AutoSourceID: Modular Astrophysical Source Detection

• AutoSourceID is a modular deep-learning framework that integrates learned segmentation with classic post‐processing for efficient astrophysical source detection across multiple wavelengths.
• It employs a U-Net for soft segmentation followed by a Laplacian-of-Gaussian filter to accurately localize sources and extract centroids.
• Extensions like ASID-C and ASID-FE enable star-galaxy classification and flux/centroid regression with uncertainty estimation, demonstrating versatility in optical and gamma-ray applications.

AutoSourceID (ASID) is a modular deep-learning framework for astrophysical source detection that combines learned segmentation or representation modules with lightweight classical post-processing and downstream characterization networks. In its original gamma-ray formulation, ASID used a U-shaped convolutional network for source segmentation, alternative clustering stages for localization, and a downstream classifier for source typing (Panes et al., 2021). In optical imaging, the best-known instantiation is AutoSourceID-Light (ASID-L), which localizes point sources in MeerLICHT images by chaining a U-Net with a single-scale Laplacian-of-Gaussian (LoG) filter (Stoppa et al., 2022). The framework was subsequently extended with AutoSourceID-Classifier (ASID-C) for star-galaxy classification and AutoSourceID-FeatureExtractor (ASID-FE) for flux and centroid regression with uncertainty estimation (Stoppa et al., 2023, Stoppa et al., 2023). A later end-to-end gamma-ray pipeline generalized the same design pattern to multi-energy count maps from Fermi-LAT and CTAO, combining segmentation, LoG candidate extraction, classification, and ensemble regressors (Pérez-Romero et al., 29 Sep 2025).

1. Origins and conceptual scope

The ASID line began with gamma-ray point-source identification in Fermi-LAT data. That system used a five-level U-shaped fully convolutional network on $64\times 64\times 5$ energy-binned patches, followed by either $k$-means post-processing or Centroid-Net for localization, and then a three-class classifier that separated AGNs, PSRs, and FAKE sources (Panes et al., 2021). In that setting, ASID was positioned as an alternative to traditional maximum-likelihood pipelines, with comparable flux thresholds and improved robustness to changes in the interstellar emission model.

ASID-L transferred the same hybrid philosophy to optical survey images. Its key design choice was to replace iterative or scale-search localization with a two-stage procedure in which a U-Net produces a soft mask of likely point sources and a fixed-scale LoG filter estimates source centers (Stoppa et al., 2022). The method was motivated by the increasing survey speed of optical wide-field telescopes and the need for rapid and reliable transient-oriented source localization.

Later work made the broader scope of AutoSourceID explicit. ASID-C attached a star-galaxy classifier to ASID-L cutouts (Stoppa et al., 2023); ASID-FE added a regression module for flux and sub-pixel position with calibrated uncertainties (Stoppa et al., 2023); and the 2025 gamma-ray pipeline described ASID as a versatile framework already tested on Fermi-LAT simulated data, MeerLICHT optical data, CTAO simulated data, and transfer experiments on HST and WISE imagery (Pérez-Romero et al., 29 Sep 2025). This suggests that “ASID” is best understood not as a single network, but as a family of pipelines built around learned source proposal generation plus task-specific post-processing.

2. Optical localization pipeline in ASID-L

ASID-L converts a raw optical image into a catalog of source coordinates in two stages. First, the full-frame CCD image of $10\,500\times 10\,500$ pixels is split into overlapping or non-overlapping $256\times 256$-pixel patches. Each patch is passed through a U-Net that outputs, for every pixel, a value in $[0,1]$ indicating the likelihood of belonging to a point source. The predicted mask is not a discrete label map; it is a soft, circular “blob” for each source (Stoppa et al., 2022).

Second, the continuous mask is thresholded at $\tau=0.20$ to suppress background noise, and a single-scale LoG filter with $\sigma=1.43$ is convolved with the mask. Local maxima in the LoG response are retained as source centers, subject to a maximum overlap rule $\omega=0.8$. Pixel coordinates from all patches are then concatenated and transformed back to sky coordinates through the standard astrometric solution (Stoppa et al., 2022).

The LoG kernel used for localization is

$$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$

The response map is computed as

$R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$

where $k$0 is the U-Net soft mask. Peaks in $k$1 correspond to blob centers of radius $k$2, and fixing $k$3 avoids expensive multi-scale loops because the LoG is run only once per patch (Stoppa et al., 2022).

Within this division of labor, the U-Net is responsible for background estimation, artifact rejection, and simple deblending, while the LoG acts as a lightweight centroiding stage. The artifacts explicitly named in the description are cosmic rays, diffraction spikes, and satellite trails (Stoppa et al., 2022).

3. Network architecture and objective functions

The ASID-L segmentation network follows the canonical U-Net design, adapted to small-object segmentation. Its encoder has five convolutional levels down to $k$4 pixels. At each level, two $k$5 convolutions with padding $k$6 and ReLU activations are followed by $k$7 max-pooling with stride 2. The number of channels doubles after each pooling step, for example $k$8. The bottleneck consists of two $k$9 convolution-plus-ReLU layers at $10\,500\times 10\,500$0 resolution with 1024 channels. The decoder mirrors the encoder: each of its five levels uses a $10\,500\times 10\,500$1 up-convolution to halve the channel count and double spatial size, concatenates the corresponding encoder feature map through a skip connection, and applies two $10\,500\times 10\,500$2 convolution-plus-ReLU layers. A final $10\,500\times 10\,500$3 convolution to one channel followed by a Sigmoid activation produces the $10\,500\times 10\,500$4 mask. The model has approximately $10\,500\times 10\,500$5 million trainable parameters (Stoppa et al., 2022).

Training uses a “Combo” loss equal to the sum of binary cross-entropy and Dice loss:

$10\,500\times 10\,500$6

$10\,500\times 10\,500$7

$10\,500\times 10\,500$8

The stated purpose of this combination is to balance pixel-wise accuracy with overlap-based segmentation quality in a highly imbalanced setting with few source pixels and many background pixels (Stoppa et al., 2022).

The broader ASID family preserves this modular logic but changes the output heads and losses according to the task. In the 2025 gamma-ray formulation, the segmentation module is a Multi-Input U-Net operating on energy-binned count maps, the classifier is a VGG-like CNN that outputs $10\,500\times 10\,500$9, and two deep-ensemble regressors estimate flux $256\times 256$0 and a residual $256\times 256$1 for location refinement (Pérez-Romero et al., 29 Sep 2025). That formulation uses binary cross-entropy for segmentation and classification, while flux and localization can use either mean-squared-error or a negative-log-Gaussian-likelihood (Pérez-Romero et al., 29 Sep 2025).

4. Training data, supervision, and empirical performance on MeerLICHT

ASID-L was trained on three fields from the MeerLICHT $256\times 256$2-band: the Omega Cen globular cluster, the Fornax galaxy cluster, and the Chandra Deep Field-South. These yielded $256\times 256$3 patches of size $256\times 256$4 pixels. Gaia EDR3 provided the reference catalog used to place “true” point-source masks; the mask geometry was a $256\times 256$5 central square plus four cardinal pixels approximating a small circle. Only Gaia sources with predicted $256\times 256$6, estimated from local zero point, PSF, and sky plus read noise, were included in the masks (Stoppa et al., 2022).

The split was $256\times 256$7 train, $256\times 256$8 validation, and $256\times 256$9 test, corresponding to approximately 4034, 504, and 504 patches. In addition, the 165 most crowded Omega Cen patches formed an independent test set for detailed evaluation (Stoppa et al., 2022). The paper does not explicitly describe standard augmentations such as rotations, flips, or photometric jitter. It also states that optimizer and learning-rate schedule are not spelled out, although convergence was assessed through stabilization of the combined loss on the validation set and training stopped when no improvement was seen after a predefined patience (Stoppa et al., 2022).

Performance was quantified through true positives, false negatives, false positives, and the Dice coefficient

$[0,1]$0

On the Omega Cen test set, Gaia EDR3 ground truth corresponded to $[0,1]$1 sources arcmin$[0,1]$2. SExtractor, with default settings except $[0,1]$3 and $[0,1]$4, achieved $[0,1]$5, $[0,1]$6, $[0,1]$7, and $[0,1]$8. ASID-L, with $[0,1]$9, $\tau=0.20$0, and $\tau=0.20$1, achieved $\tau=0.20$2, $\tau=0.20$3, $\tau=0.20$4, and $\tau=0.20$5 (Stoppa et al., 2022).

The density dependence is central to the method’s positioning. ASID-L maintained $\tau=0.20$6–$\tau=0.20$7 across $\tau=0.20$8–$\tau=0.20$9 sources arcmin$\sigma=1.43$0, whereas SExtractor’s Dice dropped sharply above approximately $\sigma=1.43$1 sources arcmin$\sigma=1.43$2 (Stoppa et al., 2022). Computation time was measured on $\sigma=1.43$3 pixel images, approximately $\sigma=1.43$4 arcmin$\sigma=1.43$5, on an Intel i9 plus RTX 2080 system. For source densities of $\sigma=1.43$6, $\sigma=1.43$7, and $\sigma=1.43$8 sources arcmin$\sigma=1.43$9, SExtractor required $\omega=0.8$0 s, $\omega=0.8$1 s, and $\omega=0.8$2 s, while ASID-L required $\omega=0.8$3 s, $\omega=0.8$4 s, and $\omega=0.8$5 s, respectively (Stoppa et al., 2022). The method therefore had essentially constant runtime with source density and was reported as giving a factor $\omega=0.8$6 speed-up in sparse fields and $\omega=0.8$7 in crowded fields.

5. Extension to classification and feature extraction

ASID-C extends the localization stage into a two-stage optical pipeline in which ASID-L first detects and centers sources and then a dedicated classifier labels the resulting cutouts as stars or galaxies (Stoppa et al., 2023). ASID-L processes full MeerLICHT images from the $\omega=0.8$8 cm telescope and outputs $\omega=0.8$9 pixel cutouts with source centers. These cutouts are matched to DECaLS DR10 morphological labels and passed to a dual-branch CNN. One branch ingests a single-band $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$0 cutout, randomly shifted by $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$1 pixel from the original $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$2 stamp; the other ingests normalized pixel coordinates $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$3 within the full frame. Their fused representation is mapped to a Sigmoid output that yields a raw score $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$4, and a modified Platt scaling stage then produces calibrated posterior probabilities $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$5 for star versus galaxy (Stoppa et al., 2023).

The ASID-C training set comprises $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$6 million images from $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$7 “green” MeerLICHT exposures across SDSS $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$8 filters, with a class imbalance of approximately $$\mathrm{LoG}(x,y;\sigma) \;=\; -\frac{1}{\pi\sigma^4}\, \Bigl[\,1 - \tfrac{x^2+y^2}{2\sigma^2}\Bigr]\, \exp\!\Bigl(-\frac{x^2+y^2}{2\sigma^2}\Bigr).$$9 stars to galaxies. Splits are $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$0 training, $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$1 validation, $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$2 test, and $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$3 calibration. The network was trained with Adam, initial learning rate $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$4, exponential decay $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$5, and early stopping after $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$6 epochs without validation improvement (Stoppa et al., 2023). Quantitatively, ASID-C maintained $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$7 ROC-AUC down to $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$8, produced sharply bimodal well-calibrated predictions, and processed approximately $R(x,y)\;=\;\bigl(\,m \;*\;\mathrm{LoG}\bigr)(x,y),$9k cutouts per second on a GeForce RTX2080, corresponding to about $k$00s per source (Stoppa et al., 2023).

ASID-FE addresses a different downstream problem: continuous feature regression with uncertainty characterization. It takes $k$01 cutouts around localized sources from ASID-L or other catalogs and estimates flux and sub-pixel center coordinates together with their uncertainties (Stoppa et al., 2023). The method uses a two-step mean-variance estimation (TS-MVE) architecture. Part I predicts $k$02, $k$03, and $k$04 through three convolutional blocks and three independent MLP tails; Part II reuses the mean branch, adds a parallel variance branch, and outputs refined means and $k$05, $k$06, and $k$07, assuming diagonal covariance (Stoppa et al., 2023).

ASID-FE was trained on synthetic cutouts derived from real MeerLICHT images. PSFEx provided position-dependent PSF models; Gaia DR2 sources were projected onto blank full-frame arrays; sky background and its $k$08 were taken from real pipeline products; and approximately $k$09 million $k$10 cutouts were generated and randomly shifted before center-cropping to $k$11 (Stoppa et al., 2023). On synthetic tests, ASID-FE achieved centroid uncertainty calibration $k$12 across $k$13, flux uncertainty calibration near $k$14, and $k$15–$k$16 lower flux MARE than SourceExtractor’s FLUX_AUTO. Relative to a small deep-ensemble MVE without the two-step procedure, TS-MVE Part I already yielded approximately $k$17–$k$18 lower MSE and MAE on the mean predictions, and Part II further lowered the negative log-likelihood by a similar margin (Stoppa et al., 2023).

6. Generalization to gamma-ray surveys, limitations, and future directions

The 2025 end-to-end ASID pipeline generalizes the AutoSourceID pattern from optical images and earlier gamma-ray work to multi-energy count maps from Fermi-LAT and CTAO (Pérez-Romero et al., 29 Sep 2025). Its first stage is a Multi-Input U-Net operating on six energy slices for Fermi-LAT or three for CTAO, each covering a $k$19 sky patch. A LoG clustering step extracts discrete source candidates from the segmentation map, a VGG-like CNN classifies each candidate as a true source or a fake fluctuation, and two deep-ensemble networks then estimate flux and refine the location (Pérez-Romero et al., 29 Sep 2025).

In that formulation, the inputs are $k$20-channel count maps with $k$21 for Fermi-LAT and $k$22 for CTAO, while the outputs for each detected source are sky location $k$23, estimated flux $k$24, and classification score $k$25 (Pérez-Romero et al., 29 Sep 2025). Reported results include a Fermi-LAT flux sensitivity of approximately $k$26 at $k$27 completeness, a $k$28 association rate on real high-latitude 4FGL-DR2 sources with $k$29 and $k$30, approximately $k$31 association for two more difficult source subsets, and CTAO recall reaching $k$32 at $k$33, consistent with standard likelihood analyses in Gammapy (Pérez-Romero et al., 29 Sep 2025).

The same paper reports latent-space clustering in which Fermi-LAT and CTAO source patches form one cluster while background patches form another, and interprets this as evidence that a single “foundation” encoder could serve multiple instruments (Pérez-Romero et al., 29 Sep 2025). This suggests a path toward jointly trained, heterogeneous source-detection models, although that remains prospective rather than established.

The documented limitations of ASID-L remain specific and important. Current U-Net training is limited to point-like masks derived from Gaia, so extended sources such as galaxies and nebulae are not represented. The method has not been exhaustively tested on images with dramatically different PSFs or backgrounds, although preliminary transfer tests on HST and WISE were described as encouraging. ASID-L currently outputs only centroid coordinates; photometry, shape parameters, and classification are delegated to follow-up modules (Stoppa et al., 2022). Proposed extensions include adding a refinement step for flux and PSF metrics, incorporating extended-source segmentation masks, developing a full real-time multi-band ASID that ingests all filters simultaneously, and carrying out more thorough transfer-learning experiments across telescopes and instruments (Stoppa et al., 2022).

Taken together, these developments define AutoSourceID as a layered astronomical analysis framework: learned localization through segmentation, lightweight centroid extraction, and modular downstream inference for classification or regression. The available evidence shows that this design is effective in both optical and gamma-ray point-source detection, especially in crowded or background-dominated regimes where classical heuristics or likelihood scans become slow, brittle, or both (Stoppa et al., 2022).