-
Channel-Based Attention for LCC Using Sentinel-2 Time Series
Authors:
Hermann Courteille,
A. Benoît,
N Méger,
A Atto,
D. Ienco
Abstract:
Deep Neural Networks (DNNs) are getting increasing attention to deal with Land Cover Classification (LCC) relying on Satellite Image Time Series (SITS). Though high performances can be achieved, the rationale of a prediction yielded by a DNN often remains unclear. An architecture expressing predictions with respect to input channels is thus proposed in this paper. It relies on convolutional layers…
▽ More
Deep Neural Networks (DNNs) are getting increasing attention to deal with Land Cover Classification (LCC) relying on Satellite Image Time Series (SITS). Though high performances can be achieved, the rationale of a prediction yielded by a DNN often remains unclear. An architecture expressing predictions with respect to input channels is thus proposed in this paper. It relies on convolutional layers and an attention mechanism weighting the importance of each channel in the final classification decision. The correlation between channels is taken into account to set up shared kernels and lower model complexity. Experiments based on a Sentinel-2 SITS show promising results.
△ Less
Submitted 31 March, 2021;
originally announced March 2021.
-
Parametric Rectified Power Sigmoid Units: Learning Nonlinear Neural Transfer Analytical Forms
Authors:
Abdourrahmane Mahamane Atto,
Sylvie Galichet,
Dominique Pastor,
Nicolas Méger
Abstract:
The paper proposes representation functionals in a dual paradigm where learning jointly concerns both linear convolutional weights and parametric forms of nonlinear activation functions. The nonlinear forms proposed for performing the functional representation are associated with a new class of parametric neural transfer functions called rectified power sigmoid units. This class is constructed to…
▽ More
The paper proposes representation functionals in a dual paradigm where learning jointly concerns both linear convolutional weights and parametric forms of nonlinear activation functions. The nonlinear forms proposed for performing the functional representation are associated with a new class of parametric neural transfer functions called rectified power sigmoid units. This class is constructed to integrate both advantages of sigmoid and rectified linear unit functions, in addition with rejecting the drawbacks of these functions. Moreover, the analytic form of this new neural class involves scale, shift and shape parameters so as to obtain a wide range of activation shapes, including the standard rectified linear unit as a limit case. Parameters of this neural transfer class are considered as learnable for the sake of discovering the complex shapes that can contribute in solving machine learning issues. Performance achieved by the joint learning of convolutional and rectified power sigmoid learnable parameters are shown outstanding in both shallow and deep learning frameworks. This class opens new prospects with respect to machine learning in the sense that learnable parameters are not only attached to linear transformations, but also to suitable nonlinear operators.
△ Less
Submitted 5 February, 2021; v1 submitted 25 January, 2021;
originally announced January 2021.
-
Clustering properties of BzK-selected galaxies in GOODS-N: environmental quenching and triggering of star formation at z ~ 2
Authors:
Lihwai Lin,
Mark Dickinson,
Hung-Yu Jian,
A. I. Merson,
C. M. Baugh,
Douglas Scott,
Sebastien Foucaud,
Wei-Hao Wang,
Chi-Hung Yan,
Hao-Jing Yan,
Yi-Wen Cheng,
Yicheng Guo,
John Helly,
Franz Kirsten,
David C. Koo,
Claudia del P. Lagos,
Nicole Meger,
Alexandra Pope,
Luc Simard,
Norman A. Grogin,
Hugo Messias,
Shiang-Yu Wang
Abstract:
Using a sample of BzK-selected galaxies at z~2 identified from the CFHT/WIRCAM near-infrared survey of GOODS-North, we discuss the relation between star formation rate (SFR), specific star formation rate (SSFR), and stellar mass (M_{*}), and the clustering of galaxies as a function of these parameters. For star-forming galaxies (sBzKs), the UV-based SFR, corrected for extinction, scales with the s…
▽ More
Using a sample of BzK-selected galaxies at z~2 identified from the CFHT/WIRCAM near-infrared survey of GOODS-North, we discuss the relation between star formation rate (SFR), specific star formation rate (SSFR), and stellar mass (M_{*}), and the clustering of galaxies as a function of these parameters. For star-forming galaxies (sBzKs), the UV-based SFR, corrected for extinction, scales with the stellar mass as SFR ~ M_{*}^{alpha} with alpha = 0.74+/-0.20 down to M_{*} ~ 10^{9} M_{solar}, indicating a weak dependence on the stellar mass of the SSFR. We also measure the angular correlation function and hence infer the correlation length for sBzK galaxies as a function of M_{*}, SFR, and SSFR, as well as K-band apparent magnitude. We show that passive galaxies (pBzKs) are more strongly clustered than sBzK galaxies at a given stellar mass, mirroring the color-density relation seen at lower redshifts. We also find that the correlation length of sBzK galaxies ranges from 4 to 20 h^{-1}Mpc, being a strong function of M_{K}, M_{*}, and SFR. On the other hand, the clustering dependence on SSFR changes abruptly at 2x10^{-9} yr^{-1}, which is the typical value for "main sequence" star-forming galaxies at z~2. We show that the correlation length reaches a minimum at this characteristic value, and is larger for galaxies with both smaller and larger SSFRs; a dichotomy that is only marginally implied from the predictions of the semi-analytical models. Our results suggest that there are two types of environmental effects at work at z~2. Stronger clustering for relatively quiescent galaxies implies that the environment has started to play a role in quenching star formation. At the same time, stronger clustering for galaxies with elevated SSFRs ("starbursts") might be attributed to an increased efficiency for galaxy interactions and mergers in dense environments.
△ Less
Submitted 6 July, 2012; v1 submitted 9 November, 2011;
originally announced November 2011.
-
The nature of faint Spitzer-selected dust-obscured galaxies
Authors:
Alexandra Pope,
R. Shane Bussmann,
Arjun Dey,
Nicole Meger,
David M. Alexander,
Mark Brodwin,
Ranga-Ram Chary,
Mark E. Dickinson,
David T. Frayer,
Thomas R. Greve,
Minh Huynh,
Lihwai Lin,
Glenn Morrison,
Douglas Scott,
Chi-Hung Yan
Abstract:
We use deep far-IR, submm, radio and X-ray imaging and mid-IR spectroscopy to explore the nature of a sample of Spitzer-selected dust-obscured galaxies (DOGs) in GOODS-N. A sample of 79 galaxies satisfy the criteria R-[24]>14 (Vega) down to S24>100 microJy. Twelve of these galaxies have IRS spectra available which we use to measure redshifts and classify these objects as being dominated by star…
▽ More
We use deep far-IR, submm, radio and X-ray imaging and mid-IR spectroscopy to explore the nature of a sample of Spitzer-selected dust-obscured galaxies (DOGs) in GOODS-N. A sample of 79 galaxies satisfy the criteria R-[24]>14 (Vega) down to S24>100 microJy. Twelve of these galaxies have IRS spectra available which we use to measure redshifts and classify these objects as being dominated by star formation or active galactic nuclei (AGN) activity in the mid-IR. The IRS spectra and Spitzer photometric redshifts confirm that the DOGs lie in a tight redshift distribution around z~2. Based on mid-IR colors, 80% of DOGs are likely dominated by star formation; the stacked X-ray emission from this sub-sample of DOGs is also consistent with star formation. Since only a small number of DOGs are individually detected at far-IR and submm wavelengths, we use a stacking analysis to determine the average flux from these objects and plot a composite IR (8-1000 microns) spectral energy distribution (SED). The average luminosity of these star forming DOGs is LIR~1e12 Lsun. We compare the average star forming DOG to the average bright (S850>5 mJy) submillimeter galaxy (SMG); the S24>100 microJy DOGs are 3 times more numerous but 8 times less luminous in the IR. The far-IR SED shape of DOGs is similar to that of SMGs (average dust temperature of around 30 K) but DOGs have a higher mid-IR to far-IR flux ratio. The average star formation-dominated DOG has a star formation rate of 200 Msun/yr which, given their space density, amounts to a contribution of 0.01 Msun/yr/Mpc3 (or 5-10%) to the star formation rate density at z~2. We use the composite SED to predict the average flux of DOGs in future Herschel/PACS 100 micron and SCUBA-2 450 micron surveys and show that the majority of them will be detected.
△ Less
Submitted 20 August, 2008;
originally announced August 2008.
