Limitations of the shower shape cuts

Cuts on the aforementioned shower shape parameters have proven to effectively reduce the hadronic cosmic-ray background independent of the observational conditions and primary particle energies. However, it is clear that the Hillas parametrisation just employs global shower properties, such as the width and length of the images. Shower properties on the pixel level are not taken into account by the Hillas approach. Also the stereoscopic information is not fully employed by the M RSW and M RSL parameters. As mentioned in Section 1.1.2, the subclass of hadronic cosmic rays, which transfer most of their energy into the electromagnetic part of the shower are not easily distinguishable from γ-ray induced showers. Information stored in the image shape parameters is not sufficient to identify this class of background events. However, there is some potential to discriminate those by

using the shower maximum Xmax as a discriminating parameter.

In the next chapter not only the shower maximum but also additional parameters with classification power are introduced and combined in a multivariate analysis method which is based on decision trees. The applicability of this approach and its enormous potential in terms of γ/hadron separation is demonstrated. Moreover, the improvement in sensitivity due to additional parameters and due to the method are examined separately.

2 Improved γ/hadron separation using a

multivariate analysis technique

As shown in Chapter 1, cuts on the MRSW and MRSL parameters can suppress a large fraction of cosmic-ray background events. This approach has proven to be robust and has been successfully applied in the analysis of H.E.S.S. data. H.E.S.S. scanned the inner parts of the Milky Way and established more than ∼ 50 new sources of different type as VHE γ-ray emitters, such as e.g. SNRs, PWNe or γ-ray binaries. VHE γ-ray emission was also detected from extragalactic objects such as Active Galactic Nuclei (AGN) or Radio galaxies (see e.g. Hinton & Hofmann 2009, for a recent review). All these sources emit VHE γ rays at a flux level, which made their detection possible with H.E.S.S. within observation times of 25 hours or less – a significantly smaller amount of time than the available dark time per year of ≈1700 hours. However, after five years of operation, H.E.S.S. is now getting into the regime, where presumably all strong VHE γ-ray sources have been detected and where a significantly larger amount of observation time is needed to establish weaker sources as VHE γ-ray emitters. Especially extragalactic objects such as Starburst galaxies or Galaxy clusters and Galactic objects such as colliding wind binaries (CWB) are predicted to emit γ rays at a flux level which is right at the edge or slightly below the H.E.S.S. sensitivity1. In order to increase the sensitivity of IACT systems, additional telescopes, spread over a larger area on ground are needed. The next generation of ground-based VHE γ-ray telescope systems is currently being studied by the CTA (Hermann et al. 2008) and AGIS (Fegan et al. 2008) consortia and will start operation within in the next 10 − 15 years.

Still, the sensitivity for existing instruments can considerably be improved by an increased background reduction. Compared to the Hillas approach which utilises the two-dimensional shape of the recorded EAS images in the camera for shower reconstruction and γ/hadron separation, more elaborated analysis methods have been developed and successfully em- ployed. The 3D Model Analysis, introduced by Lemoine-Goumard et al. (2006), compares the recorded images with a three-dimensional photosphere model of the shower, achieving a similar performance as the H.E.S.S. Standard Analysis. Beyond that, the application of multivariate analysis (MVA) techniques, such as Random Forests (RF) (Breiman 2001), have been studied and successfully applied to a single IACT recently (Bock et al. 2004; Al- bert et al. 2008). RFs have also been utilised in H.E.S.S., in an analysis especially designed for the study of cosmic-ray electrons (Egberts 2005, 2009). The H.E.S.S. result, presented in Aharonian et al. (2008a), had strong impacts on the field of dark matter physics and was highly regarded in the astronomical and physics community.

In this chapter the application of the Boosted Decision Tree (BDT) method (which is

1

Here, sensitivity is defined as the minimum flux Fγ emitted by a source which is measured on Earth as signal Nγ at a given confidence level.

integrated in the TMVA package (Hoecker et al. 2007)) to data obtained by H.E.S.S. is discussed. The method presented in the following can be applied for γ/hadron separation and the study of VHE γ-ray sources, independent of the number of telescopes. Beyond the first performance studies of a telescope-independent, decision-tree-based γ/hadron separation in ground-based VHE γ-ray astronomy, presented in Ohm (2007), this work focuses on a thorough systematic study of shower shape parameters with classification potential and their integration in the BDT method, taking into account the dynamical properties of the recorded data such as the zenith angle or event energy (Section 2.1). The basic working principle of the BDT and its advantages compared to other MVA methods is discussed in Section 2.2, before in Section 2.3 the training and evaluation of the BDT method is described. Section 2.4 presents detailed systematic tests using H.E.S.S. data based on observations of several astrophysical objects. Finally, the performance and increased sensitivity of the BDT approach using Monte Carlo γ-ray simulations and background data is discussed in Section 2.52_.

2.1 Parameters with γ/hadron separation potential

In the H.E.S.S. Standard Analysis cuts on the shower shape parameters MRSW and MRSL are applied to select γ-ray like events and to reject the hadronic background. The sepa- ration potential is illustrated in Fig. 2.1(a), where the distribution of the two parameters is shown for Monte Carlo γ rays and Off data. Although the application of these box cuts suppresses a large fraction of the charged cosmic rays, they apparently do not fully explore the available information stored in the two parameters. An improvement could already be achieved, if a cut on MRSW as a function of the MRSL value would be applied.

Apart from these two shower shape parameters, which utilise the width and length of the EAS image in the camera, more information about the shower origin, its spatial intensity distribution and interaction processes during the shower development are stored in the recorded images. In this section, additional shower shape parameters which are based on the Hillas parameters and have γ/hadron separation potential are introduced and their properties are investigated. These parameters are then fed into the BDT algorithm, which is presented in Section 2.2.