acceptodds
Withdrawn from ICLR 2027

What Should Be Shared Across Scenes? Self-Supervised Scene-Style-Conditioned Correspondence Learning for Hyperspectral Image Domain Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-scene hyperspectral image (HSI) domain adaptation aims to transfer knowledge from a labeled source scene to an unlabeled target scene under substantial spectral and spatial differences. Existing methods mainly rely on feature alignment, pseudo-labeling, or cross-scene similarity to establish transferable relations between the two scenes. However, such relations are typically inferred directly from observed features or predictions, without separating scene-dependent spectral variation from class-related structure. To address this issue, we propose scene-style-conditioned correspondence learning (SCCL), a cross-scene adaptation framework that shares only relations that remain meaningful under scene-style variation. We first model scene-dependent feature statistics with a bounded stochastic scene-style posterior and construct scene-style transport under a spectral-admissibility constraint. We then examine whether source–target semantic relations persist across the resulting admissible views, yielding a correspondence-survival probability. The surviving relations are organized by pairwise correspondence admissibility and a globally normalized correspondence plan, which directly induces soft target semantics for final learning. Experiments on Houston, Pavia, and Shanghai–Hangzhou cover six directed cross-scene tasks, demonstrating the effectiveness and superiority of the proposed SCCL method. The code is available at https://github.com/Sky-byte-box/SCCL_code.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cross-scene hyperspectral image domain adaptation, self-supervised learning, scene-style conditioning, cross-scene correspondence learning.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gaSLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

is paper proposes a transfer learning method for hyperspectral image (HSI) classification. The method consists of three main components: scene-style posterior estimation and normalization, scene-style transport, and style-survival correspondence discovery. With these components, the transfer learning problem for cross-scene HSI classification is reformulated as a scene-style estimation task. However, the model and algorithm details are not sufficiently clear, which, in my view, weakens the justification for the authors’ claims. The experimental results are competitive with those of SOTA methods.

Critical strengths

The experimental results appear promising.

Critical weaknesses

Its contributions are rather limited, since it relies solely on style-modeling to characterize scene variations. Moreover, style-modeling bears strong similarities to shared feature learning, given that the observed images are constructed from style component and shared component. The proposed method also fails to adequately substantiate its claim that the scene-type posterior distribution can be effectively estimated. Finally, the writing of paper is poor, as many critical technical details are insufficiently described and explained.

  1. Can the Gaussian distribution in Eq. (1) and affine transformation precisely characterize scene style variations?
  2. Can the intermediate style be represented by Eq. (2)?
  3. What is the theoretical foundation for using Eq. (4) to measure whether the same class assignment survives?
Questions for authors

Why is self-supervised learning capable of modeling scene-style variations, given that the target data lack sufficient class label information? Why does the proposed method outperform other related transfer learning approaches?

Reviewer_g9UGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SCCL for transferring hyperspectral image classification knowledge from a labeled source scene to an unlabeled target scene. The method generates controlled changes in scene style and identifies source-target relationships that remain stable across these changes. It then uses these reliable relationships to generate soft labels for target-scene training. Experiments on six transfer tasks demonstrate the effectiveness of the approach and report improved overall performance over the compared methods.

Critical strengths

The paper addresses the problem of hyperspectral image domain adaptation. Checking whether classifier and prototype assignments persist across constrained transformations provides a useful hypothesis for improving target supervision. Combining this check with target-memory consistency is relevant to researchers working on cross-scene HSI adaptation and pseudo-label reliability.

Critical weaknesses

The novelty of the proposed problem formulation is not sufficiently established. The central idea of transferring semantic information that remains invariant under scene-style variations has been extensively explored in domain adaptation and domain generalization. In particular, prior studies on cross-scene HSI classification have investigated style-content disentanglement, or consistency-based semantic transfer. The authors should clarify how their formulation differs from these existing paradigms.

Fig. 2 demonstrates the overview of the proposed SCCL. The part B, however, proposes a self-supervised semantic learning module without clearly demonstrating the SSL objectives in the method. The authors wrote: The contrastive, masked-modeling, and redundancy-reduction elements in Fig. 2 schematically summarize self-supervised semantic cues at the shared representation stage rather than three additional optimization branches in the formulation below. It is very confusing. If the contrastive, masked-modeling, and redundancy-reduction components in Fig. 2 are merely schematic and are not explicitly optimized, what is the actual self-supervised learning objective in the proposed SCCL framework?

The authors critically addressed the mixture of scene-dependent raviation and class-related information, but in this paper, the method and experiments did not prove the mixture was effectively removed. Section 3.2 disentangled the style and content only by channel-wise mean/std normalization, without demonstrating and clearly evaluate how the scene and class information are separated.

From Eqs. (4)-(7), the default survival threshold appears to restrict each accepted target sample to a single source class. This raises a concern about what global correspondence optimization adds to target supervision when the class assignment is already fixed. Redistributing correspondence weights within that class produces the same one-hot target label, and no other use of the correspondence plan in the training loss is evident. The substantial performance gain attributed to global normalization in Table 2 is therefore difficult to explain. The authors should clarify how this optimization affects learning and reconcile this explanation with the implementation and ablation results.

Questions for authors

Under the default threshold, each accepted target sample appears to be linked to only one source class. If redistributing correspondence weights leaves its target label unchanged, how does global optimization affect training? Please explain how this accounts for the improvement over local normalization in Table 2 and the sensitivity to correspondence parameters in Figure 4(c).

Table 2 shows that multi-view survival improves final classification accuracy compared with the single-view variant. Do the authors also have measurements of the accuracy of the target pseudo-labels selected by these two variants and the proportion of target samples retained? These results would help clarify whether the improvement comes from selecting more reliable labels.

The current experiments seem to not solve the problem of style-class entanglement. There is no strong evidencee that the selected cross-domain relationship is more accurate in semantics. The ablations only evaluate the modules like spectral constraint and target memory conribute to the final OA result.

Reviewer_YTW7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SCCL for cross-scene hyperspectral image domain adaptation. The core idea is to shift away from directly aligning source and target features, and filter out cross-scene correspondences that remain semantically valid regardless of scene-style variations. Experiments on benchmark datasets demonstrate the effectiveness of the proposed method.

Critical strengths

1、The central question—whether a relation survives admissible scene-style variation rather than whether features align—is a reasonable reframing. 2、Component ablations (Table 2) and pairwise parameter sweeps (Fig. 4) are thorough.

Critical weaknesses
  1. Negative Transfer Analysis: Negative transfer on the H13→H18 task is not discussed: SCCL achieves 64.97±3.00, whereas Source-only reaches 69.45±1.19 (Table 1)—meaning adaptation causes a drop of 4.5 points on one of the six tasks, despite the method's premise of robustness to scene variations. The manuscript only notes that CACL scores better there. Please provide a detailed analysis of this failure mode.
  2. Hyperparameter Sensitivity and Selection: The proposed method involves a large number of hyperparameters. Are these hyperparameters kept fixed across all test scenes? Furthermore, how were they selected—were they tuned using the test set?
  3. Novelty and Literature Comparison: The method lacks sufficient novelty as similar approaches already exist in the literature, which significantly diminishes its technical contribution. For example, relevant recent works include: Cross-SPECL: Cross-Scene Hyperspectral Image Classification via Spectral Stability Exploiting and Causal Learning (TGRS 2026) Decoupling and Aligning Modality-Shared Semantics for Single-Domain Generalization in Hyperspectral Image Classification (JSTARS 2026)
  4. Insufficient Benchmarks and Comparisons: Compared to the benchmarks used in studies like Cross-SPECL, the evaluation benchmarks in this paper are quite limited. The authors should complete the full evaluation suite and include comparisons against these state-of-the-art methods.
Questions for authors
  1. Regarding the baselines, did the authors re-run all baseline methods using a unified backbone architecture rather than utilizing the original networks proposed by respective baselines?
  2. Please address all the questions and concerns raised in the Weaknesses section above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.