Photometric Redshifts in JWST Deep Fields: A Pixel-Based Alternative with DeepDISC

TL;DR

The paper demonstrates that a pixel-based deep learning framework, DeepDISC, can yield reliable photometric redshifts from JWST/NIRCam images, approaching template-fitting performance and even surpassing it when filter sets are aligned. By integrating detection, deblending, and redshift estimation within a single model and using a Gaussian Mixture Model to produce redshift PDFs, the method bypasses traditional aperture photometry and leverages morphology and color information directly from images. The authors extensively compare backbone architectures and pretraining strategies, finding that in-domain galaxy-pretrained ResNet50 often provides the best performance, while transformer backbones require larger training samples. They validate on real JADES data and simulations (JAGUAR), discuss domain gaps and scaling, and provide a public photo-z catalog and open-source DeepDISC code to enable broader adoption in JWST surveys and future missions.

Abstract

Photo-z algorithms that utilize SED template fitting have matured, and are widely adopted for use on high-redshift near-infrared data that provides a unique window into the early universe. Alternative photo-z methods have been developed, largely within the context of low-redshift optical surveys. Machine learning based approaches have gained footing in this regime, including those that utilize raw pixel information instead of aperture photometry. However, the efficacy of image-based algorithms on high-redshift, near-infrared data remains underexplored. Here, we test the performance of Detection, Instance Segmentation and Classification with Deep Learning (DeepDISC) on photometric redshift estimation with NIRCam images from the JWST Advanced Deep Extragalactic Survey (JADES) program. DeepDISC is designed to produce probabilistic photometric redshift estimates directly from images, after detecting and deblending sources in a scene. Using NIRCam-only images and a compiled catalog of spectroscopic redshifts, we show that DeepDISC produces reliable photo-zs and uncertainties comparable to those estimated from template fitting using HST+JWST filters; DeepDISC even outperforms template fitting (lower scatter/fewer outliers) when the input photometric filters are matched. Compared with template fitting, DeepDISC does not require measured photometry from images, and can produce a catalog of 94000 photo-zs in ~4 minutes on a single NVIDIA A40 GPU. While current spectroscopic training samples are small and incomplete in color-magnitude space, this work demonstrates the potential of DeepDISC for increasingly larger image volumes and spectroscopic samples from ongoing and future programs. We discuss the impact of the training data on applications to broader samples and produce a catalog of photo-zs for all JADES DR2 photometric sources in the GOOD-S field, with quality flags indicating caveats.

Figures (16)

• Figure 1: JADES photometric and spectroscopic catalog footprint. The left panel shows all sources in the photometric catalog (grey) and sources where imaging is present in the 9 NIRCam filters from DR1 (blue). Red points indicate our spectroscopic sample. The right panel shows how the imaging data is partitioned into sub-images that can contain multiple sources with spec-zs. We exclude from our training set any sub-image that contains no sources with spec-zs.
• Figure 2: N(z) for objects in the training and test sets. Few sources exist at $z>6$ in either sample.
• Figure 3: Loss curves showing the training and validation set loss for for two models pretrained on ImageNet data over 150 epochs. The MViTv2 network (left) consistently learns over the course of the training, as the training and validation loss decrease over time. The ResNet50 does not learn as much with regards to object redshift, evidenced by the higher validation set loss. The losses have been smoothed with an exponential moving average. The upper panels show the total loss, defined in Equation \ref{['eq:loss']} and the lower panels show only the redshift component of the loss.
• Figure 4: Loss curves showing the training and validation set loss for for two models pretrained on Galaxies ML data over 150 epochs. The MViTv2 network (left) curves appear qualitatively similar to those in Figure \ref{['fig:IMpre_loss_curves']}, indicated pretraining does not have a large effect. The ResNet50 network (right) achieves lower validation set redshift loss compared to the previous result. Loss curves are smoothed and arranged as in Figure \ref{['fig:IMpre_loss_curves']}.
• Figure 5: Photo-z point estimate scatter plot comparing an MViT and ResNet50 (R50) model. Both models are pretrained using the GalaxiesML dataset of images and redshifts from HSC DR2. The mode of each PDF is used as a point estimate $z_{\rm phot}$ and error bars are taken from the 68% confidence intervals. The number of detected objects is indicated in the bottom right. The R50 model outperforms the MViTv2 model, in terms of bias, scatter $\sigma_{IQR}$ and outlier fraction $\eta$ as well as object detection.