Super-resolving Herschel - a deep learning based deconvolution and denoising technique

TL;DR

This work tackles the limited angular resolution of FIR/sub-mm surveys by introducing a transformer-based super-resolution/deconvolution method applied to Herschel SPIRE 500 μm data. Trained solely on realistic SIDES and SHARK simulations, the Swin-Unet-inspired network ingests multi-band priors (MIPS 24 μm, SPIRE 250/350/500 μm) and outputs a high-resolution 500 μm map at 7.9″ FWHM, achieving ≈1 s per deg^2 inference on consumer GPUs. It delivers flux accuracy of about 5% for sources ≳8 mJy, sub-arcsecond astrometry (≲1″), and reliability/completeness >90% for modest fluxes, outperforming blind extractions and aligning well with SCUBA-2 450 μm measurements in the COSMOS field after a simple flux conversion. The method enables SR over hundreds of deg^2 without fine-tuning, yielding robust multi-wavelength statistics for DSFGs and paving the way for extending SR to other SPIRE bands and upcoming surveys.

Abstract

Dusty star-forming galaxies (DSFGs) dominate the far-infrared and sub-millimetre number counts, but single-dish surveys suffer from poor angular resolution, complicating mult-wavelength counterpart identification. Prior-driven deblending techniques require extensive fine-tuning and struggle to process large fields. This work aims to develop a fast, reliable deep-learning based deconvolution and denoising super-resolution (SR) technique. We employ a transformer neural network to improve the resolution of Herschel/SPIRE 500 $μ$m observations by a factor 4.5, using Spitzer/MIPS 24$μ$m and Herschel/SPIRE 250, 350, 500$μ$m images. Trained on SIDES and SHARK simulations, we injected instrumental noise into the input simulated images, while keeping the target images noise-free to enhance de-noising capabilities of our method. We evaluated the performance on simulated test sets and real JCMT/SCUBA-2 450 $μ$m observations in the COSMOS field which have superior resolution compared to Herschel. Our SR method achieves an inference time of $1s/deg^2$ on consumer GPUs, much faster than traditional deblending techniques. Using the simulation test sets, we show that fluxes of the extracted sources from the super-resolved image are accurate to within 5% for sources with an intrinsic flux $\gtrsim$ 8 mJy, which is a substantial improvement compared to blind extraction on the native images. Astrometric error is low ($\lesssim$ 1" vs 12" pixel scale). Reliability is $\gtrsim$ 90% for sources $>$3 mJy and $>$90% of sources with intrinsic fluxes $\gtrsim5$ mJy are recovered. Applied to real 500 $μ$m observations, fluxes of the extracted sources from the super-resolved map agree well with SCUBA-2 measured fluxes for sources $\geq$10 mJy. Our technique enables SR over hundreds of $deg^2$ without the need for fine-tuning, facilitating statistical analysis of DSFGs.

Figures (16)

• Figure 1: Example image cutouts (256$^{\prime\prime}$$\times$ 256$^{\prime\prime}$) at MIPS 24 $\mu$m, SPIRE 250, 350 and 500 $\mu$m, and SCUBA-2 450 $\mu$m. The top two rows show simulated cutouts from SIDES and SHARK. The bottom row shows cutouts of real observations of COSMOS. The first four columns represents the input to the network. The final column represents either the super-resolved target image at 500 $\mu$m in the case of simulations, or the SCUBA-2 450 $\mu$m image in the case of real observations.
• Figure 2: Contours of the MIPS, SPIRE and SCUBA-2 maps projected onto the SPIRE 500$\mu$m map. Our super-resolved 500$\mu$m map falls within the MIPS coverage.
• Figure 3: Architecture of the Swin-Unet network adaptation used in this paper to super-resolve the SPIRE 500 $\mu$m images. The input consists of 24, 250, 350 and 500 $\mu$m images of the same area. The output is the super-resolved 500 $\mu$m image. The network is structured as an auto-encoder, where information is compressed/decompressed by downsampling/upsampling the spatial dimensions of the features by a factor of 2, using the Patch Merging and Patch Expanding layers. The encoder and decoder is connected by skip connections (concatenations) at various dimensional levels, with a bottleneck layer at the bottom. The image is split in patches in the Patch Partition layer and embedded together with its relative position in the Linear Embedding layer. In each encoder/decoder level, the Swin Transformer Block is applied three times in consecutive order. This Swin Transformer Block is also depicted on the right, comprised of the windowed multi-head self-attention (W-MSA) part and a part where the window is shifted (SW-MSA). Finally, in our adaptation an additional $1\times1$ convolutional layer at the output is added such that we obtain a super-resolved image.
• Figure 4: SR performance on the simulated images (Top: The native 500 $\mu$m images; Middle: The super-resolved images; Bottom: The target images at 7.9$^{\prime\prime}$ resolution, representing the ground truth). The blue boxes highlight regions that contain at least one target source. Green circles (blue crosses) indicate sources extracted from the super-resolved (target) map.
• Figure 5: Astrometric error vs. photometric error for super-resolved sources matched with the corresponding input sources using a maximum matching radius of 4$^{\prime\prime}$. The marginal PDF and CDF are estimated using the counts within the selected region.