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

TL;DR

The paper addresses the challenge of detecting extragalactic stellar streams in wide-field imaging, where streams are extremely faint. It introduces a self-supervised NNCLR framework applied to DES DR2 cutouts, augmented with a novel tiered sigmoid scaling to bias learning toward low surface brightness outskirts and complemented by saliency-based interpretability. Latent-space analyses using UMAP and densMAP show clustering by major-merger morphologies and provide a diagnostic link to Galaxy Zoo classifications, though fully separating streams remains difficult without additional supervision. The approach offers a practical blueprint for preprocessing and training in anticipation of LSST/Euclid data, enabling scalable use of unlabeled sky surveys for faint-structure detection with interpretable learning signals.

Abstract

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.

Figures (12)

• Figure 1: Histogram of objects per tiered sigmoid scaling function bin, as validated by the procedure described in Section \ref{['subsec:image_preprocessing']}. The bins represent the DES DR2 subset, with a total of 38,334 unlabeled objects used for training.
• Figure 2: From visual inspection of our dataset, the galaxy LEDA 131334 contains a candidate stellar stream shown above with the (left) arcsinh stretch used by Desmons2024 and (right) tiered sigmoid scaling function.
• Figure 3: A schematic of the NNCLR algorithm as outlined by Dwibedi2021. A batch of images is input to the augmentation pipeline described in Section \ref{['subsubsec:augmentations']}, outputting two augmented images per input image. Each augmented batch is then passed through the NNCLR encoder, where we chose the residual neural network architecture (ResNet-18 Kaiming2016, see Section \ref{['subsec:nnclr']}). The resulting embedding vectors, referred to as "views", are $\ell_2$ normalized (Euclidean norm) and compared to the nearest neighbor of a queue formed by the embedding vectors of other images in the dataset. The NNCLR loss function quantifies the similarity between the views and the nearest neighbor via cosine similarity, and is used by the stochastic gradient descent optimizer to update the model weights accordingly.
• Figure 4: A randomly selected image (top row) prior to the star simulator augmentation and (bottom row) after applying the augmentation. For this example, we include blue apertures for ease of identifying the generated fake stars; the apertures are not present for the images used in training the NNCLR models.
• Figure 5: The saliency map of the model trained on the DES DR2 subset with the (top) tiered sigmoid scaling function and (bottom) the arcsinh stretch of Desmons2024. These maps represent the gradients of the model output with respect to each input pixel. We used the maximum gradient values across color channels as per Simonyan2014 and averaged over the entire DES DR2 subset. The gradients were scaled to be between zero and unity. We observe the model trained with data which has been tiered-sigmoid-scaled being most sensitive to changes within a broader area of the center, such that the model extracts less information from the central galaxy in each cutout. Note that the repeating patterns are artifacts due to the convolutional operations of the NNCLR encoder along with the skip connections of the residual neural network architecture.