Contrastive Learning of Extragalactic Stellar Streams: Sculpting a Latent Space of Representations with DES DR2 Photometry

We present a self-supervised approach for characterizing low surface brightness tidal features in wide-field imaging data by applying the nearest-neighbor contrastive learning of visual representations (NNCLR) algorithm to a curated subset of the Dark Energy Survey Data Release 2 (DES DR2). We construct 38,334 cutouts of well-resolved galaxies in the g, r, i bands, applying a novel “tiered sigmoid scaling function” to dynamically adjust image contrast according to the object’s signal-to-noise and background level. A supplemental labeled sample of 366 galaxies enables qualitative assessment of the learned embeddings. We train a convolutional neural network with image augmentations including injection of simulated background stars, and project the resulting 512-dimensional representations into two dimensions using uniform manifold approximation and projection (UMAP) and its local density preserving variant (densMAP). We find that the NNCLR latent space recovers global trends corresponding to major merger features, yet does not reliably separate stellar streams without further supervision. To interpret the network’s implicit attention, we compute gradient-based saliency maps averaged over the full dataset: these reveal that the tiered sigmoid scaling effectively attenuates information from the center of the image cutouts, thereby suppressing the learning of high surface brightness features of each image cutout’s central galaxy. Our study provides a blueprint for leveraging contrastive methods to mine forthcoming survey data for faint tidal substructure, and highlights key preprocessing and interpretability considerations for robust stream detection.

💡 Research Summary

**
This paper presents a self‑supervised framework for detecting and characterizing low surface‑brightness tidal features, especially extragalactic stellar streams, in wide‑field imaging from the Dark Energy Survey Data Release 2 (DES DR2). The authors assemble a curated dataset of 38 334 well‑resolved galaxy cutouts in the g, r, and i bands and a supplemental labeled set of 366 objects (47 positives containing known streams). To enhance faint structures, they introduce a novel “tiered sigmoid scaling” function that dynamically adjusts image contrast based on each object’s signal‑to‑noise ratio and background level. The scaling is applied per magnitude tier (four bins per band) with hyper‑parameters µ and σ tuned using photometric statistics from known streams; a 30 % random tier selection is also used to prevent the network from learning tier‑specific cues.

The core representation learner is the Nearest‑Neighbor Contrastive Learning of Visual Representations (NNCLR) algorithm, implemented with a ResNet‑18 encoder and a 512‑dimensional latent space. For each input image, two augmented views are generated using a pipeline that includes random rotations, color jitter, injection of simulated background stars, and the tiered sigmoid scaling. The two views are encoded, ℓ₂‑normalized, and compared to the nearest neighbor from a queue of embeddings via a cosine‑similarity loss, encouraging instance discrimination while leveraging the nearest‑neighbor operation to improve generalization.

After training, the high‑dimensional embeddings are projected to two dimensions using Uniform Manifold Approximation and Projection (UMAP) and its density‑preserving variant (densMAP). The resulting plots reveal clear global trends: galaxies undergoing major mergers (tidal tails, shells) form distinct clusters, while normal galaxies occupy another region. However, the 47 stream‑positive examples are dispersed throughout the space and are not isolated as a separate cluster, indicating that NNCLR alone cannot reliably separate stellar streams without additional supervision.

To interpret what the network focuses on, the authors compute gradient‑based saliency maps averaged over the entire dataset. These maps show that the tiered sigmoid scaling effectively suppresses information from the central, high‑surface‑brightness part of each cutout, causing the encoder to attend primarily to the low‑brightness outskirts. While this behavior aligns with the goal of emphasizing faint tidal features, it also unintentionally reduces the network’s ability to learn discriminative cues that differentiate streams from other low‑brightness structures.

The study highlights several key insights and future directions. First, preprocessing choices—particularly the contrast‑scaling function—have a profound impact on what features are learned. A more balanced approach, such as multi‑scale inputs or adaptive scaling that preserves some central information, may improve stream detection. Second, contrastive pre‑training is valuable for extracting meaningful representations from massive unlabeled surveys, but downstream fine‑tuning with even a modest number of labeled streams (or prototype‑based methods) appears necessary for robust classification. Third, the current image resolution (256 × 256 px) may limit the detectability of very thin streams; higher resolution cutouts and more sophisticated background modeling could yield better performance.

In summary, the paper provides a practical blueprint for applying contrastive self‑supervised learning to forthcoming large‑scale surveys (e.g., LSST, Euclid) to mine faint tidal substructures. It demonstrates that while NNCLR can capture broad merger‑driven morphology, additional methodological refinements—particularly in image scaling and supervised fine‑tuning—are essential to achieve reliable stellar‑stream detection.