Exoplanet detection in angular differential imaging: combining a statistics-based learning with a deep-based learning for improved detections

Olivier Flasseur, Théo Bodrito, Julien Mairal, Jean Ponce, M. Langlois, Anne-Marie LagrangeAdaptive 2022 · DOI

TL;DR

This paper introduces a novel exoplanet detection algorithm that combines the statistical modeling of the PACO method with the power of deep learning. By using PACO for preprocessing and a U-Net architecture for detection, the method significantly improves sensitivity and reduces false alarms compared to existing techniques, achieving up to a half-magnitude contrast improvement.

What Problem Does This Paper Address?

Detecting exoplanets through direct imaging is incredibly challenging due to the immense contrast difference between a star and its orbiting planets (often >10^5 in the infrared). This high contrast, coupled with residual starlight and instrumental artifacts, creates "speckles" that mimic planetary signals. While techniques like Angular Differential Imaging (ADI) and post-processing algorithms like PACO, cADI, and PCA help, they have limitations. PACO, while effective at modeling spatial correlations in speckles, relies on a statistical model that is an approximation of reality, leading to room for improvement in detection sensitivity. This paper aims to bridge this gap by enhancing detection sensitivity and reliability.

Methods

The proposed algorithm is a three-step process:

  1. Preprocessing with PACO: The raw ADI data is first preprocessed using the PACO framework. This involves centering and locally whitening the data to improve stationarity and contrast. PACO achieves this by modeling the nuisance component (speckles) with a locally learned, data-driven statistical model (a multivariate Gaussian) and applying a shrinkage estimator for regularization.
  2. Deep Learning Detection: A convolutional neural network (CNN), specifically a U-Net with a ResNet18 encoder, is trained in a supervised manner. The network learns to identify the signature of synthetic exoplanet sources within the preprocessed science data.
  3. Detection Map Generation: The trained CNN is then applied to the preprocessed observations to generate a detection map, where each pixel indicates the probability of a source being present.

A key innovation is the custom data augmentation strategy for training the CNN. To overcome the scarcity of real ground-truth data and the variability of observing conditions, the authors generate synthetic training samples by:

  • Randomly permuting the temporal sequence of images in an observation.
  • Injecting synthetic point-like sources with contrasts drawn from a distribution that includes sources detectable by PACO and those below its detection limit.
  • Updating the local statistics and re-preprocessing patches affected by injected sources.
  • Derotating the images to co-align synthetic source signals along the temporal axis, making them suitable for the CNN's receptive field.

The loss function used for training is a Dice2 loss, which is well-suited for highly unbalanced classes (many background pixels vs. few source pixels) and aligns with the goal of maximizing detection accuracy.

Datasets

The study utilizes 11 datasets from the VLT/SPHERE-IRDIS instrument, acquired in K1 or H2 spectral bands under various observing conditions. These datasets were pre-reduced by the SPHERE instrument's data reduction pipeline.

For training, a synthetic training strategy is employed, generating a large set of synthetic sources injected into real observation data. This approach differs from relying solely on real data due to the limited number of confirmed exoplanets and the difficulty in defining true negative samples.

Key Insights and Ideas

  • Synergistic Power of Statistics and Deep Learning: The paper demonstrates that combining a statistically-grounded preprocessing step (PACO) with a data-driven deep learning model (U-Net) can outperform either approach alone. PACO's statistical model provides a robust foundation, while deep learning handles the residual complexities and potential inaccuracies of that model.
  • Data Augmentation for Robustness: The custom data augmentation strategy is crucial for training a reliable deep learning model in this domain. By simulating realistic scenarios with varying source properties and temporal permutations, the model becomes more robust to real-world observational challenges.
  • Addressing Class Imbalance with Dice Loss: The choice of Dice2 loss is a smart move for this problem, as it directly tackles the severe imbalance between background and potential exoplanet pixels, leading to more effective training.
  • PACO's Statistical Model as a Feature Extractor: The PACO framework, beyond its direct detection capabilities, serves as an effective feature extractor by improving data stationarity and contrast, making the subsequent deep learning task more tractable.

Key Results

  • The proposed method performs on-par with or better than established algorithms like cADI, PCA, and PACO.
  • It achieves a contrast improvement of up to half a magnitude compared to PACO.
  • Qualitative results show the proposed method detecting four out of six known real sources without false alarms on a typical dataset.
  • Quantitative results, particularly ROC curves, demonstrate a significant improvement in the precision-recall trade-off across various angular separation ranges, with AUC values for the proposed method being consistently higher (e.g., AUC of 0.95 for real sources in the [0; 9]" range, compared to PACO's 0.88).
  • The method shows fewer false alarms than expected for standard algorithms, particularly compared to cADI and PCA.

Limitations and Future Work

  • Control of Uncertainties: The authors acknowledge that the current detection scores are "pseudo-probabilities." A critical future step is to develop methods to control these uncertainties, allowing the scores to be interpreted as statistically-grounded probabilities.
  • Extension to Multi-Spectral Data: The team plans to extend the model to handle multi-spectral data at low resolution, such as data from the VLT/SPHERE-IFS instrument.
  • Hyper-parameter Tuning: While the paper focuses on a specific U-Net architecture and training strategy, further exploration of hyper-parameter tuning for both the PACO preprocessing and the deep learning model could yield additional improvements.

Summary updated

Abstract

Direct imaging is an active research topic in astronomy for the detection and the characterization of young substellar objects. The very high contrast between the host star and its companions makes detection particularly challenging. In addition to the use of an extreme adaptive optics system and a coronagraph to strongly attenuate the starlight contamination, dedicated post-processing methods combining several images recorded with the pupil tracking mode of the telescope are needed. In previous works, we have presented the PACO algorithm capturing the spatial correlations of the data with a multi-variate Gaussian model whose parameters are estimated in a data-driven fashion at the scale of a patch of a few tens of pixels. PACO is parameter free and delivers reliable detection confidences with an improved sensitivity compared to the standard methods of the field (e.g., cADI, PCA, TLOCI ). However, there is a room for improvement in the detection sensitivity due to the approximate fidelity of the PACO statistical model with respect to the observations. We propose to combine the statistics-based model of PACO with a deep learning approach in a three-step algorithm. First, the data are centered and whitened locally using the PACO framework to improve the stationarity and the contrast in a preprocessing step. Second, a convolutional neural network is trained in a supervised fashion to detect the signature of synthetic sources in the preprocessed science data. The network is trained from scratch with a custom data augmentation strategy allowing to generate a large training set from a single spatio-temporal dataset. Finally, the trained network is applied to the preprocessed observations and delivers a detection map. We apply our method on eleven datasets from the VLT/SPHERE-IRDIS instrument and compare our method with PACO and other baselines of the field (cADI, PCA). Our results show that the proposed method performs on-par with or better than these algorithms, with a contrast improvement up to half a magnitude with respect to PACO.
BibTeX
@inproceedings{flasseur2022exoplanet,
  author = {Olivier Flasseur and Théo Bodrito and Julien Mairal and Jean Ponce and M. Langlois and Anne-Marie Lagrange},
  title = {{Exoplanet detection in angular differential imaging: combining a statistics-based learning with a deep-based learning for improved detections}},
  booktitle = {Adaptive},
  year = {2022},
  doi = {10.1117/12.2629849},
}
Figure