Supervised Domain Adaptation

• Supervised domain adaptation is a transfer-learning approach that uses plentiful labeled source data and minimal target labels to mitigate distribution shifts.
• It employs methods like contrastive semantic alignment, adversarial pairing, and graph embedding to align features across domains in few-shot settings.
• Applications span image recognition, disaster response, and spacecraft pose estimation, emphasizing explicit cross-domain alignment and rigorous evaluation protocols.

Supervised domain adaptation (SDA) is the transfer-learning setting in which a model leverages labeled data from a source domain and a smaller amount of labeled data from a target domain to mitigate domain shift and improve target-domain performance. In the deep-learning literature, SDA is typically formulated as learning an embedding or predictor that is simultaneously discriminative for the task and aligned across domains, especially in few-shot regimes where only a handful of target labels are available (Motiian et al., 2017). A substantial line of work also interprets SDA as representation learning under graph constraints, in which same-class cross-domain samples are pulled together and different-class samples are pushed apart (Morsing et al., 2020). Later analysis showed that several prominent SDA objectives can be expressed within a common graph-embedding formalism and that reported few-shot results depend strongly on evaluation protocol design (Hedegaard et al., 2020).

1. Problem setting and formal objectives

In the supervised domain adaptation setting, the source domain $D_S$ provides labeled samples $\{(x_i^s,y_i^s)\}$, the target domain $D_T$ provides a much smaller labeled set $\{(x_i^t,y_i^t)\}$, and the learning problem is complicated by a distribution gap between domains. The central objective is to learn a feature transformation that makes samples of the same class close regardless of domain and samples of different classes well separated, so that target risk is reduced despite the shift (Morsing et al., 2020).

A particularly influential formulation is the Classification and Contrastive Semantic Alignment (CCSA) framework. It uses a Siamese architecture with a shared embedding function $g:\mathcal{X}\rightarrow\mathcal{Z}$, a classifier $h:\mathcal{Z}\rightarrow\mathcal{Y}$, and a total objective

$$\mathcal{L}_{CCSA}(f)=\mathcal{L}_C(h\circ g)+\mathcal{L}_{SA}(g)+\mathcal{L}_S(g),$$

where $\mathcal{L}_C$ is the supervised classification loss, $\mathcal{L}_{SA}$ aligns same-class source and target samples, and $\mathcal{L}_S$ separates different-class cross-domain pairs (Motiian et al., 2017). Because class-conditional distributions are difficult to estimate in the few-shot regime, the method replaces distribution distances with point-wise surrogates: $\{(x_i^s,y_i^s)\}$0 for positive pairs, and

$\{(x_i^s,y_i^s)\}$1

for negative pairs with margin $\{(x_i^s,y_i^s)\}$2 (Motiian et al., 2017).

This formulation made the few-shot SDA setting especially prominent. The reported finding that even one labeled target sample per category can be effective was later echoed by adversarial SDA work, which also emphasized “high speed of adaptation” in the sense that accuracy improves sharply with extremely limited target supervision (Motiian et al., 2017).

2. Principal methodological families

Several methodological families recur in SDA research. They differ mainly in how they encode cross-domain supervision and how directly they impose class-conditional alignment.

Family	Core mechanism	Representative paper
Contrastive semantic alignment	Siamese embedding with alignment and separation losses	(Motiian et al., 2017)
Few-shot adversarial pairing	Four-group pair discriminator for semantic alignment and domain confusion	(Motiian et al., 2017)
Graph-embedding SDA	Intrinsic and penalty graphs optimized end-to-end	(Morsing et al., 2020)

The contrastive family is exemplified by CCSA, where the same-class attraction and different-class repulsion terms are applied explicitly to source-target pairs. The method was presented as a unified framework for supervised domain adaptation and domain generalization, with the supervised setting becoming attractive when only few target data samples need to be labeled (Motiian et al., 2017).

The adversarial family is represented by Few-Shot Adversarial Domain Adaptation (FADA). Rather than using a binary domain discriminator, FADA augments the discriminator to distinguish four groups of sample pairs: source same-class, source-target same-class, source different-class, and source-target different-class. The discriminator is trained with

$\{(x_i^s,y_i^s)\}$3

while the feature extractor is trained adversarially so that source-target same-class pairs resemble source-source same-class pairs and source-target different-class pairs resemble source-source different-class pairs (Motiian et al., 2017). This design explicitly combines domain confusion with semantic alignment.

The graph-embedding family is represented by Domain Adaptation using Graph Embedding (DAGE). In this framework, source and target features are concatenated into $\{(x_i^s,y_i^s)\}$4, and within-class compactness and between-class separability are expressed through intrinsic and penalty graph Laplacians. The core loss is a trace ratio,

$\{(x_i^s,y_i^s)\}$5

optimized jointly with cross-entropy losses on source and target (Morsing et al., 2020). A simple LDA-inspired graph construction, DAGE-LDA, connects same-class cross-domain samples in the intrinsic graph and different-class cross-domain samples in the penalty graph.

A recurrent theme across these families is that “SDA” in foundational work usually denotes more than source pretraining followed by ordinary fine-tuning. The learning objective typically contains an explicit cross-domain alignment mechanism, whether contrastive, adversarial, or graph-theoretic. By contrast, some application papers use “SDA” for transfer by pretraining on the source and fine-tuning on labeled target data; this difference in usage is methodologically significant, though both remain within supervised adaptation.

3. Graph embedding as a unifying perspective and the question of evaluation

A major conceptual consolidation came from the graph-embedding reinterpretation of SDA. The central claim is that SDA methods using pair-wise relationships between source and target data can be formulated as graph embedding, where domain labels are incorporated into intrinsic and penalty graphs (Hedegaard et al., 2020). In this view, the intrinsic graph encodes within-class affinities and the penalty graph encodes between-class relationships; domain membership determines which edges are emphasized.

Within this framework, CCSA, $\{(x_i^s,y_i^s)\}$6-SNE, and Neural Embedding Matching (NEM) become special cases that differ mainly in graph construction. CCSA uses cross-domain same-class pairs as intrinsic edges and cross-domain different-class pairs within a margin as penalty edges. $\{(x_i^s,y_i^s)\}$7-SNE focuses only on the furthest same-class and closest different-class cross-domain pairs. NEM extends CCSA by adding target-target neighborhood preservation (Hedegaard et al., 2020). This perspective suggests that many ostensibly different SDA losses differ less in principle than in the sparsity pattern and weighting of graph edges.

The same work identified substantial reproducibility problems in standard few-shot SDA evaluation. Three issues were highlighted: lack of validation split, test-validation leakage, and overestimated performance due to repeated resampling in which target instances used for testing in one run may appear in training or validation in another (Hedegaard et al., 2020). The proposed rectified protocol introduced a fixed independent test set, training and validation splits drawn only from the remaining target data, and model selection based only on validation data. Under this rectified protocol, reported performance on Office31, Digits, and VisDA dropped slightly; on Office31, the paper reported decreases of about $\{(x_i^s,y_i^s)\}$8, including CCSA from $\{(x_i^s,y_i^s)\}$9 to $D_T$0, $D_T$1-SNE from $D_T$2 to $D_T$3, and DAGE-LDA from $D_T$4 to $D_T$5 (Hedegaard et al., 2020).

This episode established a persistent methodological caution in SDA: gains in the extreme low-label regime are inseparable from the rigor of the data-splitting protocol. A plausible implication is that benchmark leadership in SDA should be interpreted jointly with the evaluation design and not only with the adaptation loss.

4. Explainability and semantically structured adaptation

Explainability has been comparatively rare in SDA, but one explicit attempt is XSDA-Net, described as an “explainable by design supervised domain adaptation framework.” Its abstract states that the method integrates a case-based reasoning mechanism to explain the prediction of a test instance in terms of similar-looking regions in the source and target train images, and that the domain adaptation settings were curated on datasets popularly known to exhibit part-based explainability (Kamakshi et al., 2022). On the basis of the abstract alone, the concrete contribution is thus the integration of case-based reasoning into supervised adaptation rather than a post hoc explanation layer.

Semantically structured adaptation is also visible in semi-supervised extensions that preserve the supervised core. For semantic segmentation, SSDDA addresses a setting with a large amount of labeled source data, a small amount of labeled target data, and a large amount of unlabeled target data. Its two-step design separates cross-domain and intra-domain adaptation: first, image-level mixing between labeled source and labeled target is used to align the feature distribution shift; second, a student-teacher framework performs category-level mixing on unlabeled target images while respecting predicted object boundaries (Chen et al., 2021). The unlabeled-target consistency step uses

$D_T$6

followed by a cross-entropy consistency loss against the mixed pseudo-label (Chen et al., 2021).

Although SSDDA is explicitly semi-supervised domain adaptation rather than purely supervised domain adaptation, it is relevant because it makes the supervised structure of SDA more explicit: cross-domain alignment is performed with labeled source and labeled target data, while unlabeled target data is added only in a second stage. This suggests a broader design pattern in which supervised alignment anchors the representation and additional target data refine it.

5. Application domains and empirical behavior

Recent work shows that SDA is not confined to standard visual benchmarks. It has been deployed in astronomy, remote sensing, disaster response, and spacecraft perception, often under severe source-to-target mismatch.

Application	SDA role	Reported outcome
Gravitational lens finding	Align simulated training with real survey data	SDA + ENN reached AUROC $D_T$7, $D_T$8 (Parul et al., 2024)
Building extraction from off-nadir imagery	Adapt encoder-decoder networks from noisy to clean data	F1 up to $D_T$9, $\{(x_i^t,y_i^t)\}$0, $\{(x_i^t,y_i^t)\}$1, $\{(x_i^t,y_i^t)\}$2 across building types (Neupane et al., 2023)
Cross-disaster building damage detection	Fine-tune xBD models on Ida-BD	Unsharp-enhanced RGB + SDA achieved Macro-F1 $\{(x_i^t,y_i^t)\}$3 (Mouradi et al., 16 Mar 2026)
Spacecraft 6-DoF pose estimation	Joint invariant representation and risk minimization	LIRR-SDA reached PCK $\{(x_i^t,y_i^t)\}$4 on Sunlamp$\{(x_i^t,y_i^t)\}$5 (Singh et al., 17 Sep 2025)

In gravitational lens finding, three domain adaptation techniques were assessed: ADDA, WDGRL, and SDA. The paper reports that WDGRL combined with an ENN-based encoder gives the best performance in an unsupervised setting, while supervised domain adaptation enhances the model’s ability to distinguish between lenses and common similar-looking false positives such as spiral galaxies (Parul et al., 2024). In the reported table, SDA with an ENN backbone achieved AUROC $\{(x_i^t,y_i^t)\}$6, $\{(x_i^t,y_i^t)\}$7, and $\{(x_i^t,y_i^t)\}$8. When spiral contaminants were added to the target set, SDA+ENN reached AUROC $\{(x_i^t,y_i^t)\}$9 and $g:\mathcal{X}\rightarrow\mathcal{Z}$0 (Parul et al., 2024).

In aerial-image building extraction, the target problem is label misalignment in off-nadir imagery. The proposed SDA method pre-trains an encoder-decoder network on a large noisy dataset and fine-tunes all layers on a smaller clean dataset. After benchmarking 43 lightweight CNNs, 7 encoder-decoder networks, 5 optimizers, and 9 loss functions, the study reported that SDA of the best-performing EDN significantly outperformed knowledge distillation and deep mutual learning, with F1 scores up to $g:\mathcal{X}\rightarrow\mathcal{Z}$1, $g:\mathcal{X}\rightarrow\mathcal{Z}$2, $g:\mathcal{X}\rightarrow\mathcal{Z}$3, and $g:\mathcal{X}\rightarrow\mathcal{Z}$4 for low-rise, mid-rise, high-rise, and skyscrapers respectively (Neupane et al., 2023).

In cross-disaster building damage detection, the proposed two-stage ensemble adapts the xView2 first-place method to the Ida-BD dataset using supervised fine-tuning on labeled target data. The ablation study reported that SDA is indispensable: removing it caused damage detection to “fail entirely.” The most robust configuration used SDA with unsharp-enhanced RGB input and attained Macro-F1 $g:\mathcal{X}\rightarrow\mathcal{Z}$5 on the unseen Ida-BD test split (Mouradi et al., 16 Mar 2026).

In spacecraft pose estimation, the proposed framework extends SDA to keypoint regression using the Learning Invariant Representation and Risk paradigm. It combines empirical losses on synthetic and real labeled data with adversarial representation alignment and a risk-invariance term. On SPEED+, the paper reports that with only $g:\mathcal{X}\rightarrow\mathcal{Z}$6 labeled target data the method matches or surpasses oracle performance trained on larger fractions of labeled data; on Sunlamp$g:\mathcal{X}\rightarrow\mathcal{Z}$7, LIRR-SDA achieved PCK $g:\mathcal{X}\rightarrow\mathcal{Z}$8, attitude error $g:\mathcal{X}\rightarrow\mathcal{Z}$9, and ESA score $h:\mathcal{Z}\rightarrow\mathcal{Y}$0 (Singh et al., 17 Sep 2025).

Across these applications, a consistent pattern emerges. In some domains, SDA denotes explicit domain-alignment objectives; in others, it denotes supervised transfer from source to target by fine-tuning. The empirical record provided here shows that both usages can materially improve performance when a modest but nonzero labeled target set is available.

6. Scope, acronym ambiguity, and adjacent variants

The abbreviation “SDA” is not unique to supervised domain adaptation. In “Self-Adversarial Disentangling for Specific Domain Adaptation,” SDA denotes Specific Domain Adaptation, a setting in which the domain shift is aligned along a demanded-specific dimension such as fog, rainfall, or field of view (Zhou et al., 2021). The proposed SAD framework uses a Domainness Creator to generate source samples with controllable domainness and a Self-Adversarial Regularizer to disentangle domainness-specific and domainness-invariant features. Its total objective is

$h:\mathcal{Z}\rightarrow\mathcal{Y}$1

and the paper states that it introduces no extra costs at inference time (Zhou et al., 2021). This is conceptually adjacent to supervised DA, but it is a different problem setting.

Another adjacent line is semi-supervised domain adaptation with dynamic distribution alignment for EEG emotion recognition. SDA-DDA aligns marginal and conditional distributions with MMD and CMMD, dynamically shifts weight from marginal to conditional alignment through training, and filters pseudo-labels by confidence. The paper states that the framework explicitly minimizes cross-entropy on labeled data and is directly applicable to supervised domain adaptation settings when some target labels exist (Tang, 23 Feb 2025). Here the relation to SDA is architectural and methodological rather than definitional.

A hybrid domain-adaptive direction appears in ultrasound computed tomography. SDA-UCT uses supervised pre-training on simulation data, followed by physics-informed self-supervised fine-tuning on in-vivo data, with LoRA updating only $h:\mathcal{Z}\rightarrow\mathcal{Y}$2 of parameters while achieving comparable performance to full fine-tuning (Liu et al., 24 May 2026). Because the second stage is self-supervised rather than target-labeled supervised adaptation, this is not a pure SDA method. It nonetheless indicates a broader trend in which supervised pretraining on the source domain is combined with more label-efficient target adaptation.

These neighboring usages clarify two common misconceptions. First, SDA is not a single algorithmic recipe: the term spans contrastive, adversarial, graph-based, and fine-tuning-centered methods. Second, the acronym itself is ambiguous across the literature. This suggests that reading “SDA” in isolation is insufficient; the exact supervision regime, alignment mechanism, and target-data assumptions must be specified to determine what kind of adaptation is actually being studied.