baseline values.

GEOMAGNETIC OBSERVATIONS AT SYOWA STATION
The three-axis fluxgate magnetometer used by JARE at Syowa Station since 1966 measures three components of geomagnetic variation: the components parallel and perpendicular to the geomagnetic meridian and the vertical component. The variation data are sampled every 1, 2, or 10 s, and the resolving power is 0.1 nT. During the same period, JARE has made absolute geomagnetic observations to investigate the secular variation of declination, inclination, and geomagnetic intensity. These observations are obtained approximately monthly with a magnetic theodolite and proton magnetometer. A search coil magnetometer, which is the G.S.I. (Geographical Survey Institute of Japan) type magnetometer, was used from Mar., 1966 to May, 1997, and after that, a fluxgate declinometer/inclinometer has been used (Ookawa, 1999). The G.S.I. type magnetometer consists of a rotating search coil and a Helmholtz coil that are mounted on a steel-free theodolite. Its minimum detectable angle is 0.2 minutes. The fluxgate declinometer/inclinometer consists of a single-axis magnetometer with a fluxgate probe mounted on a steel-free theodolite. Its observation accuracy is 0.1 nT. The observation accuracy of the proton magnetometer is better than 0.1 nT, and the accuracy of the absolute observation depends not only on the accuracy of the instruments but also on the individual observer's skill at the operation. In each absolute observation session, four observed baseline values are averaged to determine the baseline value for that session. Baseline values for the horizontal (H) and vertical (Z) components and the declination (D) of the continuous variation data are calculated by using the absolute observations. Figure 3 shows the calculated baseline values from 1997 to 2011. It can be seen in Figure 3 that all components of the baseline values measured during some absolute observation sessions had a large variance. Then a statistical approach is applied to detect outliers in the observed baseline values.

DETECTING OUTLIERS IN THE DATABASE OF BASELINE VALUES
For objective detection of outliers in a series of observed baseline values, Ito and Fujii (2003) proposed a robust estimation procedure that uses the median and median absolute deviation of the observed data. However, it was difficult to apply this robust estimation procedure to our database because only four baseline values were obtained per observation session.
We therefore investigated the distribution of residuals (i.e., differences between the mean and each of the four observed baseline values) for all observations from 1997 to 2011. During this period, 720 observed baseline values were obtained for each component. The frequency distribution of the residuals of each component is bell-shaped, with only a few large-amplitude samples (Figure 4), which suggests that we can assume a Gaussian-type distribution. In addition, we examined the distribution of the residuals of each component in a

Geomagnetic variations
Data Science Journal, Volume 12, 13 May 2013 WDS244 normal quantile-quantile plot ( Figure 5), in which residuals normalized by the standard deviation of the residuals on the vertical axis are compared to the theoretical residuals predicted given a standard Gaussian distribution. For all components, the normalized residuals with a normalized residual size of about 3 or larger deviate from the 45° reference line. Therefore, we regarded the observed baseline values with normalized residuals in excess of 3 as outliers because these residuals are larger than would be expected given a Gaussian distribution. This is equivalent to determining as outliers residual amplitudes of the H-, Z-, and D-components in excess of 2.8 nT, 2.0 nT, and 0.7 minutes, respectively. In this way, we identified a total of 42 residuals as outliers: 16 in the H-component, 10 in the Z-component, and 16 in the D-component (Figure 3).