2. AUTOMATED MORPHOLOGICAL CLASSIFICATION IN DEEP HST FIELDS

2.1. WFPC2 observations, reduction, and ANN classifications

2.1.1. Primary and parallel observations in B, V, I, and/or U

In Cycle 4-5, a single deep 63-orbit dithered WFPC2 field (hereafter the "W02" field) was imaged surrounding the radio galaxy 53W002 at z = 2.39, with 12 orbits in both V₆₀₆ & I₈₁₄ (Driver et al. 1995; D95a), plus 24 orbits in B₄₅₀ (Odewahn et al. 1996; O96), and 15 orbits in the medium-band filter F410M (_eff 4090 Å or Ly at z 2.36±0.06; Pascarelle et al. 1996b; P96b). The 48-orbit stack of BVI images is shown as a color Plate in W98. Calibration followed D95a, D95b & O96, using the best available WFPC2 super-sky flats. These deep WFPC2 images provided a 5 point source detection limit for B₄₅₀, V₆₀₆ 28.4, & I₈₁₄ 27.6 mag, and a 1 surface brightness (SB) sensitivity of B, V 27.5 & I 26.7 mag arcsec^-2 (see W98 & D95a). O96 compare the galaxy counts in the 53W002 field to those from UBVI images in the HDF, which reach ~ 0.8-1.2 mag deeper. In Cycle 6, 29 shallower B₄₅₀ fields were collected in the WFPC2 B-band Parallel Survey (S. Cohen et al. 1997, Co97), as well as from several other Archival WFPC2 B₄₅₀ and I₈₁₄-band fields from the Medium Deep Survey (R97), and from other Archival sources to get better statistics at the bright-end (19 B 25.5 mag, see Fig. 1 - 3).

Figure 1 and 2. (1) [Left] The B450 magnitude vs. half-light radius rhl for all classified galaxies in the W02 field and the HDF. Symbols indicate membership in the categories E/S0, Sabc, & Sd/Irr+M classified by the rest-frame ANN (Section 2.1.4). The solid almost-horizontal curves indicates the B450-detection limits. The almost-vertical curves indicate how the median scale-length of RC3 galaxies of given Hubble type and MB decline towards fainter magnitude (see O96). Galaxies classified as E/S0 have on average smaller observed scale-lengths than Sabc's, which are generally smaller than Sd/Irr galaxies. The observed scale-lengths reach a median of rhl 0".25-0".3 at B 27 mag; (2) [Right] The (B - R) vs. B450 color-magnitude diagram for all classified galaxies, as in Fig. 1. The two (slanted) dashed lines indicate the completeness limits in (R and) B for the two fields. The same ANN galaxy classes are indicated as in Fig. 1, which are reliable for B 27 mag. The reddest objects at B 27 mag are mostly classified as E/S0's and Sabc's, and do not increase as rapidly towards the completeness limits, suggesting that the formation of early-type galaxies was largely complete by z ~ 1. The late-type+merging population galaxies increases rapidly to the detection and classification limit (B ~ 27 mag; see Fig. 3d). A color version of Figs. 1-3 is given by O96.

2.1.2. Data processing and catalog generation

All images in each filter were spatially registered, CR-clipped, and averaged to create high S/N composites (see W94a, W98), from which object catalogs were computed (cf. D95a, O96). Potential objects were located and measured using aperture magnitudes grown-to-total (cf. P96b, O96). These (U), B, V, I catalogs were used to produce a final catalog by selecting any object present in any filter with S/N > 3. Hence, very red or very blue objects are not excluded. Surface photometry was carried out with the automated image analysis package MORPHO (O95), giving ellipse fits over a range of isophotal levels, elliptically averaged SB-profiles, and a set of type-dependent photometric parameters (O96).

2.1.3. ANN classifications

O96 assigned a morphological type to each galaxy in several independent ways. First, 173 galaxies were classified by eye with U₃₆₀ 26 mag, 372 with B₄₅₀ 26.5 mag, and 542 with V₆₀₆, I₈₁₄ 26 mag, obtaining results mutually consistent within ±2 (rms) Hubble classes, and scale errors 10% on the 16-step Hubble scale. All objects with B₄₅₀ 27.5 mag were classified using an ANN analysis of the photometric parameters (O92, O95). ANN's are systems of weight vectors, whose component values are established through an iterative learning algorithm such as back-propagation. They take as input a linear set of patterns, and produce as output a numerical pattern encoding an object classification. O95 show that this pattern can be used to assign a confidence value to the estimated galaxy type. Fundamental to the development of such pattern classifiers is the existence of a large sample of examples (or "training set"). A set of ANN classifiers was developed for the I₈₁₄ and V₆₀₆ images using the morphological I₈₁₄ eye-ball classifications as a training set, as described by D95a, D95b, & O96. Each classifier inspected both the digital image and the SB profile for each galaxy in each filter. Hence, some profile information was incorporated in these eyeball estimates. With the mean visual types from our samples in U, B, V and/or I, ANN galaxy classifiers were developed using six primary parameters: the SB at the 25% and 75% quartile radii, the mean SB within the effective radius and within an isophotal radius, as well as the slope and intercept of a linear fit to the SB-profile (in r^1/4 space). These ANN's are based on the observed galaxy SB-profile, but not on color, nor on scale-length. ANN classifiers were also developed to classify galaxies in their appropriate rest-frame UBV filters.

Figure 3. (a) B-band counts from the W02 field (open circles), the HDF (open triangles), the WFPC2 B-band parallel survey (filled circles), and other Archival B450 fields (open squares), as well as previous ground-based BJ-band counts (crosses). Panels show: (a) all galaxies, (b) E/S0's, (c) early-type spirals (Sabc's), and (d) late-type/Irregulars Sd/Irr+M's. The Merger (M) fraction is given by filled blue or shaded squares. Plotted are mean types from: (1) our visually classified images (color-coded or shaded lines); (2) the single-filter ANN types (open data points), and (3) from the rest-frame ANN types (heavy lines and filled circles). All models are plotted as thin curves (see 2.2.3). A color version of Fig. 3 is shown in O96.

2.1.4. Tests of the classification systems as a function of wavelength and redshift

Since for faint galaxies (B 24 mag) the expected median redshift is z 0.7 (KK92; C96), the B₄₅₀ filter could look back into the rest-frame mid-UV. At higher redshifts (and fainter fluxes), V₆₀₆ and even I₈₁₄ may have the same problem. One therefore has to be concerned how the cosmological (1 + z)⁴ SB-dimming and the wavelength-dependence of the rest-frame (UV-)morphology may affect the ANN classifications. As discussed by O96, little or no systematic change was observed in the eyeball morphological types between the U, B, V & I filters (see Fig. 3 here and Fig. 1 of Burg et al., this volume). This suggests that the properties on which our human classifiers rely (morphology and light profile) are more uniform with spectral bandpass than we would expect (cf. dJ94). Alternatively, it could suggest that the median redshift for the faint galaxy sample is somewhat smaller (z ~ 1.0 for B 25) and the redshift distribution rather much wider than expected (SLY97), so that cosmological effects possibly do not yet differ greatly between the B, V, & I filters. To investigate this further, a set of ANN classifiers was designed that is valid for rest-frame filters. For this, the following steps were used: (1) the likely redshift range was estimated for each B₄₅₀ interval using the B redshift distributions of KK92 and E96, and the models for Gunn g 26 mag of NW95a; (2) using the known UBVI filter responses, these redshift ranges were used to estimate the most likely rest-frame central wavelength for each galaxy image; (3) a series of rest-frame U, B, and/or V images was selected, and the corresponding rest-frame ANN's were derived. For galaxies with B₄₅₀ 26 mag, these rest-frame ANN types were compared to the UBVI eyeball classifications, as well as to the types predicted by the single-filter ANN classifiers. The rest-frame predictions for B 27 (heavy black lines in Fig. 3) do not differ significantly from the mean eyeball estimates (color-coded or shaded lines), nor from the single-filter ANN estimates (open data points). The four-color counts in Fig. 3 show no obvious evidence that the shorter wavelength filters result in a larger fraction of late-type galaxies. The W02 and HDF samples yield consistent classes as a function of flux (see Fig. 3), yet have S/N that differs by 1.2-1.5 mag at a given flux. For B 27 mag both methods of ANN classification are therefore reliable, but for B 27 significant discrepancies may exist between the I₈₁₄ or V₆₀₆ ANN's and the rest-frame ANN's, indicating that Z_med may be 2 and that the uncertain far-UV morphology becomes important. Further tests of systematic errors are discussed by O96. About 15% of all galaxies with B₄₅₀ 25.5 mag and classified by the ANN systems as late-type spirals or irregulars were classified as merging systems by the human classifiers. For B₄₅₀ 25.5 mag, the number of merging systems increases to 35% of those ANN-classed Sd/Irr galaxies. Hence, a good fraction of "late-type" systems as classified by the ANN are actually merger morphologies, so this class is designated as Sd/Irr+M systems.