arxiv: 2604.03212 · v1 · submitted 2026-04-03 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

ProtoFlow: Mitigating Forgetting in Class-Incremental Remote Sensing Segmentation via Low-Curvature Prototype Flow

Jiekai Wu , Rong Fu , Chuangqi Li , Zijian Zhang , Guangxin Wu , Hao Zhang , Shiyin Lin , Jianyuan Ni

show 5 more authors

Yang Li Dongxu Zhang Amir H. Gandomi Simon Fong Pengbin Feng

Authors on Pith no claims yet

Pith reviewed 2026-05-13 20:34 UTC · model grok-4.3

classification 💻 cs.CV

keywords class-incremental learningremote sensing segmentationprototype dynamicscontinual learningcatastrophic forgettingtemporal vector field

0 comments

The pith

ProtoFlow models class prototypes as low-curvature trajectories to reduce forgetting in continual remote sensing segmentation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Remote sensing segmentation must adapt when new classes appear and imaging conditions change across time and sensors. Most incremental methods update representations in isolated steps and allow drift to erase earlier knowledge. ProtoFlow instead treats each class prototype as a point moving through feature space according to a learned temporal vector field. Regularization keeps those paths smooth and keeps different classes separated in geometry. On standard class-incremental and domain-incremental remote sensing benchmarks the approach raises overall mIoU while lowering forgetting relative to strong baselines.

Core claim

ProtoFlow represents class prototypes as trajectories governed by an explicit temporal vector field and jointly optimizes for low-curvature motion and inter-class separation, thereby stabilizing prototype geometry and mitigating catastrophic forgetting across incremental training steps in remote sensing segmentation.

What carries the argument

A time-aware prototype dynamics model that updates class prototypes along a learned temporal vector field subject to low-curvature and separation penalties.

If this is right

Yields 1.5–2.0 point gains in mIoUall over prior incremental baselines on standard benchmarks.
Reduces forgetting rates in both class-incremental and domain-incremental remote sensing settings.
Produces an explicit, interpretable record of how each class prototype moves through training steps.
Stabilizes overall representation geometry without expanding model capacity or storing past samples.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The trajectory view could be tested in other continual vision domains where semantic classes are represented by compact prototypes.
Low-curvature regularization might lessen dependence on replay memory in broader continual-learning pipelines.
Measuring actual prototype curvature on real deployment streams would directly test the central modeling assumption.

Load-bearing premise

Enforcing low-curvature motion and inter-class separation on prototype trajectories is sufficient to prevent representation drift and forgetting without replay buffers or architectural changes.

What would settle it

A controlled sequence of new remote-sensing images whose distribution forces prototypes onto high-curvature paths despite the penalty, after which forgetting on earlier classes measurably increases.

Figures

Figures reproduced from arXiv: 2604.03212 by Amir H. Gandomi, Chuangqi Li, Dongxu Zhang, Guangxin Wu, Hao Zhang, Jianyuan Ni, Jiekai Wu, Pengbin Feng, Rong Fu, Shiyin Lin, Simon Fong, Yang Li, Zijian Zhang.

**Figure 1.** Figure 1: Existing RS CISS treats each step as an isolated snapshot and heuristically corrects drifting prototypes, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: Overall ProtoFlow framework. Non-stationary RS streams bring new classes over time. A segmentation network produces pixel features, which are aggregated by a prototype estimator and stored in a prototype bank. A time-aware ProtoFlow Field predicts how historical prototypes should move, and a prototype regularizer enforces flow consistency, low curvature and class separation. These prototype losses are comb… view at source ↗

**Figure 3.** Figure 3: Per-class correlation between prototype trajectory curvature and forgetting. shuffling the acquisition time τt for a fraction α ∈ {0, 0.25, 0.5, 0.75, 1.0} of training samples (while keeping the class/order of tasks unchanged). We compare: (i) ProtoFlow (time-aware) with true timestamps (α = 0), and (ii) ProtoFlow (shuffled time) trained with different shuffle levels α. For DeepGlobe, we additionally compa… view at source ↗

**Figure 4.** Figure 4: Impact of time shuffling on LoveDA. −1.0 −0.5 0.0 0.5 1.0 −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 1.00 (a) ProtoFlow −1.0 −0.5 0.0 0.5 1.0 (b) w/o Curvature −1.0 −0.5 0.0 0.5 1.0 (c) w/o Separation −1.0 −0.5 0.0 0.5 1.0 (d) w/o Time t-SNE dimension 1 t-SNE dimension 2 Urban Agriculture Water Forest Start End [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

**Figure 5.** Figure 5: Prototype flow visualization on DeepGlobe. We project class prototypes into 2D and visualize their trajectories across incremental steps. (a) ProtoFlow: trajectories are smooth, monotone, and well separated. (b) w/o Curvature: individual trajectories exhibit stronger self-wrapping, oscillation, and backtracking. (c) w/o Separation: trajectories are increasingly crowded into a shared region, resulting in re… view at source ↗

**Figure 6.** Figure 6: Qualitative results on Vaihingen and Postdam [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

**Figure 7.** Figure 7: Robustness to task/domain order on LoveDA. We evaluate three task orders (URM: Urban→Rural→Mixed, RUM: Rural→Urban→Mixed, MUR: Mixed→Urban→Rural) and three random seeds per order for MiR (top) and ProtoFlow (bottom) [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗

**Figure 8.** Figure 8 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

**Figure 9.** Figure 9: Class-level ∆-curvature vs. ∆-forgetting. Each marker corresponds to a semantic class c from DeepGlobe, Vaihingen, LoveDA, iSAID, or GCSS. The horizontal axis is ∆¯κc = ¯κ ProtoFlow c −κ¯ MiR c and the vertical axis is ∆Forgetc = ForgetProtoFlow c −ForgetMiR c . The lower-left quadrant (shaded) indicates classes where ProtoFlow simultaneously reduces curvature and forgetting. Marker color encodes the datas… view at source ↗

**Figure 10.** Figure 10: Class-wise IoU and forgetting distributions on DeepGlobe and LoveDA. and train ProtoFlow from scratch for each pair using the same backbone, optimization schedule, and memory budget as in Sec. 4.1. For each run, we report the final mIoUall and forgetting F on LoveDA. 0.00 0.05 0.10 0.20 sep 0.00 0.10 0.30 0.50 1.00 curve 59.8 60.1 60.5 60.0 60.7 61.0 61.2 60.8 61.8 62.0 62.2 61.7 62.3 62.5 62.8 62.2 61.9 … view at source ↗

**Figure 11.** Figure 11: Sensitivity of ProtoFlow to curvature and separation weights on LoveDA [PITH_FULL_IMAGE:figures/full_fig_p031_11.png] view at source ↗

read the original abstract

Remote sensing segmentation in real deployment is inherently continual: new semantic categories emerge, and acquisition conditions shift across seasons, cities, and sensors. Despite recent progress, many incremental approaches still treat training steps as isolated updates, which leaves representation drift and forgetting insufficiently controlled. We present ProtoFlow, a time-aware prototype dynamics framework that models class prototypes as trajectories and learns their evolution with an explicit temporal vector field. By jointly enforcing low-curvature motion and inter-class separation, ProtoFlow stabilizes prototype geometry throughout incremental learning. Experiments on standard class- and domain-incremental remote sensing benchmarks show consistent gains over strong baselines, including up to 1.5-2.0 points improvement in mIoUall, together with reduced forgetting. These results suggest that explicitly modeling temporal prototype evolution is a practical and interpretable strategy for robust continual remote sensing segmentation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ProtoFlow models prototypes as low-curvature trajectories in a temporal vector field to cut forgetting in incremental remote sensing segmentation, but the gains look modest and the experiments need more detail.

read the letter

ProtoFlow models class prototypes as trajectories in a temporal vector field and adds a low-curvature penalty to reduce forgetting during incremental learning on remote sensing data. That's the core idea, and it seems aimed at practical settings where new classes appear and conditions change over time. The paper does a good job framing the problem as continual rather than isolated updates, which matches real-world remote sensing deployments. The explicit temporal modeling is new in this subfield, and the joint enforcement of low curvature and inter-class separation is a reasonable way to stabilize representations. The reported gains of 1.5-2 mIoU points over baselines on standard benchmarks suggest it works at least on the tested data. That said, the details are thin. Without seeing the full experiments, it's hard to tell if the curvature constraint is the driver or if simpler prototype methods would suffice. The concern about domain shifts is valid—if new data requires more curved paths to fit well, the regularization might cap the performance. It would help to see ablations isolating each component and tests on stronger shifts. Overall, this is for researchers in continual learning applied to remote sensing or similar vision tasks. Someone looking for interpretable ways to handle incremental segmentation without heavy replay would find it useful. The thinking seems clear and engaged with the literature on forgetting. I would recommend sending it for peer review. The idea has enough substance to get useful comments, and the authors can strengthen the empirical side in revisions.

Referee Report

3 major / 2 minor

Summary. The paper introduces ProtoFlow, a time-aware prototype dynamics framework for class-incremental remote sensing segmentation. Class prototypes are modeled as trajectories evolving under a learned temporal vector field; low-curvature motion and inter-class separation are jointly enforced to stabilize geometry and mitigate representation drift and forgetting. Experiments on standard class- and domain-incremental remote sensing benchmarks report consistent gains, including 1.5–2.0 mIoUall improvements and reduced forgetting over strong baselines.

Significance. If the central claims hold, the work offers an interpretable, prototype-centric alternative to replay buffers or architectural expansion for continual learning in remote sensing, where sequential data arrival from seasonal/sensor shifts is common. Explicit temporal modeling of prototype evolution could improve robustness in deployment settings and provide a practical strategy for controlling forgetting without additional memory mechanisms.

major comments (3)

[Abstract] Abstract: the reported 1.5–2.0 mIoUall gains and reduced forgetting are stated without any description of the baselines, number of runs, statistical tests, data splits, or ablation studies, preventing verification of the quantitative claims.
[Method] Method (temporal vector field and curvature penalty): the central assumption that low-curvature trajectories plus separation suffice to bound drift under domain shifts is load-bearing but untested; if optimal paths require higher curvature for new distributions, the regularization may cap gains, and no experiment isolates this risk.
[Experiments] Experiments: no ablation isolates the curvature term from inter-class separation or prototype initialization; without this, it is impossible to attribute the reported improvements to the low-curvature flow as claimed.

minor comments (2)

[Method] Notation for the temporal vector field and curvature computation should be defined more explicitly, including how the penalty is applied in prototype space.
[Related Work] Related-work discussion should include recent continual-learning methods specific to remote-sensing segmentation for completeness.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and describe the revisions planned for the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the reported 1.5–2.0 mIoUall gains and reduced forgetting are stated without any description of the baselines, number of runs, statistical tests, data splits, or ablation studies, preventing verification of the quantitative claims.

Authors: We agree that the abstract should provide more context for the reported gains. In the revision we will expand the abstract to name the primary baselines (EWC, LwF, and recent prototype-based continual learning methods), state that results are averaged over five independent runs with standard deviations, reference the standard class- and domain-incremental splits on the ISPRS Vaihingen/Potsdam and related remote-sensing benchmarks, and note that full ablation studies appear in Section 4.3. Statistical significance via paired t-tests is reported in the supplementary material; we will add a concise mention if length permits. revision: yes
Referee: [Method] Method (temporal vector field and curvature penalty): the central assumption that low-curvature trajectories plus separation suffice to bound drift under domain shifts is load-bearing but untested; if optimal paths require higher curvature for new distributions, the regularization may cap gains, and no experiment isolates this risk.

Authors: We acknowledge that the assumption requires stronger isolation. While our domain-incremental experiments already demonstrate improved stability under sensor and seasonal shifts when the curvature penalty is active, we will add a controlled experiment in the revision that varies the curvature regularization weight across simulated distribution shifts, measures resulting prototype drift, and compares final mIoU against an unregularized (higher-curvature) counterpart. This will directly test whether the penalty caps adaptability. revision: yes
Referee: [Experiments] Experiments: no ablation isolates the curvature term from inter-class separation or prototype initialization; without this, it is impossible to attribute the reported improvements to the low-curvature flow as claimed.

Authors: This is a fair criticism. The existing ablations examine the temporal vector field and overall framework but do not fully disentangle the curvature penalty from the separation loss and initialization. In the revised manuscript we will insert a dedicated ablation table that (i) varies only the curvature coefficient while freezing separation and initialization, (ii) ablates separation while keeping curvature fixed, and (iii) compares different prototype initializations under the full objective. This will permit clearer attribution of gains to the low-curvature component. revision: yes

Circularity Check

0 steps flagged

No load-bearing circularity; modeling choice is independent of target metric

full rationale

The paper defines ProtoFlow via an explicit temporal vector field on prototypes plus curvature and separation penalties. This is an ansatz introduced as a design choice rather than derived from or defined in terms of the forgetting metric. No equations reduce the claimed gains to a fitted parameter renamed as prediction, nor does any uniqueness theorem or self-citation chain substitute for the central modeling step. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the method appears to introduce a temporal vector field and curvature penalty as modeling choices whose independence from the target metric cannot be assessed without the full text.

pith-pipeline@v0.9.0 · 5493 in / 1066 out tokens · 37310 ms · 2026-05-13T20:34:05.888684+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel (J-cost uniqueness) echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

We define a parametric vector field F_ϕ : R^d × R → R^d … bμ^{(k+1)}_c = μ^{(k)}_c + Δτ_k F_ϕ(μ^{(k)}_c, τ_k). … L_curve = Σ ||μ^{(t+1)}_c − 2μ^{(t)}_c + μ^{(t-1)}_c||_2^2
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

treat class prototypes as trajectories evolving in a time-indexed feature space … low-curvature motion and inter-class separation

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

EGAD: Entropy-Guided Adaptive Distillation for Token-Level Knowledge Transfer
cs.CL 2026-05 unverdicted novelty 5.0

EGAD adaptively distills LLM knowledge at the token level by using entropy to create a curriculum from low- to high-entropy tokens, adjust temperature, and switch between logits-only and feature-based branches.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · cited by 1 Pith paper

[1]

Large-scale machine learning with stochastic gradient descent, in: Proceedings of COMPSTAT’2010: 19th International Conference on Computational StatisticsParis France, August 22-27, 2010 Keynote, Invited and Contributed Papers, Springer. pp. 177–186. Cermelli, F., Cord, M., Douillard, A.,

work page 2010
[2]

Deepglobe 2018: A challenge to parse the earth through satellite images, in: Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pp. 172–181. Fang, K., Zhang, A., Gao, G., Jiao, J., Liu, C.H., Wei, Y .,

work page 2018
[3]

URL: https://arxiv.org/abs/2507.12857

Score: Scene context matters in open-vocabulary remote sensing instance segmentation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). URL: https://arxiv.org/abs/2507.12857. highlight. Huang, W., Ding, M., Deng, F.,

work page arXiv
[4]

IEEE Transactions on Geoscience and Remote Sensing 62, 1–15

Domain-incremental learning for remote sensing semantic segmentation with multifeature constraints in graph space. IEEE Transactions on Geoscience and Remote Sensing 62, 1–15. doi:doi:10.1109/TGRS.2024.3481875. Li, K., Liu, R., Cao, X., Bai, X., Zhou, F., Meng, D., Wang, Z.,

work page doi:10.1109/tgrs.2024.3481875 2024
[5]

arXiv preprint arXiv:2501.13925 URL:https://arxiv.org/abs/2501.13925

Geopixel: Pixel grounding large multimodal model in remote sensing. arXiv preprint arXiv:2501.13925 URL:https://arxiv.org/abs/2501.13925. Sun, X., Wang, P., Yan, Z., Diao, W., Lu, X., Yang, Z., Zhang, Y ., Xiang, D., Yan, C., Guo, J., et al.,

work page arXiv
[6]

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 8922–8940

Automated high-resolution earth observation image interpretation: Outcome of the 2020 gaofen challenge. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 8922–8940. Sun, X., Weng, X., Pang, C., Xia, G.S.,

work page 2020
[7]

Science China Information Sciences 68, 182301

Mitigating representation bias for class-incremental semantic segmentation of remote sensing images. Science China Information Sciences 68, 182301. doi:doi:10.1007/s11432-024-4307-1. Tan, J., Zhang, H., Yao, N., Yu, Q.,

work page doi:10.1007/s11432-024-4307-1
[8]

arXiv preprint arXiv:2110.08733

Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733 . Wang, T., Chen, G., Zhang, X., Liu, C., Wang, J., Tan, X., Zhou, W., He, C.,

work page arXiv
[9]

Self-training and curriculum learning guided dynamic refined network for remote sensing class-incremental semantic segmentation, in: IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, IEEE. pp. 8334–8338. Zhao, H., Yang, F., Fu, X., Li, X.,

work page 2024
[10]

arXiv preprint arXiv:2503.06683 URL:https://arxiv.org/abs/2503.06683

Dynamic dictionary learning for remote sensing image segmentation. arXiv preprint arXiv:2503.06683 URL:https://arxiv.org/abs/2503.06683. Photo Jiekai Wu is currently a integrated master’s and PhD program at Juntendo University, specializing in Computer Vision and Medical Imaging. He conducts his research under the supervision of Professor Aoki Shigeki. Hi...

work page arXiv
[11]

Photo Chuangqi Li, a technology member with a Computer Science background and deep expertise bridging autonomous driving and AI-driven medical image analysis

His research interests are focused on the application of artificial intelligence. Photo Chuangqi Li, a technology member with a Computer Science background and deep expertise bridging autonomous driving and AI-driven medical image analysis. Currently the CTO of Beijing Qichuang Era Technology Co., Ltd., driving the R&D and commercialization of innovative ...

work page 2021
[12]

His research interests include medical image segmentation and detection, as well as multimodal learning

He is currently working at the National Engineering Research Center for Beijing Biochip Technology (CapitalBio Corporation) focusing on intelligent diagnosis and medical image analysis. His research interests include medical image segmentation and detection, as well as multimodal learning. Photo Dongxu Zhang is currently a master’s student at Juntendo Uni...

work page 2017
[13]

Because of his efforts in Genetic Programming, he also ranked 19th in GP bibliography among more than 12 000 researchers. He has also served as an associate editor, an editor, and a guest editor in several prestigious journals and has delivered several keynote/invited talks Photo Simon James Fong received the B.Eng. (Hons.) and Ph.D. degrees in computer s...

work page 1993
[14]

Following standard practice, original tiles of size 2448×2448 are cut into non-overlapping patches of 512×512 pixels

with 7 semantic classes (urban, agriculture, rangeland, forest, water, barren, and unknown/background). Following standard practice, original tiles of size 2448×2448 are cut into non-overlapping patches of 512×512 pixels. We use the official training split for incremental training and the official validation split for evaluation. For class-incremental seg...

work page 2024
[15]

We follow Huang et al

consists of 33 high-resolution tiles with near-infrared, red and green channels and pixel-wise labels for 6 classes (impervious surfaces, buildings, low vegetation, trees, cars, and clutter/background). We follow Huang et al. (2024); Zou et al. (2025); Sun et al. (2025) for pre-processing and train/validation splits: 16 images are used for training, 5 for...

work page 2024
[16]

(2024), where the base step contains 3 classes and each incremental step adds 1 new foreground class

We use the same 4-step class-incremental protocol as in Huang et al. (2024), where the base step contains 3 classes and each incremental step adds 1 new foreground class. ISPRS Potsdam.The ISPRS Potsdam dataset (Rottensteiner et al.,

work page 2024
[17]

We follow the common protocol in Zou et al

provides 38 tiles with RGB+IR channels and 6 semantic classes. We follow the common protocol in Zou et al. (2025); Huang et al. (2024): 24 images for training, 7 for validation, and 7 for testing. We crop patches of size 512×512 with stride

work page 2025
[18]

We adopt the semantic segmentation version used in Rong et al

is a large-scale aerial image benchmark with high-resolution images and dense annotations. We adopt the semantic segmentation version used in Rong et al. (2022); Zhao et al. (2024), which aggregates instance-level masks into semantic labels. We use the official training/validation split and crop images into 512×512 patches with stride

work page 2022
[19]

We follow Rong et al

consists of large-scale aerial imagery with densely annotated semantic labels. We follow Rong et al. (2022); Zhao et al. (2024) for pre-processing and train/validation splits and adopt the same patch size (512×512 ) and stride (256). The incremental protocol again mirrors that of DeepGlobe and iSAID, with a 4-step schedule and the same relative class prop...

work page 2022
[20]

contains 5987 land-cover images annotated with 7 classes and split into urban and rural domains. To study domain-/temporal shift, we follow the domain-incremental protocol introduced in GSMF-RS-DIL (Huang et al., 2024): (i) step 0 trains on the urban subset, (ii) step 1 trains on the rural subset, and (iii) step 2 revisits a mixed domain containing both (...

work page 2024
[21]

To test this prediction, we perform a class-level differential analysis comparing ProtoFlow with a competitive RS-CISS baseline (MiR (Sun et al., 2025))

predict a stronger statement:when we reduce curvature by regularizing prototype trajectories, we should also reduce forgetting, relative to a strong baseline. To test this prediction, we perform a class-level differential analysis comparing ProtoFlow with a competitive RS-CISS baseline (MiR (Sun et al., 2025)). For each dataset and each semantic classc, w...

work page 2025