-
Self-Supervised Pretraining on Satellite Imagery: a Case Study on Label-Efficient Vehicle Detection
Authors:
Jules BOURCIER,
Thomas Floquet,
Gohar Dashyan,
Tugdual Ceillier,
Karteek Alahari,
Jocelyn Chanussot
Abstract:
In defense-related remote sensing applications, such as vehicle detection on satellite imagery, supervised learning requires a huge number of labeled examples to reach operational performances. Such data are challenging to obtain as it requires military experts, and some observables are intrinsically rare. This limited labeling capability, as well as the large number of unlabeled images available…
▽ More
In defense-related remote sensing applications, such as vehicle detection on satellite imagery, supervised learning requires a huge number of labeled examples to reach operational performances. Such data are challenging to obtain as it requires military experts, and some observables are intrinsically rare. This limited labeling capability, as well as the large number of unlabeled images available due to the growing number of sensors, make object detection on remote sensing imagery highly relevant for self-supervised learning. We study in-domain self-supervised representation learning for object detection on very high resolution optical satellite imagery, that is yet poorly explored. For the first time to our knowledge, we study the problem of label efficiency on this task. We use the large land use classification dataset Functional Map of the World to pretrain representations with an extension of the Momentum Contrast framework. We then investigate this model's transferability on a real-world task of fine-grained vehicle detection and classification on Preligens proprietary data, which is designed to be representative of an operational use case of strategic site surveillance. We show that our in-domain self-supervised learning model is competitive with ImageNet pretraining, and outperforms it in the low-label regime.
△ Less
Submitted 21 October, 2022;
originally announced October 2022.
-
Evaluating the Label Efficiency of Contrastive Self-Supervised Learning for Multi-Resolution Satellite Imagery
Authors:
Jules BOURCIER,
Gohar Dashyan,
Jocelyn Chanussot,
Karteek Alahari
Abstract:
The application of deep neural networks to remote sensing imagery is often constrained by the lack of ground-truth annotations. Adressing this issue requires models that generalize efficiently from limited amounts of labeled data, allowing us to tackle a wider range of Earth observation tasks. Another challenge in this domain is develo** algorithms that operate at variable spatial resolutions, e…
▽ More
The application of deep neural networks to remote sensing imagery is often constrained by the lack of ground-truth annotations. Adressing this issue requires models that generalize efficiently from limited amounts of labeled data, allowing us to tackle a wider range of Earth observation tasks. Another challenge in this domain is develo** algorithms that operate at variable spatial resolutions, e.g., for the problem of classifying land use at different scales. Recently, self-supervised learning has been applied in the remote sensing domain to exploit readily-available unlabeled data, and was shown to reduce or even close the gap with supervised learning. In this paper, we study self-supervised visual representation learning through the lens of label efficiency, for the task of land use classification on multi-resolution/multi-scale satellite images. We benchmark two contrastive self-supervised methods adapted from Momentum Contrast (MoCo) and provide evidence that these methods can be perform effectively given little downstream supervision, where randomly initialized networks fail to generalize. Moreover, they outperform out-of-domain pretraining alternatives. We use the large-scale fMoW dataset to pretrain and evaluate the networks, and validate our observations with transfer to the RESISC45 dataset.
△ Less
Submitted 13 October, 2022;
originally announced October 2022.
-
Improving performance of aircraft detection in satellite imagery while limiting the labelling effort: Hybrid active learning
Authors:
Julie Imbert,
Gohar Dashyan,
Alex Goupilleau,
Tugdual Ceillier,
Marie-Caroline Corbineau
Abstract:
The earth observation industry provides satellite imagery with high spatial resolution and short revisit time. To allow efficient operational employment of these images, automating certain tasks has become necessary. In the defense domain, aircraft detection on satellite imagery is a valuable tool for analysts. Obtaining high performance detectors on such a task can only be achieved by leveragi…
▽ More
The earth observation industry provides satellite imagery with high spatial resolution and short revisit time. To allow efficient operational employment of these images, automating certain tasks has become necessary. In the defense domain, aircraft detection on satellite imagery is a valuable tool for analysts. Obtaining high performance detectors on such a task can only be achieved by leveraging deep learning and thus us-ing a large amount of labeled data. To obtain labels of a high enough quality, the knowledge of military experts is needed.We propose a hybrid clustering active learning method to select the most relevant data to label, thus limiting the amount of data required and further improving the performances. It combines diversity- and uncertainty-based active learning selection methods. For aircraft detection by segmentation, we show that this method can provide better or competitive results compared to other active learning methods.
△ Less
Submitted 10 February, 2022;
originally announced February 2022.
-
QUBIC II: Spectro-Polarimetry with Bolometric Interferometry
Authors:
L. Mousset,
M. M. Gamboa Lerena,
E. S. Battistelli,
P. de Bernardis,
P. Chanial,
G. D'Alessandro,
G. Dashyan,
M. De Petris,
L. Grandsire,
J. -Ch. Hamilton,
F. Incardona,
S. Landau,
S. Marnieros,
S. Masi,
A. Mennella,
C. O'Sullivan,
M. Piat,
G. Ricciardi,
C. G. Scóccola,
M. Stolpovskiy,
A. Tartari,
J. -P. Thermeau,
S. A. Torchinsky,
F. Voisin,
M. Zannoni
, et al. (106 additional authors not shown)
Abstract:
Bolometric interferometry is a novel technique that has the ability to perform spectral imaging. A bolometric interferometer observes the sky in a wide frequency band and can reconstruct sky maps in several sub-bands within the physical band in post-processing of the data. This provides a powerful spectral method to discriminate between the cosmic microwave background (CMB) and astrophysical foreg…
▽ More
Bolometric interferometry is a novel technique that has the ability to perform spectral imaging. A bolometric interferometer observes the sky in a wide frequency band and can reconstruct sky maps in several sub-bands within the physical band in post-processing of the data. This provides a powerful spectral method to discriminate between the cosmic microwave background (CMB) and astrophysical foregrounds. In this paper, the methodology is illustrated with examples based on the Q \& U Bolometric Interferometer for Cosmology (QUBIC) which is a ground-based instrument designed to measure the B-mode polarization of the sky at millimeter wavelengths. We consider the specific cases of point source reconstruction and Galactic dust map** and we characterize the point spread function as a function of frequency. We study the noise properties of spectral imaging, especially the correlations between sub-bands, using end-to-end simulations together with a fast noise simulator. We conclude showing that spectral imaging performance are nearly optimal up to five sub-bands in the case of QUBIC.
△ Less
Submitted 28 March, 2022; v1 submitted 28 October, 2020;
originally announced October 2020.
-
Cosmic ray feedback from supernovae in dwarf galaxies
Authors:
Gohar Dashyan,
Yohan Dubois
Abstract:
The regulation of the baryonic content in dwarf galaxies is a long-standing problem. Supernovae (SNe) are supposed to play a key role in forming large-scale galactic winds by removing important amounts of gas from galaxies. SNe are efficient accelerators of non-thermal particles, so-called cosmic rays (CRs), which can substantially modify the dynamics of the gas and conditions to form large-scale…
▽ More
The regulation of the baryonic content in dwarf galaxies is a long-standing problem. Supernovae (SNe) are supposed to play a key role in forming large-scale galactic winds by removing important amounts of gas from galaxies. SNe are efficient accelerators of non-thermal particles, so-called cosmic rays (CRs), which can substantially modify the dynamics of the gas and conditions to form large-scale galactic winds. We investigate how CR injection by SNe impacts the star formation and the formation of large-scale winds in dwarf galaxies, and whether it can produce galaxy star-formation rates (SFR) and wind properties closer to observations. We ran CR magneto-hydrodynamical simulations of dwarf galaxies at high resolution (9 pc) with the adaptive mesh refinement code ramses. Disc galaxies are embedded in isolated halos of mass of $10^{10}$ and $10^{11} \, \rm M_{\odot}$, and CRs are injected by SNe. We included CR isotropic and anisotropic diffusion with various diffusion coefficients, CR radiative losses, and CR streaming. The injection of CR energy into the interstellar medium smooths out the highest gas densities, which reduces the SFR by a factor of 2-3. Mass outflow rates are significantly greater with CR diffusion, by 2 orders of magnitudes for the higher diffusion coefficients. Without diffusion and streaming, CRs are inefficient at generating winds. CR streaming alone allows for the formation of winds but which are too weak to match observations. The formation of galactic winds strongly depends on the diffusion coefficient: for low coefficients, CR energy stays confined in high density regions where CR energy losses are highest, and higher coefficients, which allow for a more efficient leaking of CRs out of dense gas, produce stronger winds. CR diffusion leads to colder and denser winds than without CRs, and brings outflow rates and mass loading factors much closer to observations.
△ Less
Submitted 24 April, 2020; v1 submitted 22 March, 2020;
originally announced March 2020.
-
AGN-driven quenching of satellite galaxies
Authors:
Gohar Dashyan,
Ena Choi,
Rachel S. Somerville,
Thorsten Naab,
Amanda C. N. Quirk,
Michaela Hirschmann,
Jeremiah P. Ostriker
Abstract:
We explore the effect of active galactic nucleus (AGN) feedback from central galaxies on their satellites by comparing two sets of cosmological zoom-in runs of 27 halos with masses ranging from $10^{12}$ to $10^{13.4}$ solar masses at z=0, with (wAGN) and without (noAGN) AGN feedback. Both simulations include stellar feedback from multiple processes, including powerful winds from supernovae, stell…
▽ More
We explore the effect of active galactic nucleus (AGN) feedback from central galaxies on their satellites by comparing two sets of cosmological zoom-in runs of 27 halos with masses ranging from $10^{12}$ to $10^{13.4}$ solar masses at z=0, with (wAGN) and without (noAGN) AGN feedback. Both simulations include stellar feedback from multiple processes, including powerful winds from supernovae, stellar winds from young massive stars, AGB stars, radiative heating within Strömgren spheres and photoelectric heating. Our wAGN model is identical to the noAGN model except that it also includes a model for black hole seeding and accretion, as well as AGN feedback via high-velocity broad absorption line winds and Compton/photoionization heating. We show that the inclusion of AGN feedback from the central galaxy significantly affects the star formation history and the gas content of the satellite galaxies. AGN feedback starts to affect the gas content and the star formation of the satellites as early as z=2. The mean gas rich fraction of satellites at z=0 decreases from 15% in the noAGN simulation to 5% in the wAGN simulation. The difference between the two sets extends as far out as five times the virial radius of the central galaxy at z=1. We investigate the quenching mechanism by studying the physical conditions in the surroundings of pairs of satellites matched across the wAGN and noAGN simulations and find an increase in the temperature and relative velocity of the intergalactic gas.
△ Less
Submitted 18 June, 2019;
originally announced June 2019.
-
AGN feedback in dwarf galaxies?
Authors:
Gohar Dashyan,
Joseph Silk,
Gary A. Mamon,
Yohan Dubois,
Tilman Hartwig
Abstract:
Dwarf galaxy anomalies, such as their abundance and cusp-core problems, remain a prime challenge in our understanding of galaxy formation. The inclusion of baryonic physics could potentially solve these issues, but the efficiency of stellar feedback is still controversial. We analytically explore the possibility of feedback from Active Galactic Nuclei (AGN) in dwarf galaxies and compare AGN and su…
▽ More
Dwarf galaxy anomalies, such as their abundance and cusp-core problems, remain a prime challenge in our understanding of galaxy formation. The inclusion of baryonic physics could potentially solve these issues, but the efficiency of stellar feedback is still controversial. We analytically explore the possibility of feedback from Active Galactic Nuclei (AGN) in dwarf galaxies and compare AGN and supernova (SN) feedback. We assume the presence of an intermediate mass black hole within low mass galaxies and standard scaling relations between the relevant physical quantities. We model the propagation and properties of the outflow and explore the critical condition for global gas ejection. Performing the same calculation for SNe, we compare the ability of AGN and SNe to drive gas out of galaxies. We find that a critical halo mass exists below which AGN feedback can remove gas from the host halo and that the critical halo mass for AGN is greater than the equivalent for SNe in a significant part of the parameter space, suggesting that AGN could provide an alternative and more successful source of negative feedback than SNe, even in the most massive dwarf galaxies.
△ Less
Submitted 16 October, 2017;
originally announced October 2017.
-
Active Galactic Nuclei outflows in galaxy discs
Authors:
Tilman Hartwig,
Marta Volonteri,
Gohar Dashyan
Abstract:
Galactic outflows, driven by active galactic nuclei (AGN), play a crucial role in galaxy formation and in the self-regulated growth of supermassive black holes (BHs). AGN feedback couples to and affects gas, rather than stars, and in many, if not most, gas-rich galaxies cold gas is rotationally supported and settles in a disc. We present a 2D analytical model for AGN-driven outflows in a gaseous d…
▽ More
Galactic outflows, driven by active galactic nuclei (AGN), play a crucial role in galaxy formation and in the self-regulated growth of supermassive black holes (BHs). AGN feedback couples to and affects gas, rather than stars, and in many, if not most, gas-rich galaxies cold gas is rotationally supported and settles in a disc. We present a 2D analytical model for AGN-driven outflows in a gaseous disc and demonstrate the main improvements, compared to existing 1D solutions. We find significant differences for the outflow dynamics and wind efficiency. The outflow is energy-driven due to inefficient cooling up to a certain AGN luminosity ($\sim 10^{43}$erg/s in our fiducial model), above which the outflow remains momentum-driven in the disc up to galactic scales. We reproduce results of 3D simulations that gas is preferentially ejected perpendicular to the disc and find that the fraction of ejected interstellar medium is lower than in 1D models. The recovery time of gas in the disc, defined as the freefall time from the radius to which the AGN pushes the ISM at most, is remarkably short, of the order 1Myr. This indicates that AGN-driven winds cannot suppress BH growth for long. Without the inclusion of supernova feedback, we find a scaling of the black hole mass with the halo velocity dispersion of $M_\mathrm{BH} \propto σ^{4.8}$.
△ Less
Submitted 9 April, 2018; v1 submitted 12 July, 2017;
originally announced July 2017.
