acceptodds
Under review as a conference paper at ICLR 2027

HyperFlow: Class-Incremental Semantic Change Detection via Hyperbolic Flow

20 views

Abstract

Semantic change detection in remote sensing aims to identify and characterize land-cover transitions observed in pairs of satellite or aerial images. Existing methods assume a fixed set of change categories, limiting their applicability in real-world scenarios where changes of interest may vary, for example across geographies or over time. In this work, we tackle the problem of class-incremental semantic change detection, where a model progressively learns to detect new types of semantic change. To do so, we exploit the hierarchical structure of land-cover semantics, where newly introduced categories correspond to fine-grained refinements of existing ones, and represent said semantic categories in hyperbolic space, which is particularly well suited to represent such hierarchical relationships. To capture semantic changes, we introduce HyperFlow, a method relying on Busemann prototypes, modeling change as an anomaly in a learned semantic flow within the hyperbolic embedding space induced by segmentation. HyperFlow couples change detection to class-incremental semantic segmentation, enabling class-incremental semantic change detection. We also present HiChange, a new large-scale, high resolution remote sensing dataset with detailed, fine-grained hierarchical labels. Experimental results demonstrate that HyperFlow achieves both strong change detection and semantic segmentation performance across incremental steps. In particular, we show that progressively refining semantic categories, rather than considering them all directly, yields improved change detection, highlighting the benefits of incremental learning for semantic change detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Semantic Change Detection, Remote Sensing, Hyperbolic space, Class-Incremental Learning, Busemann prototypes, Hyperbolic flow

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Kx4ROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a hyperbolic flow-based method for remote sensing change detection. While this approach offered a novel perspective on the problem and demonstrated strong performance on hierarchical datasets, it underperformed on conventional datasets and failed to outperform existing CISS methods. Given that most publicly available remote sensing change detection datasets lack a hierarchical structure, further discussion is needed regarding how to address this limitation and improve performance in such contexts.

Critical strengths
  • The authors successfully applied the CISS technique to change detection tasks using remote sensing imagery.
  • Furthermore, they demonstrated the potential for incremental change detection by representing semantic information in hyperbolic space and utilizing its flow, achieving strong performance on two prominent public datasets.
  • To validate this approach, they introduced a dataset named "HiChange" and conducted a validation study.
Critical weaknesses
  • Table 1 suggests that the proposed method achieves overwhelming performance. However, while the difference in accuracy compared to AnyChange or SAM3 is merely 0.2 to 0.3, the figures for F1-score, Precision, and Recall show disparities of several-fold magnitude. The numerical values ​​presented in this table warrant a thorough verification.

  • Although the proposed method was well-suited to the HiChange dataset, which features a hierarchical structure, most existing datasets lack such a structure. Consequently, the method performs poorly on the second dataset listed in Table 1.

  • Table 3 compares the proposed method with other CISS-based approaches; despite being tailored specifically for remote sensing, the proposed method shows performance that is either similar to or lower than that of existing methods. Additionally, the specific significance of the hyperbolic representation used here needs to be clarified.

Questions for authors
  • There are many change detection datasets (whether binary or non-binary) that do not have a hierarchical structure; would the validity of the proposed method be verified using these datasets?

  • As noted in the section regarding weaknesses, the figures in Table 1 appear questionable; I would appreciate it if you could verify them again.

  • Regarding Table 3, the proposed method does not appear to outperform existing CISS methods. A detailed analysis and explanation of the proposed method's advantages over these existing approaches are required.

3 comments

We thank the reviewer for the feedback and for recognizing the potential of applying CISS techniques to incremental change detection, the novelty of hyperbolic flow formulation which achieves strong performance, as well as the value of the proposed HiChange dataset. We address each concern below.

Weakness 1 / Question 2 - Discrepancy between Accuracy and other metrics in Table 1

The values in Table 1 are correct. The observation of small differences in accuracy despite large differences in F1, Precision and Recall is a consequence of the large class imbalance in CD datasets. The SECOND dataset presents only 20% of changed pixels, against 6.6% for Hi-UCD-mini and 5.5% for HiChange. Consequently, accuracy alone is not informative: a trivial classifier that always predicts “unchanged” could already achieve 95% accuracy on HiChange, without detecting any changes. This also explains why small differences in accuracy can be related to large ones in F1, Precision and Recall. We report accuracy only for completeness, as it is a commonly used metric, but since F1, Precision and Recall focus directly on the detection of the minority (=changed) class, they provide a better assessment of CD performance under such class imbalance. In particular, F1-score is the most relevant metric for evaluating CD performance, as it jointly accounts for Precision and Recall, and on this metric HyperFlow clearly outperforms competing methods. To avoid any further confusion, we have removed Accuracy from the reported results.

Weakness 2 / Question 1 - Performance on non-hierarchical datasets

We respectfully disagree. We evaluate HyperFlow on two datasets that do not present such a hierarchical structure (SECOND and Hi-UCD-mini), and on these datasets HyperFlow clearly outperforms the baselines on F1. On SECOND, Hyperflow achieves a F1 of 43.0 against 34.7 for the best unsupervised baseline, and on Hi-UCD-mini 44.3 against 26.5.

Weakness 3 / Question 3 - Comparison with CISS methods in Table 3

The goal of our work is to tackle the problem of class-incremental semantic change detection (CI-SCD), not to outperform dedicated CISS methods on CISS, which is not the task that we are addressing. Table 3 is meant to show that HyperFlow achieves state-of-the-art CISS performance, while additionally being able to perform change detection, which competing CISS methods cannot do. Table 3 shows that this added capacity to perform CD does not come at the cost of incremental segmentation.

Summary

We sincerely appreciate the reviewer's careful assessment. Most of the concerns stem from evaluating the results through accuracy only, a metric that is not suited to the highly imbalanced class distribution of CD. When the results are read through the relevant metrics (F1, precision, and recall), our method outperforms the baselines on both hierarchical and non-hierarchical datasets. In our response, we addressed all comments raised.

Thank you for author's response. I have read the author's response to my questions, but it has not changed my evaluation. Then I have decided to maintain my score.

We acknowledge the reviewer's decision of keeping the score. Out of respect for the time and investment we put in addressing all reviewer's concern, we believe however that the reviewer should provide justifications for their choice. We appeal to the principle of "constructive reviews" and ask the reviewer to provide a justification for their decision, in the same way as the reviewer asked us to provide justifications and address their concerns.

A good review should be constructive and help the authors to improve the quality of their paper, even when rejected. It is upon such principles that science is built.

Reviewer_n6BbOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates semantic change detection under the condition of gradual refinement of class hierarchies and proposes HyperFlow, which combines hyperbolic semantic segmentation with prediction-residual-based change detection. The method adopts Busemann prototypes to represent semantic classes, initializes new subclasses via parent-class prototypes, and trains the segmentation model with hierarchical loss. The segmentation model is then frozen, and a vector field induced by Busemann features is used to predict semantic embeddings at another time phase, so as to identify changes via reconstruction residuals. This paper also constructs the HiChange dataset with hierarchical labels, evaluates incremental refinement on this dataset, and assesses static change detection on SECOND and Hi-UCD-mini.

Critical strengths
  1. The research problem has certain practical value. As monitoring demands evolve, land cover categories may need to be gradually refined from coarse granularity to fine granularity. This paper introduces this requirement into semantic change detection and provides a hierarchically structured dataset with geographically separated splits, offering a potential resource for further research.
  2. The system is relatively fully integrated and accompanied by useful ablation analyses. The paper presents change detection results on three datasets and compares Euclidean and hyperbolic constructions, different prototype sources, backbones, and embedding dimensions. The appendix also discusses threshold sensitivity and the limitations of automatically generated labels.
Critical weaknesses
  1. The main algorithmic components of this paper have direct counterparts in existing work: Section 3.2 adopts the horospherical classifier from Berg et al., Section 3.4 uses the existing tree-min hierarchical loss, and the feature-mapping induced vector field and exponential map update in Section 3.5 and Appendix C.3 inherit the Riemannian residual network construction from Katsman et al.; hierarchical category refinement and incremental semantic segmentation in hyperbolic space have also been studied by works such as TOPICS. The novel contribution of this paper mainly lies in the combination of these components for semantic change detection and the reuse of segmentation prototypes, yet existing experiments have not clearly established substantial advantages of this combination over simpler alternatives.
  2. In particular, in Table 10(b) of Appendix F.4, after applying optimal thresholds respectively, the F1 score of the ordinary Euclidean predictor is 35.5 while that of the hyperbolic version is 35.9, a difference of only 0.4 percentage points. The paper does not report the variance across repeated experiments, and this comparison is limited to the initial four-class task, making it impossible to verify whether the geometric advantage persists stably during incremental refinement. In addition, the paper lacks a vector-field-free baseline that directly computes bi-temporal distances on the same frozen semantic embedding; comparisons with discrete Segm. diff. alter both the representation form and detection mechanism simultaneously, failing to adequately isolate the effect of the learned flow. The main model also trains features with target-domain semantic supervision, so the performance gap from ordinary unsupervised CD methods cannot be directly attributed to the proposed geometric design. These gaps collectively limit the assessment of the significance of the method’s contributions. 3.The experimental protocol in Section 3.1 reuses the same set of images across all stages, with only progressive label refinement; Appendix D further states that the change detection network is trained from scratch at each stage. While this constitutes a valuable taxonomy-refinement setup, its results cannot directly support effectiveness in more general continual learning scenarios where old data is inaccessible and new data arrives continuously. Table 2 only reports the full incremental performance of HyperFlow itself, and Table 3 compares only segmentation mIoU against other methods.
  3. Consequently, complete CI-SCD comparisons under identical training curricula and label access conditions remain absent. Meanwhile, the HiChange dataset, the only one supporting incremental evaluation, partially relies on FLAIR automatic predictions to generate semantic labels, and change labels are constrained to the intersection of OCSGE change masks and semantic discrepancies. This mechanism cannot recover fine-grained changes that are not marked by OCSGE. Although the authors acknowledge the associated noise and missing annotations, no independently manually reviewed subsets or stratified label quality statistics are provided. It is therefore difficult at present to assess the annotation coverage of emerging fine-grained class transitions and the impact of baseline label errors on the conclusions of incremental experiments.
Questions for authors
  1. When keeping the same frozen semantic embeddings and threshold selection rules, what CD and SCS results can be obtained by directly using the bi-temporal embedding distance, i.e., setting the vector field to zero? Is the performance gap from full HyperFlow stable across all incremental stages?
  2. In Table 10(b), the hyperbolic and standard Euclidean versions have only a 0.4 percentage point F1 gap. Does this gap exceed random seed fluctuations? When adopting the same coarse-to-fine training curriculum, how does the Euclidean version perform in subsequent incremental stages? Could you provide multiple runs along with PR curves or AP values to further isolate the effect of threshold selection?
  3. What are the respective proportions of BDTopo labels and FLAIR-predicted labels in HiChange? Are there results from independent manual review, especially the precision and omission rate for newly added fine-grained semantic transitions? These results will directly affect the judgments of the dataset and incremental experiments.
  4. Equation (14) in Appendix C.1 seems to omit the conformal factor (λx/2)(\lambda_x/2) in the general base-point exponential map. Please clarify whether this is merely a typo, and specify the exponential map and Riemannian gradient adopted in the implementation.
  5. According to Table 6, 11 subclasses are newly added and three parent classes are replaced in the final step. Following Equation (1), the number of classes should change from 22 to 30 instead of 32. Please provide the complete list of active classes for each stage, and align Figure 5, Table 6 and the experimental setup.
  6. Appendix B states that the semantic labels of SECOND only cover change regions. How are the remaining pixels handled during segmentation training? What target-domain label information is accessible to the main model and each comparison method respectively? This is critical for understanding the experimental setting without change label supervision.
6 comments

We thank the reviewer for the careful and detailed feedback, and for recognizing the relevance of our problem, the value of the HiChange dataset and the breadth of our experimental analyses. We note that the concerns raised are minor and fixable, mostly concerning clarification of experimental protocols or technical details, which supports the correctness of the overall study. We address them one by one below.

## Weakness 1 - Novelty and positioning While we agree that HyperFlow does draw inspiration from existing components, we believe our contribution goes way beyond a mere combination of them. This contribution can be reformulated as:

  1. A new mathematical formulation for unsupervised CD, where the expected evolution of the pixels is simulated by a learned vector field.
  2. A new problem. Our proposed new formulation enables incremental taxonomy refinement, solving an important limitation of CD and enabling CI-SCD, which, to our knowledge, has not been addressed yet.
  3. A new methodological framework based on hyperbolic geometry. To implement our formulation, we consider hyperbolic vector fields and exploit the geometric properties of hyperbolic space to represent hierarchical semantic relationships. We couple segmentation and change detection in a geometrically grounded manner.
  4. A new dataset, HiChange.

The reviewer asks for evidence that our proposed combination of methodological blocks beats simpler alternatives. Results from the paper and the rebuttal support it:

  • Hierarchical prototype initialization improves CD in the non-incremental setting (Table 4 in the original and revised manuscripts)
  • The Busemann vector field outperforms both the no-flow distance and a simple UNET (Table R8 in Weakness 2.2 below, Table 5 in the revised manuscript)
  • The hyperbolic formulation beats the Euclidean one (Table R6 in Weakness 2.1 below, Table 14 in the revised manuscript)
  • The benefits of reusing segmentation prototypes over randomly instantiated ones, demonstrated in Appendix F.3

## Weakness 2.1 / Question 2 - Evidence for a hyperbolic advantage

We thank the reviewer for pointing out the small difference between Hyperbolic and Euclidean variants when using optimal thresholds, and to ask for the effect of the threshold selection on those results. We agree that the comparison should be performed independently of the selection of a particular threshold. We therefore evaluated AUPRC for Hyperbolic, Euclidean and Euclidean-Busemann variants across all incremental steps, as requested by the reviewer. Results are provided in Table R6 below. We also conducted in Table R7 repeated experiments with multiple random seeds to quantify the variability of the results and ensure that the differences between variants are significant.

SpaceT40T^0_4T71T^1_7T132T^2_{13}T163T^3_{16}T204T^4_{20}T225T^5_{22}T306T^6_{30}
Hyp0.2960.2110.2180.2270.2280.2500.216
EucBus0.2000.2250.2060.2070.2070.2140.204
Euc0.2110.2270.2080.2090.2070.2190.204

Table R6: Comparison of hyperbolic and Euclidean formulations for HyperFlow on all increments of the HiChange dataset. For the results to not be tight to a single threshold, AUPRC is reported.

Experiment#1#2#3#4#5MEANSTD
AUPRC0.2960.3070.2980.3220.3140.3070.011

Table R7: Multi-seed experiments results on T40T^0_4. AUPRC is reported for all experiments, together with the mean and standard deviation over all experiments.

We observe that the hyperbolic variant outperforms both euclidean variants across all incremental steps except T1. These improvements are significant, exceeding random seed fluctuations (STD).

We incorporated the additional results and updated discussion into the revised manuscript in the Supplementary Materials F4.

Weakness 2.2 / Question 1 - Contribution of the learned flow

We agree that comparing HyperFlow only against segmentation differencing does not fully isolate the contribution of the learned vector field. To address this concern, we conducted additional experiments with two complementary baselines:

  • Direct embedding-distance baseline: Instead of learning a vector field, we directly compute the distance between the embeddings at the two timestamps, i.e., d(z1,z2)d(z_1,z_2).
  • Simplified vector-field baseline: We replace the proposed vector-field parameterization with a simpler predictor, namely a UNET predicting directly the embeddings of the second image (z2z_2) based on the embeddings of the first image (z1z_1). Results are reported in the table below.

We observe that Busemann vector fields outperform simpler approaches.

Vector fieldF1.Pre.Rec.AUPRC
None, simple difference28.021.042.40.247
Simple UNET21.421.021.90.135
Busemann Vector field29.620.552.80.296

Table R8: Comparison of vector field formulations on HiChange for the task T40T^0_4.

The revised manuscript now includes this additional experiment in section 5.3.

## Weakness 2.3 - Incremental claims

We would first like to clarify that our setting is progressive taxonomy refinement on a single image corpus, as already explicitly stated in section 2.2 of the original submission. We are dealing with class-incremental learning without data-incremental learning. This setting has already been considered in past literature, see Lin et al. (2022) [1]. We have made the distinction between class- and data-incremental learning clearer in the abstract and introduction.

This setting is also required by the nature of remote sensing images that are quite different from the natural images that are considered in traditional CISS studies. Indeed, for natural images, new classes can be introduced through entirely new images. In remote sensing, however, a new category necessarily appears on images presenting objects of previously learned categories, and vice versa: newly added categories were present in previously considered images, just labeled differently. It is not possible to build a dataset that contains only images of the new classes, because a given image always contains both old and new categories. Moreover, in remote sensing applications, the objective is not simply to observe different geographical areas once, but rather to monitor the same areas repeatedly to ensure continuous tracking of land-cover and land-use changes, justifying the choice of this setting.

As for the comparison with other CISS baselines, reusing images does not favor our method. Nothing in the methodology of MiB, PLOP, DKD or TOPICS requires new images, and with the same images available, forgetting and background shifts are, if anything, easier to handle.

Clarification of the protocol for baseline methods.

An important aspect of our approach is that parent prototypes are kept after a category is refined, but are no longer used for prediction. Instead, predictions are made only among the active child prototypes. For a fair comparison, we apply the same rule to all methods: logits corresponding to parent classes are set to 0. This ensures that comparison methods are not penalized for predicting parent categories over children ones.

We clarified the comparison protocol in the Supplementary Materials D, Comparison protocol for CISS methods.

The reviewer also points out that Table 2 only reports the full incremental performance of HyperFlow itself, while Table 3 compares only segmentation mIoU against other methods. We understand the concern to be that no single table compares HyperFlow with baselines on change detection and segmentation jointly. To our knowledge, HyperFlow is the first work to address CI-SCD; no existing other method can be evaluated on both incremental CD and segmentation. We therefore have to evaluate the two capabilities separately: First, CD capacity compared with state-of-the-art CD methods (Table 1), and incremental change detection (Table 2); Second, incremental segmentation against CISS methods (Table 3).

Therefore, we respectfully do not agree with the fact that complete CI-SCD comparisons under identical training curricula and label access conditions remain absent.

[1] Lin, Zhiqiu, et al. "Continual learning with evolving class ontologies." Advances in Neural Information Processing Systems 35 (2022): 7671-7684.

Weakness 3 / Question 3 - Reliability of HiChange annotations

Exhaustive manual annotation of a semantic change detection dataset of this size and with such fine-grained annotations is not realistic. For comparison, LEVIR-CD required about 120 person-days for 637 image pairs of size 1024x1024 pixels [2], labeled only for building changes, i.e. without having semantic annotations nor the large semantic diversity considered with HiChange. The HiChange dataset being approximately 40 times larger, the amount of manual labeling is infeasible, especially accounting for the extra effort to manually label hierarchical semantic classes. This same reasoning also makes a manual audit of a test subset for the rebuttal not realistic either. For such an audit to be meaningful, at least 1% of the images would need to be reviewed, requiring 1.5 person-months for the change labels alone, and even more for the segmentation labels. We therefore decided not to pursue this request and hope the reviewer will understand our reasoning.

That being said, we agree that the reliability should be discussed. We instead propose to estimate the proportion of pixel errors using available information on the OSCGE, BDTOPO and FLAIR products. Specifically, we quantify in the table below the proportion of benchmark pixels originating from each source (BD TOPO, and FLAIR).

OrigineTraining setValidation setTest set
BDTOPO7.94 %8.65 %7.87 %
FLAIR92.06 %91.35 %92.13 %

Table R9: Percentage of pixels with semantic segmentation annotations from FLAIR and BDTopo,respectively.

For simplicity, let’s round the numbers and assume that 92% of the pixels are labeled by FLAIR and 8% by BDTOPO. The overall accuracy of the semantic information of the dataset can be approximated as: Accsem=0.92⋅AccFLAIR+0.08⋅AccBDTOPO.Acc_{sem} = 0.92\cdot Acc_{FLAIR} + 0.08\cdot Acc_{BDTOPO}.

The FLAIR-version we used (FLAIR-HUB_LC-A_swinbase-upernet) has an accuracy of 77.45%. For BDTOPO, a direct accuracy is not available. Nevertheless, its quality specifications indicate low inter-class confusion rates (below 1%) and completeness requirements of 95% for buildings, 100% for natural parks and reserves, and 98% for paved road networks, a planimetric accuracy of 0.1 m. The 0.1 m planimetric accuracy of the selected BDTOPO data is below pixel resolution; therefore, we assume 100% planimetric accuracy for these annotations. Overall, for our calculations, we therefore consider an accuracy of 95% for BDTOPO. The overall accuracy of the segmentation annotations of the dataset can therefore be estimated at: Accsem=0.92⋅0.7745+0.08⋅0.95=0.79Acc_{sem} = 0.92\cdot 0.7745 + 0.08\cdot 0.95 = \boldsymbol{0.79}

Regarding change annotations, the documentation of OCS GE reports omission and excess rates of 12.1% and 14.2%, respectively. Since changed pixels represent approximately 5.5% of the dataset, the combined omission and excess errors correspond to approximately 0.055⋅0.121+0.055⋅0.142=0.0140.055\cdot 0.121 + 0.055\cdot 0.142 = 0.014 = 1.4% of the dataset with potential errors, i.e an accuracy of 98.6%. Note that this accuracy does not account for the filtering step, which would reduce the excess detection of changes and further increase the accuracy.

We included this analysis in the revised manuscript (Supplementary Materials A4 - Limitations).

[2] Chen, Hao, and Zhenwei Shi. "A spatial-temporal attention-based method and a new dataset for remote sensing image change detection." Remote sensing 12.10 (2020): 1662.

Reviewer_Q2BvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies class-incremental semantic change detection in remote sensing, where land-cover categories are progressively refined into finer subclasses. Its proposed method, HyperFlow, combines supervised semantic segmentation in hyperbolic space with a change detector trained without explicit change labels. Semantic classes are represented using Busemann prototypes. When a class is refined, child prototypes are initialized near their parent. A separate network learns a vector field that predicts second-timestamp embeddings from first-timestamp embeddings; large reconstruction residuals indicate change. The paper also introduces HiChange, comprising 1,000 high-resolution image pairs with a hierarchy supporting 4–32 active categories. The contribution primarily lies in connecting established horospherical classification and Riemannian vector-field constructions to evolving semantic change detection, together with the new benchmark. The underlying classifier and vector-field machinery have clear predecessors in Berg et al., BMVC 2024 and Katsman et al., NeurIPS 2023.

Critical strengths

1.The problem formulation addresses a meaningful limitation of existing change-detection systems.

  1. The complete pipeline improves upon segmentation differencing across all three datasets.

  2. The progressive-training comparison includes meaningful controls.

  3. HiChange could be a useful research resource, and its construction is reasonably transparent.

Critical weaknesses
  1. The evidence for a specifically hyperbolic advantage is substantially weaker than the paper’s interpretation. In Appendix F.4, Table 10a shows a large advantage under triangle thresholding: 29.6 F1 for HyperFlow versus 19.7 for the direct Euclidean predictor. However, with individually optimized thresholds in Table 10b, these become 35.9 versus 35.5, a difference of only 0.4 percentage points. The Euclidean Busemann variant reaches 34.9. These results suggest that much of the reported advantage is associated with the interaction between score distributions and threshold selection. That interaction may itself be useful, but it does not establish a substantial representational advantage. No repeated-run uncertainty is reported, and this comparison covers only the four-class base task, despite hierarchical refinement being the main motivation. The smallest convincing resolution would be precision–recall/AUPRC comparisons, a clearly specified validation-based threshold protocol, and a matched Euclidean comparison at a later refinement stage.

  2. The incremental-learning claims extend beyond what the evaluation establishes. Sections 2.2 and 3.1 explicitly reuse the same image corpus across all steps while refining its labels. Appendix D resumes segmentation training and retrains the change detector from scratch. This demonstrates progressive taxonomy learning on a retained corpus; it does not establish adaptation when historical imagery becomes unavailable or quantify retention of previously learned capabilities. This distinction matters for Table 3: the cited TOPICS protocol uses disjoint data subsets when refining known classes and explicitly excludes showing the same image under different labeling taxonomies. The adaptation of these baselines to HyperFlow’s protocol therefore needs explanation. There is also a concrete comparison problem in §5.2: later-step HyperFlow results are compared against Table 1’s base-task CD baselines, although the ground-truth change masks evolve with the taxonomy. These numbers do not establish superiority on the later tasks. Baselines should be evaluated against the same step-specific masks, and broader continual-learning claims should be narrowed or supported with retention measurements.

  3. The reliability of the sole incremental benchmark is insufficiently quantified. Appendix A.3 constructs semantic labels from BDTopo where available and FLAIR predictions elsewhere. Change labels are then defined by intersecting OCSGE change annotations with disagreement between these semantic labels at the selected taxonomy level. Consequently, semantic-label errors can affect both segmentation evaluation and whether a pixel is considered changed. Fine-grained transitions outside the OCSGE change mask cannot be recovered through this filtering. The authors acknowledge annotation noise and minimum-mapping-unit limitations, and Figure 16 illustrates annotation errors, but the paper provides no quantitative independent audit or breakdown of label-source coverage. Because HiChange is the only dataset supporting the central incremental experiments, these uncertainties materially affect the conclusions. A stratified manual audit of a manageable test subset—covering annotation sources, hierarchy depths, and change types—would be more informative than simply increasing dataset size.

  4. The experiments do not cleanly isolate the proposed change-detection mechanism from semantic supervision. Section 5.1 appropriately clarifies that “unsupervised” applies only to the CD stage. Nevertheless, the principal HyperFlow model uses target-domain semantic supervision, whereas the unsupervised baselines do not receive equivalent semantic information. Their comparison establishes performance under different supervision conditions, rather than isolating detector quality. The foundation-model variants are helpful but do not fully resolve this issue. For example, HyperFlow–AnySat obtains 20.1 F1 on HiChange versus 20.2 for KPCAMNet, and 23.7 on Hi-UCD-mini versus 26.5 for AnyChange. Moreover, segmentation differencing is not a sufficient control for the learned vector field: it compares hard labels, while HyperFlow uses continuous distances and thresholding. A direct zero-flow baseline, thresholding d(z1,z2)d(z_1,z_2) with the same frozen embeddings and calibration procedure, is needed to establish how much the learned displacement contributes. A matched simple predictor would further test whether the Busemann construction provides added value.

  5. The stated exponential map contains a mathematical error whose implementation impact is unresolved. Appendix C.1, Eq. (14), omits the position-dependent conformal factor. Under the metric defined immediately beforehand, the exponential map should be exp⁡xc(v)=x⊕c[tanh⁡ ⁣(c λx∥v∥22)vc ∥v∥2],λx=21−c∥x∥22.\exp_x^c(v)=x\oplus_c \left[ \tanh\!\left(\frac{\sqrt c\,\lambda_x\|v\|_2}{2}\right) \frac{v}{\sqrt c\,\|v\|_2} \right], \qquad \lambda_x=\frac{2}{1-c\|x\|_2^2}. The printed formula lacks λx/2\lambda_x/2 inside the hyperbolic tangent. This agrees with the standard expression in Ganea et al., Lemma 2. It is correct at the origin, but the CD update evaluates the map at generally nonzero embeddings. This could be a manuscript typo if the implementation uses a correct library operation. If implemented literally, however, it changes the displacement used by the central method. The authors should identify the implemented operation and clarify whether any reported results require correction.

Questions for authors
  1. Geometry versus calibration: On which split were the “optimal” thresholds in Table 10b selected? Can you report AUPRC and repeated-run uncertainty for the hyperbolic and Euclidean models, including one later incremental stage?
  2. Incremental protocol: At each step, are labels for all previously learned active classes still available on the entire training corpus? How were MiB, PLOP, DKD, and TOPICS adapted to this access pattern? Can later-step CD baselines be evaluated using the corresponding updated change masks?
  3. Benchmark validity: What fraction of evaluation pixels is labeled by FLAIR versus BDTopo, broken down by class? Can an independently checked subset quantify semantic-label accuracy and change-mask precision/recall, especially for fine-grained transitions?
  4. Value of the learned flow: How does HyperFlow compare with thresholding d(z1,z2)d(z_1,z_2) directly using identical embeddings and threshold selection? Does the learned transformation reduce residuals primarily on unchanged pixels, or also suppress genuine changes?
  5. Benefit of progressive refinement: Table 4 reports only the combined SCS score. Can you provide its component scores and CD F1 for all three training strategies? This would establish whether progressive training improves change detection itself, semantic recognition, or both.
  6. Mathematical implementation: Does the code include the missing λx/2\lambda_x/2 factor in Eq. (14)? Please provide the exact exponential-map and Riemannian-gradient implementation, and clarify the effect on results if the printed formula was used.
4 comments

We thank the reviewer for the careful and detailed feedback, and for recognizing that HyperFlow addresses a meaningful limitation of existing change-detection methods, that our study includes meaningful ablations and experiments and that HiChange is a valuable contribution.

We note that the concerns raised are minor and fixable, mostly concerning clarification of experimental protocols or technical details, which supports the correctness of the overall study. We address them one by one below.

Weakness 1 / Question 1 - Evidence for a hyperbolic advantage

We thank the reviewer for pointing out the difference between the results obtained with a fixed triangle threshold and those obtained with individually optimized thresholds. We agree that the comparison should be performed independently of the selection of a particular threshold. We therefore evaluated AUPRC for Hyperbolic, Euclidean and Euclidean-Busemann variants across all incremental steps, as requested by the reviewer. Results are provided in Table R1 below. We also conducted in Table R2 repeated experiments with multiple random seeds to quantify the variability of the results and ensure that the differences between variants are significant.

SpaceT40T^0_4T71T^1_7T132T^2_{13}T163T^3_{16}T204T^4_{20}T225T^5_{22}T306T^6_{30}
Hyp0.2960.2110.2180.2270.2280.2500.216
EucBus0.2000.2250.2060.2070.2070.2140.204
Euc0.2110.2270.2080.2090.2070.2190.204

Table R1: Comparison of hyperbolic and Euclidean formulations for HyperFlow on all increments of the HiChange dataset. For the results to not be tight to a single threshold, AUPRC is reported.

Experiment#1#2#3#4#5MEANSTD
AUPRC0.2960.3070.2980.3220.3140.3070.011

Table R2: Multi-seed experiments results on T40T^0_4. AUPRC is reported for all experiments, together with the mean and standard deviation over all experiments.

We observe that the hyperbolic variant outperforms both euclidean variants across all incremental steps except T1. These improvements are significant, exceeding random seed fluctuations (STD).

Regarding threshold selection, we clarify that both triangle and optimal thresholds reported in the paper are selected using the validation set.

We incorporated the additional results and updated discussion into the revised manuscript in the Supplementary Materials F4.

Weakness 2 / Question 2 - Incremental claims

We would first like to clarify that our setting is progressive taxonomy refinement on a single image corpus, as already explicitly stated in section 2.2 of the original submission. We are dealing with class-incremental learning without data-incremental learning. This setting has already been considered in past literature, see Lin et al. (2022) [1]. We have made the distinction between class- and data-incremental learning clearer in the abstract and introduction.

This setting is also required by the nature of remote sensing images that are quite different from the natural images that are considered in traditional CISS studies. Indeed, for natural images, new classes can be introduced through entirely new images. In remote sensing, however, a new category necessarily appears on images presenting objects of previously learned categories, and vice versa: newly added categories were present in previously considered images, just labeled differently. It is not possible to build a dataset that contains only images of the new classes, because a given image always contains both old and new categories. Moreover, in remote sensing applications, the objective is not simply to observe different geographical areas once, but rather to monitor the same areas repeatedly to ensure continuous tracking of land-cover and land-use changes, justifying the choice of this setting.

As for the comparison with other CISS baselines, reusing images does not favor our method. Nothing in the methodology of MiB, PLOP, DKD or TOPICS requires new images, and with the same images available, forgetting and background shifts are, if anything, easier to handle.

Clarification of the protocol for baseline methods.

An important aspect of our approach is that parent prototypes are kept after a category is refined, but are no longer used for prediction. Instead, predictions are made only among the active child prototypes. For a fair comparison, we apply the same rule to all methods: logits corresponding to parent classes are set to 0. This ensures that comparison methods are not penalized for predicting parent categories over children ones.

We clarified the comparison protocol in the Supplementary Materials D, Comparison protocol for CISS methods.

Regarding the comment on the comparability of CD results across incremental steps, we thank the reviewer for highlighting this issue. We have removed the statement claiming that HyperFlow outperforms the unsupervised CD baselines at later incremental steps, as the evolving ground-truth change masks do not allow for a directly comparable evaluation. Note that since, to our knowledge, HyperFlow is the first work to address CI-SCD, no existing other method can be evaluated on both incremental CD and segmentation. We therefore have to evaluate the two capabilities separately: First, CD capacity compared with state-of-the-art CD methods (Table 1), and incremental change detection (Table 2); Second, incremental segmentation against CISS methods (Table 3).

[1] Lin, Zhiqiu, et al. "Continual learning with evolving class ontologies." Advances in Neural Information Processing Systems 35 (2022): 7671-7684.

Weakness 3 / Question 3 - Reliability of HiChange annotations

Exhaustive manual annotation of a semantic change detection dataset of this size and with such fine-grained annotations is not realistic. For comparison, LEVIR-CD required about 120 person-days for 637 image pairs of size 1024x1024 pixels [2], labeled only for building changes, i.e. without having semantic annotations nor the large semantic diversity considered with HiChange. The HiChange dataset being approximately 40 times larger, the amount of manual labeling is infeasible, especially accounting for the extra effort to manually label hierarchical semantic classes. This same reasoning also makes a manual audit of a test subset for the rebuttal not realistic either. For such an audit to be meaningful, at least 1% of the images would need to be reviewed, requiring 1.5 person-months for the change labels alone, and even more for the segmentation labels. We therefore decided not to pursue this request and hope the reviewer will understand our reasoning.

That being said, we agree that the reliability should be discussed. We instead propose to estimate the proportion of pixel errors using available information on the OSCGE, BDTOPO and FLAIR products. Specifically, we quantify in the table below the proportion of benchmark pixels originating from each source (BD TOPO, and FLAIR).

OrigineTraining setValidation setTest set
BDTOPO7.94 %8.65 %7.87 %
FLAIR92.06 %91.35 %92.13 %

Table R3: Percentage of pixels with semantic segmentation annotations from FLAIR and BDTopo,respectively.

For simplicity, let’s round the numbers and assume that 92% of the pixels are labeled by FLAIR and 8% by BDTOPO. The overall accuracy of the semantic information of the dataset can be approximated as: Accsem=0.92⋅AccFLAIR+0.08⋅AccBDTOPO.Acc_{sem} = 0.92\cdot Acc_{FLAIR} + 0.08\cdot Acc_{BDTOPO}.

The FLAIR-version we used (FLAIR-HUB_LC-A_swinbase-upernet) has an accuracy of 77.45%. For BDTOPO, a direct accuracy is not available. Nevertheless, its quality specifications indicate low inter-class confusion rates (below 1%) and completeness requirements of 95% for buildings, 100% for natural parks and reserves, and 98% for paved road networks, a planimetric accuracy of 0.1 m. The 0.1 m planimetric accuracy of the selected BDTOPO data is below pixel resolution; therefore, we assume 100% planimetric accuracy for these annotations. Overall, for our calculations, we therefore consider an accuracy of 95% for BDTOPO. The overall accuracy of the segmentation annotations of the dataset can therefore be estimated at: Accsem=0.92⋅0.7745+0.08⋅0.95=0.79Acc_{sem} = 0.92\cdot 0.7745 + 0.08\cdot 0.95 = \boldsymbol{0.79}

Regarding change annotations, the documentation of OCS GE reports omission and excess rates of 12.1% and 14.2%, respectively. Since changed pixels represent approximately 5.5% of the dataset, the combined omission and excess errors correspond to approximately 0.055⋅0.121+0.055⋅0.142=0.0140.055\cdot 0.121 + 0.055\cdot 0.142 = 0.014 = 1.4% of the dataset with potential errors, i.e an accuracy of 98.6%. Note that this accuracy does not account for the filtering step, which would reduce the excess detection of changes and further increase the accuracy.

We included this analysis in the revised manuscript (Supplementary Materials A4 - Limitations). Regarding the proportion of pixels originating from each data source, broken down by class, each semantic class is derived exclusively from either BD TOPO or FLAIR, with no overlap between the two sources. We have added a detailed list of the classes associated with each source in the Supplementary Material A3 - Table 8.

[2] Chen, Hao, and Zhenwei Shi. "A spatial-temporal attention-based method and a new dataset for remote sensing image change detection." Remote sensing 12.10 (2020): 1662.

Weakness 4 / Question 4 - Contribution of the learned flow

We agree that comparing HyperFlow only against segmentation differencing does not fully isolate the contribution of the learned vector field. To address this concern, we conducted additional experiments with two complementary baselines:

  • Direct embedding-distance baseline: Instead of learning a vector field, we directly compute the distance between the embeddings at the two timestamps, i.e., d(z1,z2)d(z_1,z_2).
  • Simplified vector-field baseline: We replace the proposed vector-field parameterization with a simpler predictor, namely a UNET predicting directly the embeddings of the second image (z2z_2) based on the embeddings of the first image (z1z_1). Results are reported in the table below.

We observe that Busemann vector fields outperform simpler approaches.

Vector fieldF1.Pre.Rec.AUPRC
None, simple difference28.021.042.40.247
Simple UNET21.421.021.90.135
Busemann Vector field29.620.552.80.296

Table R4: Comparison of vector field formulations on HiChange for the task T40T^0_4.

The revised manuscript now includes this additional experiment in section 5.3.

Regarding the comment about whether the flow reduces residuals mainly on unchanged pixels or also suppress genuine changes, we observe in the above table that the recall increases, meaning that the Busemann formulation does not suppress genuine changes.

Weakness 5 / Question 6 - Mathematical formulation of the exponential map

We thank the reviewer for identifying the missing conformal factor in Eq. 14 of Appendix C.1. We confirm that it is an error in the manuscript that we corrected in the revised version. The implementation uses the correct exponential map and experiments are not affected by this omission.

Question 5 - Benefit of progressive refinement.

We thank the reviewer for this suggestion. We have updated Table 4 to report segmentation (mIoU) and change detection (F1) scores, in addition to the combined SCS score. The updated results show that progressive taxonomy refinement consistently improves both semantic segmentation and change detection (CD) performance compared with non-incremental training. Regarding the effect of hierarchical initialization, we observe that it benefits CD but, somewhat surprisingly, does not improve segmentation performance. Overall, despite this hierarchical organization, both CD and semantic change detection (SCD) performances remain below those achieved in the full incremental setting. Thus, considering CD and segmentation scores separately does not alter the overall conclusion of this study. We updated the discussion of the results in section 5.2 to account for these additional metrics.

Incr.Hier.MetricT40T^0_{4}T71T^1_{7}T132T^2_{13}T163T^3_{16}T204T^4_{20}T225T^5_{22}T306T^6_{30}
YesYesSegm. [mIoU]62.056.233.728.423.520.716.1
CD [F1]29.629.629.829.227.731.626.0
SCD [SCS]39.034.422.419.917.317.613.8
NoNoSegm. [mIoU]62.051.826.820.114.910.86.5
CD [F1]29.624.220.218.219.416.520.9
SCD [SCS]39.030.916.613.711.711.79.1
NoYesSegm. [mIoU]62.051.423.017.413.812.27.9
CD [F1]29.627.328.430.130.129.923.4
SCD [SCS]39.031.717.816.214.414.110.0

Table R5: Revised Table 4 of the original manuscript.

Summary

We sincerely appreciate the reviewer's careful assessment. The additional experiments directly address the concerns raised: threshold-independent AUPRC comparisons across all incremental steps for Hyperbolic vs Euclidean variants, with multi-seed variability, a vector-field-free baseline and a simpler vector-field baseline, a quantitative estimate of HiChange label reliability. Additional details about clarity and incremental protocol have also been provided.

Comments on the paper

2 comments

We thank the reviewers for their thoughtful and detailed feedback. We appreciate the recognition of the relevance of CI-SCD (Kx4R, Q2Bv, n6Bb), the novelty of our formulation (Kx4R) and the breadth of our experimental analyses (Q2Bv, n6Bb). We are glad that the HiChange dataset is perceived as a meaningful contribution ((Kx4R, Q2Bv, n6Bb). The concerns raised are minor, highlighting the correctness of the overall study, and we are confident that these concerns can be fully addressed through our point-by-point responses during the discussion period. We have uploaded a revised version of the paper and appendix, with all changes highlighted in blue, and remain at your disposal if any open questions remain.

Dear Reviewers,

As the discussion period is coming to an end, we would like to kindly follow up on our rebuttal as we have not yet received further feedback. We have responded to all the points raised in your reviews, but we remain available to answer any additional questions or provide other clarifications. We would greatly appreciate any further feedback!

Thank you, again, for your time and consideration!

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.