Multiple-Source Detection (MSD)

• Multiple-Source Detection (MSD) is a framework for separating and interpreting observations produced by multiple latent sources across diverse domains, including object detection and network diffusion.
• It utilizes varied methodologies—such as domain adaptation, Bayesian signal detection, and subspace unmixing—to extract structured multi-source explanations from partial or noisy data.
• MSD approaches have demonstrated practical performance gains, for example improving AP scores in multi-source domain adaptation and achieving robust source enumeration in sensor networks.

Searching arXiv for recent and foundational uses of “Multiple-Source Detection” across domains. Multiple-Source Detection (MSD) denotes a family of inference problems in which observations produced by more than one latent source, domain, node, target, or subspace must be separated, counted, localized, or fused. The term is therefore not attached to a single canonical task. In current arXiv literature it appears in unsupervised multi-source domain adaptation for 2D and 3D object detection, source inference from diffusion snapshots on graphs, radio source count estimation from array covariances, Bayesian signal detection and localization in sensor networks, distributed multisensor ISAC, astronomical multi-band source finding, and marginal subspace detection for subspace unmixing [(Yao et al., 2021); (Lu et al., 11 May 2026); (Lee et al., 13 May 2026); (Li et al., 11 Jul 2025); (Vijayamohanan et al., 2022); (Nguyen et al., 2015); (0811.0764); (Thomä et al., 17 Nov 2025); (Maddox et al., 2020); (Bajwa et al., 2014)]. A unifying theme is recovery of structured multi-source explanations from partial, noisy, quantized, or distribution-shifted observations.

1. Terminological scope and formal variants

The surveyed literature uses “Multiple-Source Detection” in several non-equivalent but structurally related senses. In each case, the observable data are insufficient for direct source readout, so the problem is posed through a model, surrogate representation, or fusion rule.

Setting	Observation model	Inferred quantity
Multi-source domain adaptation for object detection	Labeled source domains and one unlabeled target domain	Target-domain detections under domain shift
Diffusion on graphs	Single-time infected snapshot on a known network	Source set of the diffusion
Array/radio sensing	Snapshot covariance or received array data	Number of impinging sources
Wireless sensor networks	Quantized sensor reports over imperfect channels	Number, locations, and powers of sources
Distributed ISAC	Multi-link delay/Doppler returns across Tx/Rx pairs	Multiple passive objects and tracks
Astronomical imaging	Multi-band maps with PSFs and confusion noise	Source positions and fluxes
Subspace unmixing	Single observation in an ambient space	Active subspaces among many candidates

A recurring distinction is between source multiplicity and source heterogeneity. In radio array processing and Bayesian collaborative detection, MSD primarily means deciding how many signal sources are present [(Vijayamohanan et al., 2022); (0811.0764)]. In graph diffusion, it means recovering a set of initiating nodes (Li et al., 11 Jul 2025). In object detection, however, “multiple source” refers to multiple source domains rather than multiple physical emitters; the target task is still object detection, but under multi-source domain shift (Yao et al., 2021, Lee et al., 13 May 2026). This suggests that MSD is best understood as a cross-domain label for inverse problems with more than one explanatory origin, rather than as a single standardized benchmark.

2. Multi-source domains in 2D and 3D object detection

In object detection, MSD refers to multi-source domain adaptation: training with labeled data from multiple distinct source domains and adapting to an unlabeled target domain. “Multi-Source Domain Adaptation for Object Detection” formalizes the setting with $M$ labeled source domains $S_1,\dots,S_M$ and one unlabeled target domain $T$, under homogeneity and closed-set assumptions, and proposes the Divide-and-Merge Spindle Network (DMSN), a Faster R-CNN–based architecture with a shared low-level extractor $G_1$, $M$ source-specific high-level subnets, and a pseudo target subnet (Yao et al., 2021). Its objective is

$$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$

where $\mathcal{L}_{det}$ is supervised detection loss, $\mathcal{L}_l$ strong low-level alignment, $\mathcal{L}_h$ weak high-level alignment, and $\mathcal{L}_{con}$ RPN consistency regularization. The design logic is explicit: strongly align low-level features to enhance localization and domain invariance, but weakly align high-level features to preserve per-source discriminative power and avoid negative transfer. Pseudo subnet learning merges source parameters through a similarity-weighted EMA, with $S_1,\dots,S_M$0 and weights $S_1,\dots,S_M$1. On cross-camera adaptation, DMSN reaches 49.2 AP for car, compared with 46.5 for the best single-source DA baseline and 44.1 for $S_1,\dots,S_M$2SDA; on cross-time adaptation it reaches 35.0 mAP, compared with 30.9 for SCL and 27.6 for MDAN (Yao et al., 2021). The paper is also explicit that MSD here concerns multi-source domain distributions, not multi-sensor fusion.

The 3D variant extends the same logic to multi-source, multi-modality domain adaptation. MUSDA assumes multiple labeled source domains, one unlabeled target domain, and two sensing modalities, camera and LiDAR, within a BEVFusion architecture (Lu et al., 11 May 2026). It introduces Hierarchical Spatially-Conditioned (HSC) domain classifiers that align fused BEV features at Level-1 and per-modality BEV features at Level-2, using a class-agnostic heatmap and a multi-modality domain probability map as spatial conditioning signals. After training, MUSDA constructs a class-wise, multi-level prototype graph and performs Prototype Graph Weighted (PGW) fusion. For a prediction from source-conditioned head $S_1,\dots,S_M$3 with score $S_1,\dots,S_M$4 and class $S_1,\dots,S_M$5, the fusion weight is

$S_1,\dots,S_M$6

where $S_1,\dots,S_M$7 is the prototype distance between source $S_1,\dots,S_M$8 and target $S_1,\dots,S_M$9 for class $T$0. On nuScenes+Lyft $T$1 Waymo, HSC-DC + PGW-MF yields Car 68.2/58.2, Pedestrian 63.5/52.7, and Cyclist 28.0/27.0 in mAP L1/L2; gains over HSC-DC are especially pronounced for Pedestrian (+6.2 L1 mAP) and Cyclist (+5.0 L1 mAP) (Lu et al., 11 May 2026).

A later development, MS-DePro, argues that deriving both domain-agnostic and domain-specific signals from RGB alone creates a training conflict, and instead injects domain-agnostic characteristics at the input level through depth and text (Lee et al., 13 May 2026). The method combines depth-guided localization with multi-modal guided prompt learning in a RegionCLIP/Faster R-CNN framework. Learnable prompts are decomposed into domain-agnostic tokens $T$2 and domain-specific tokens $T$3, with default $T$4 and $T$5. The total student loss is

$T$6

with $T$7, while teacher parameters are updated by EMA with $T$8. On cross-time adaptation, MS-DePro reaches 53.7 mAP versus 47.9 for ACIA and 50.8 for Oracle All-combined; on cross-camera adaptation it reaches 68.4 mAP versus 59.1 for ACIA (Lee et al., 13 May 2026). Relative to DMSN and MUSDA, this suggests a shift from adversarial alignment and post hoc source fusion toward modality-level invariance and prompt-conditioned classification.

3. Diffusion snapshots and overlapping communities in networks

In network analysis, MSD is the problem of inferring which nodes initiated a diffusion process from a partial, single-time snapshot of infection. “Addressing overlapping communities in multiple-source detection” assumes a known network $T$9, an infected sub-network $G_1$0, and uninfected boundary nodes $G_1$1 adjacent to infected nodes, forming an extended infected network $G_1$2 with $G_1$3 (Li et al., 11 Jul 2025). The method is model-agnostic with respect to the diffusion mechanism and instead exploits the topology of the infected region and the infected/boundary arrangement.

The central claim is that overlapping communities complicate MSD because node clustering forces each node into a single cluster, even when boundary nodes are influenced by multiple seeds. To address this, the paper integrates automated Latent Space Edge Clustering (aLSEC; Pham and Sewell, 2024) with Community-based Label Propagation (CLP; Zhang et al., 2023). Infected nodes receive a prominence-based “age”

$G_1$4

where $G_1$5 is the number of infected neighbors and $G_1$6 is the degree in the full network, while uninfected boundary nodes receive an exoneration age

$G_1$7

These initialize a label matrix $G_1$8 over $G_1$9 edge communities plus one exoneration column, and propagation proceeds via

$M$0

which converges to

$M$1

For each community $M$2, the source estimate is the infected node with maximal score in column $M$3 of the row-normalized $M$4.

Evaluation uses three ADD HEALTH social networks—addhealth15, addhealth20, and addhealth75—with diffusion simulated at infection probability $M$5 until more than 10% of nodes are infected, $M$6 sources, and 200 independent simulations per configuration (Li et al., 11 Jul 2025). Precision, Recall, and F1-Measure are reported. Across all three networks and all $M$7, aLSEC achieves the highest F1-Measure relative to Louvain and Leading Eigenvector within the same CLP framework, with the gain especially pronounced for $M$8. The result is presented not as a change in the diffusion model, but as a representational correction for mixed-membership source regions.

4. Array-based source enumeration and collaborative signal detection

In sensor arrays and radio processing, MSD commonly denotes deciding how many signal sources impinge on an array. “Source detection via multi-label classification” reformulates this as a deep multi-class classification problem over source count $M$9, using a centrosymmetric linear array of $$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$0 omni-directional elements and the normalized upper triangle of the sample covariance as input (Vijayamohanan et al., 2022). The underlying model is

$$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$1

with sample covariance

$$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$2

For coherent sources, the paper applies forward–backward spatial smoothing (FBSS), constructing

$$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$3

Two detectors are proposed: CNNDetector, a 1D CNN with five stacked convolutional layers, and RadioNet, a ResNet-34 adapted to 1D signals. Source count is recovered by $$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$4. For uncorrelated sources with $$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$5 and $$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$6, CNNDetector reaches 89% overall test accuracy; for correlated sources, CNNDetector degrades to 56.7% validation accuracy, whereas RadioNet + FBSS reaches validation accuracy $$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$7 after 100 epochs and $$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$8 in a 3 non-coherent + 2 coherent case with 1 million samples (Vijayamohanan et al., 2022). The paper reports that learning-based detectors outperform MDL/AIC at low SINR and small $$\mathcal{L} = \mathcal{L}_{det} + \lambda \big( \mathcal{L}_l + \mathcal{L}_h + \mathcal{L}_{con} \big),$$9, while FBSS-MDL/AIC remain competitive at high SINR.

A more explicitly Bayesian formulation appears in “A Bayesian Framework for Collaborative Multi-Source Signal Detection,” which considers $\mathcal{L}_{det}$0 sensors and $\mathcal{L}_{det}$1 snapshots assembled into $\mathcal{L}_{det}$2 under

$\mathcal{L}_{det}$3

with $\mathcal{L}_{det}$4 and $\mathcal{L}_{det}$5 assigned maximum-entropy Gaussian priors (0811.0764). After marginalization over channels and signals using finite random matrix theory, the detector depends only on the eigenvalues $\mathcal{L}_{det}$6 of the sample covariance $\mathcal{L}_{det}$7. Closed-form Bayes factors $\mathcal{L}_{det}$8 and $\mathcal{L}_{det}$9 are derived, enabling both source-presence decisions and source enumeration through $\mathcal{L}_l$0. In a reported example with $\mathcal{L}_l$1, $\mathcal{L}_l$2, $\mathcal{L}_l$3, and SNR $\mathcal{L}_l$4 dB, the Bayesian detector achieves up to about 10% absolute improvement in correct detection rate over the classical power detector at low false-alarm rates, and remains robust when noise power is only known to lie in wide bounded intervals (0811.0764). The methodological contrast with CNNDetector and RadioNet is sharp: one line replaces eigendecomposition-based model selection with learned covariance-feature classification, while the other keeps the spectral statistic and changes the decision rule through prior-driven marginal likelihoods.

5. Localization, multisensor fusion, and multi-band detection

Several MSD formulations move beyond counting to joint model selection and localization. In wireless sensor networks, “New Perspectives on Multiple Source Localization in Wireless Sensor Networks” seeks the unknown number of sources $\mathcal{L}_l$5, their locations $\mathcal{L}_l$6, and powers $\mathcal{L}_l$7 from quantized sensor reports observed at a fusion center through imperfect channels (Nguyen et al., 2015). The sensing model is

$\mathcal{L}_l$8

with quantizer output probabilities expressed through Gaussian $\mathcal{L}_l$9-functions and a channel transition matrix $\mathcal{L}_h$0. The paper runs $\mathcal{L}_h$1 independent, parallel SMC samplers, one per fixed-dimension model $\mathcal{L}_h$2, on tempered targets

$\mathcal{L}_h$3

and estimates model evidence for $\mathcal{L}_h$4. When the truth is $\mathcal{L}_h$5, SMC selects $\mathcal{L}_h$6 96 times versus 85 for importance sampling; the scaled ESS for $\mathcal{L}_h$7 is 0.6276 for SMC versus 0.0019 for IS (Nguyen et al., 2015). The paper also derives the Posterior Cramér–Rao Bound and uses it to analyze the effects of quantization depth, sensor count, and channel reliability.

Distributed multisensor ISAC generalizes MSD to passive object detection across multiple Tx/Rx links. “Distributed Multisensor ISAC” models the received signal on link $\mathcal{L}_h$8 as

$\mathcal{L}_h$9

where $\mathcal{L}_{con}$0 and $\mathcal{L}_{con}$1 are bistatic delay and Doppler, $\mathcal{L}_{con}$2 is clutter, and $\mathcal{L}_{con}$3 is AWGN (Thomä et al., 17 Nov 2025). The paper emphasizes Cooperative Passive Coherent Location (CPCL), excess time-of-flight and excess Doppler estimation, sparse OFDMA/TDMA model-based range–Doppler recovery, per-link CFAR, cross-link coherent or noncoherent integration, and fusion through information filters or distributed consensus. The coherent and noncoherent test statistics are given as

$\mathcal{L}_{con}$4

Here MSD is inseparable from synchronization, clutter suppression, multilink access, and downstream tracking.

Astronomical imaging provides a different multisource-fusion interpretation. MADX detects sources in Herschel-ATLAS maps by matched filtering each band and then forming an inverse-variance weighted sum guided by a chosen SED (Maddox et al., 2020). For band $\mathcal{L}_{con}$5, the matched filter is

$\mathcal{L}_{con}$6

with $\mathcal{L}_{con}$7 the total instrumental-plus-confusion noise spectrum, and multi-band weights are

$\mathcal{L}_{con}$8

Peaks in the combined image determine positions, while fluxes are measured from filtered single-band images at sub-pixel positions. In H-ATLAS-like simulations, the multi-band approach allows reliable source detection a factor 1.2 to 3 lower in flux compared to single-band source detection, reduces the false detection rate by a factor between 4 and 10, and reduces the variance of source position errors by about a factor 1.5; using confusion-aware matched filters yields an overall factor of 1.5 to 3 improvement in catalogue depth relative to a single-band PSF filter approach (Maddox et al., 2020). In this usage, MSD is not source enumeration under a parametric array model, but statistically optimal extraction of many point sources from confusion-limited multi-band fields.

6. Subspace unmixing, error control, and recurring assumptions

“A Multiple Hypothesis Testing Approach to Low-Complexity Subspace Unmixing” uses the acronym MSD for Marginal Subspace Detection, a multiple hypothesis testing procedure for identifying active subspaces under the parsimonious subspace-sum (PS3) model (Bajwa et al., 2014). The ambient observation is

$\mathcal{L}_{con}$9

where only a small subset $S_1,\dots,S_M$00 of subspaces is active. For each candidate subspace $S_1,\dots,S_M$01, the test statistic is

$S_1,\dots,S_M$02

and the estimated active set is

$S_1,\dots,S_M$03

The theory develops thresholds that control the family-wise error rate at any level $S_1,\dots,S_M$04 under bounded deterministic noise or Gaussian noise, using computable geometric quantities such as subspace coherence, local 2-subspace coherence, and quadratic-mean subspace coherence. A central analytical claim is that MSD allows linear scaling of the number of active subspaces as a function of the ambient dimension, thereby breaking the “square-root bottleneck” (Bajwa et al., 2014).

Across fields, the mathematical structures differ, but the assumptions recur. Multi-source domain adaptation for detection usually assumes homogeneous feature spaces and closed-set categories (Yao et al., 2021). Graph-based MSD assumes a known network, an observed infected region, and identifiable boundary nodes (Li et al., 11 Jul 2025). Radio-array MSD assumes narrowband, far-field, stationary sources and calibrated arrays (Vijayamohanan et al., 2022). Bayesian collaborative detection assumes spatially and temporally white Gaussian noise and maximum-entropy priors on channel and sources (0811.0764). WSN localization assumes an energy attenuation model, scalar quantization, and a known channel transition law (Nguyen et al., 2015). MADX assumes known PSFs and approximately Gaussian instrumental noise, while distributed ISAC depends on synchronization and clean reference recovery (Maddox et al., 2020, Thomä et al., 17 Nov 2025). These assumptions are not incidental; they determine identifiability, error control, and whether multi-source structure can be separated from clutter, overlap, or domain shift.

A common misconception is that MSD always means counting physical emitters. The surveyed literature shows a broader usage. In some papers it means detecting objects in an unlabeled target domain given multiple labeled source domains (Yao et al., 2021, Lee et al., 13 May 2026); in others it means recovering infection sources on graphs (Li et al., 11 Jul 2025), identifying active subspaces (Bajwa et al., 2014), or fusing multisensor/multiband evidence to improve detectability (Thomä et al., 17 Nov 2025, Maddox et al., 2020). Taken together, these formulations suggest that MSD is best viewed as a general inverse-inference template: multiple latent causes generate structured observations, and the task is to recover those causes under limited supervision, limited resolution, or imperfect sensing.