arxiv: 2604.08965 · v1 · submitted 2026-04-10 · 💻 cs.CV

Recognition: unknown

Dynamic Class-Aware Active Learning for Unbiased Satellite Image Segmentation

Gadi Hemanth Kumar , Athira Nambiar , Pankaj Bodani

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:51 UTC · model grok-4.3

classification 💻 cs.CV

keywords active learningsemantic segmentationsatellite imageryclass imbalanceland coveruncertainty samplingannotation efficiency

0 comments

The pith

A dynamic class-aware active learning method reduces bias in satellite image segmentation by tracking per-class performance in real time.

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

The paper proposes a new active learning strategy for semantic segmentation of satellite imagery that adapts sample selection based on how well the model is performing on each land cover class. Instead of using fixed global uncertainty measures, it monitors class-wise metrics and boosts the selection of samples likely to help struggling classes. This addresses the common problem of class imbalance in large satellite datasets, where dominant classes overshadow rare ones during training. A sympathetic reader would care because it promises higher accuracy across all classes and lower annotation costs for applications like land cover mapping and environmental monitoring.

Core claim

The authors claim that their Dynamic Class-Aware Uncertainty based Active Learning (DCAU-AL) overcomes limitations of standard active learning by continuously tracking segmentation performance per class and dynamically adjusting sampling weights to prioritize poorly performing or underrepresented classes. This adaptive mechanism leads to superior per-class IoU and improved annotation efficiency, as demonstrated on the OpenEarth land cover dataset under conditions of severe class imbalance.

What carries the argument

DCAU-AL acquisition function that uses real-time class-wise performance gaps to adjust sampling probabilities dynamically.

If this is right

The proposed method outperforms existing active learning approaches in per-class accuracy on imbalanced satellite data.
It achieves better overall segmentation quality with fewer annotated samples.
Continuous adjustment prevents the model from ignoring rare classes as training progresses.
Annotation efficiency improves particularly when class distributions are highly skewed.

Where Pith is reading between the lines

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

Similar dynamic weighting could help in other domains with imbalanced data, such as medical image segmentation.
Combining this with diversity-based selection might further reduce bias in very large datasets.
Testing on datasets with different imbalance levels could reveal the method's robustness limits.

Load-bearing premise

That class-wise performance can be measured accurately enough in real time to guide sample selection without creating new biases or excessive computation.

What would settle it

Running the method on a held-out satellite dataset with known severe imbalance and finding no improvement in per-class IoU compared to standard uncertainty sampling would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.08965 by Athira Nambiar, Gadi Hemanth Kumar, Pankaj Bodani.

**Figure 2.** Figure 2: Progression of mIoU scores across acquisition cycles for Fully Supervised, [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Figure 3: Segmentation results for Fully Supervised, Naive-AL, and DCAU-AL [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

read the original abstract

Semantic segmentation of satellite imagery plays a vital role in land cover mapping and environmental monitoring. However, annotating large-scale, high-resolution satellite datasets is costly and time consuming, especially when covering vast geographic regions. Instead of randomly labeling data or exhaustively annotating entire datasets, Active Learning (AL) offers an efficient alternative by intelligently selecting the most informative samples for annotation with the help of Human-in-the-loop (HITL), thereby reducing labeling costs while maintaining high model performance. AL is particularly beneficial for large-scale or resource-constrained satellite applications, as it enables high segmentation accuracy with significantly fewer labeled samples. Despite these advantages, standard AL strategies typically rely on global uncertainty or diversity measures and lack the adaptability to target underperforming or rare classes as training progresses, leading to bias in the system. To overcome these limitations, we propose a novel adaptive acquisition function, Dynamic Class-Aware Uncertainty based Active learning (DCAU-AL) that prioritizes sample selection based on real-time class-wise performance gaps, thereby overcoming class-imbalance issue. The proposed DCAU-AL mechanism continuously tracks the performance of the segmentation per class and dynamically adjusts the sampling weights to focus on poorly performing or underrepresented classes throughout the active learning process. Extensive experiments on the OpenEarth land cover dataset show that DCAU-AL significantly outperforms existing AL methods, especially under severe class imbalance, delivering superior per-class IoU and improved annotation efficiency.

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

The paper sketches a plausible class-aware dynamic weighting scheme for active learning in imbalanced satellite segmentation but supplies zero experimental evidence for its performance claims.

read the letter

The core idea is straightforward: instead of using a single global uncertainty score, DCAU-AL tracks per-class IoU or similar metrics during training and boosts the sampling probability for classes that are lagging. This targets the common remote-sensing problem where dominant classes like vegetation swamp out rare ones like specific urban features or water bodies. The approach is presented as a lightweight heuristic that updates weights in real time without new model architectures, which keeps the implementation burden low for practitioners already running segmentation models on large tile sets.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes Dynamic Class-Aware Uncertainty based Active Learning (DCAU-AL), a novel acquisition function for active learning in semantic segmentation of satellite imagery. It claims to address class imbalance by continuously tracking per-class segmentation performance and dynamically adjusting sampling weights to prioritize poorly performing or underrepresented classes. Experiments on the OpenEarth land cover dataset are asserted to show significant outperformance over existing AL methods, with improved per-class IoU and annotation efficiency.

Significance. If the empirical claims can be substantiated with detailed experiments, the method could provide a targeted improvement for active learning under severe class imbalance in remote sensing, potentially enhancing model fairness and reducing annotation costs in large-scale land cover mapping applications.

major comments (2)

[Abstract] Abstract: The central empirical claim that DCAU-AL 'significantly outperforms existing AL methods, especially under severe class imbalance, delivering superior per-class IoU and improved annotation efficiency' is stated without any experimental details, baseline methods, evaluation metrics, result tables, or statistical tests, leaving the primary contribution unsupported.
[Method description] Method: The DCAU-AL mechanism is described at a high level as tracking class-wise performance gaps and adjusting sampling weights, but no equations, formal definition of the acquisition function, algorithm, or pseudocode is supplied to specify how real-time class performance is measured or how weights are dynamically computed.

minor comments (1)

[Abstract] Abstract: The expansion of the acronym DCAU-AL contains minor inconsistencies in wording ('Uncertainty based' vs. title phrasing) that could be clarified for precision.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below and outline the revisions we will incorporate to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: The central empirical claim that DCAU-AL 'significantly outperforms existing AL methods, especially under severe class imbalance, delivering superior per-class IoU and improved annotation efficiency' is stated without any experimental details, baseline methods, evaluation metrics, result tables, or statistical tests, leaving the primary contribution unsupported.

Authors: We agree that the abstract, as a concise summary, does not include the specific experimental details, baselines, metrics, or statistical tests. These are provided in full in the Experiments and Results sections of the manuscript, including comparisons against standard AL methods on the OpenEarth dataset with per-class IoU metrics and efficiency analysis. To better support the claim at the abstract level, we will revise the abstract to briefly reference the key baselines (e.g., uncertainty sampling, core-set selection) and primary metrics (per-class IoU, annotation budget). revision: partial
Referee: [Method description] Method: The DCAU-AL mechanism is described at a high level as tracking class-wise performance gaps and adjusting sampling weights, but no equations, formal definition of the acquisition function, algorithm, or pseudocode is supplied to specify how real-time class performance is measured or how weights are dynamically computed.

Authors: This observation is correct; the initial submission presented the DCAU-AL mechanism at a conceptual level without the required formalization. We will add the mathematical formulation of the class-wise performance gap (defined as the difference between current per-class IoU and a target threshold), the dynamic weighting function, the overall acquisition score, and a complete algorithm pseudocode in the revised Method section. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper introduces DCAU-AL as a novel adaptive acquisition function that tracks per-class performance gaps in real time and adjusts sampling weights accordingly. No equations, derivations, or self-citations are provided in the available text that reduce the proposed mechanism to fitted inputs, self-definitions, or prior author results by construction. The central claim rests on an empirical comparison on the OpenEarth dataset rather than any load-bearing mathematical reduction or renamed known pattern. The method is presented as a new heuristic without internal self-reference that would force the outcome.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on the domain assumption that class-wise performance can be tracked in real time during AL and that focusing on underperforming classes improves overall segmentation without side effects; no free parameters or invented physical entities are specified.

axioms (2)

domain assumption Active learning reduces labeling costs while maintaining high model performance compared to random sampling.
Invoked when stating AL benefits for large-scale satellite applications.
domain assumption Standard AL strategies lack adaptability to target underperforming or rare classes, leading to bias.
Stated as the key limitation that DCAU-AL overcomes.

invented entities (1)

DCAU-AL acquisition function no independent evidence
purpose: Dynamically prioritizes samples based on real-time class-wise performance gaps
New mechanism introduced to address class imbalance in AL for segmentation.

pith-pipeline@v0.9.0 · 5553 in / 1509 out tokens · 51220 ms · 2026-05-10T16:51:49.281453+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 2 canonical work pages

[1]

In: Proceedings of the European Conference on Computer Vision (ECCV)

Agarwal, S., Arora, H., Anand, S., Arora, C.: Contextual diversity for active learn- ing. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 137–153. Springer (2020)

2020
[2]

In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

Aggarwal, U., Popescu, A., Hudelot, C.: Active learning for imbalanced datasets. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1410–1419. IEEE (2020)

2020
[3]

In: Proceedings of the ACM International Con- ference on AI in Finance

Barata, R., Santana, A., Costa, H., Martins, P., Cardoso, J.: Active learning for imbalanced data under cold start. In: Proceedings of the ACM International Con- ference on AI in Finance. pp. 1–9. ACM (2021)

2021
[4]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Cui, Y., Song, Y., Sun, C., Howard, A., Belongie, S.: Large scale fine-grained categorization and domain-specific transfer learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4109–4118 (2018)

2018
[5]

In: Proceedings of the International Conference on Machine Learning (ICML)

Gal, Y., Islam, R., Ghahramani, Z.: Deep bayesian active learning with image data. In: Proceedings of the International Conference on Machine Learning (ICML). PMLR (2017)

2017
[6]

IEEE Transactions on Knowl- edge and Data Engineering21(9), 1263–1284 (2009)

He, H., Garcia, E.: Learning from imbalanced data. IEEE Transactions on Knowl- edge and Data Engineering21(9), 1263–1284 (2009)

2009
[7]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Huang, C., Li, Y., Loy, C., Tang, X.: Learning deep representation for imbalanced classification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5375–5384 (2016)

2016
[8]

In: Pro- ceedings of the International Conference on Artificial Intelligence

Japkowicz, N.: The class imbalance problem: Significance and strategies. In: Pro- ceedings of the International Conference on Artificial Intelligence. vol. 56, pp. 111– 117 (2000)

2000
[9]

IEEE Journal of Selected Topics in Applied Earth Observations and Re- mote Sensing15, 3376–3389 (2022)

Lenczner, G., Lee, H., Rodriguez, A., Yao, A., Paheding, S., Zhu, Q., Kwon, H.: Dial: Deep interactive and active learning for semantic segmentation in remote sensing. IEEE Journal of Selected Topics in Applied Earth Observations and Re- mote Sensing15, 3376–3389 (2022)

2022
[10]

Medical Image Analysis88, 102862 (2023)

Li, X., Xia, M., Jiao, J., Zhou, S., Chang, C., Wang, Y., Guo, Y.: Hal-ia: A hybrid active learning framework using interactive annotation for medical image segmen- tation. Medical Image Analysis88, 102862 (2023)

2023
[11]

In: Proceedings of the International Confer- ence on Pattern Recognition (ICPR)

Mandalika, S., Nambiar, A.: Segxal: Explainable active learning for semantic seg- mentation in driving scene scenarios. In: Proceedings of the International Confer- ence on Pattern Recognition (ICPR). Springer (2024) DCAU: Active Learning for Imbalanced Satellite Segmentation 15

2024
[12]

International Journal of Computer Vision133(7), 4294– 4316 (2025)

Mittal, S.: Realistic evaluation of deep active learning for image classification and semantic segmentation. International Journal of Computer Vision133(7), 4294– 4316 (2025)

2025
[13]

In: Proceedings of the International Conference on Machine Learning (ICML)

Ren, M., Zeng, W., Yang, B., Urtasun, R.: Learning to reweight examples for robust deep learning. In: Proceedings of the International Conference on Machine Learning (ICML). pp. 4334–4343 (2018)

2018
[14]

ACM Computing Surveys54(9), 1–40 (2021)

Ren, Y., Xiao, X., Chang, X., Huang, P., Li, Z., Gupta, B., Chen, X., Wang, X.: A survey of deep active learning. ACM Computing Surveys54(9), 1–40 (2021)

2021
[15]

In: Medical Image Computing and Computer-Assisted Intervention (MICCAI)

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI). Lecture Notes in Computer Science, vol. 9351, pp. 234–
[16]

Computers & Geosciences192, 105704 (2024)

Scala, P., Manno, G., Ciraolo, G.: Semantic segmentation of coastal aerial/satellite images using deep learning techniques. Computers & Geosciences192, 105704 (2024)

2024
[17]

Active Learning for Convolutional Neural Networks: A Core-Set Approach

Sener, O., Savarese, S.: Active learning for convolutional neural networks: A coreset approach. In: arXiv preprint arXiv:1708.00489 (2017)

work page Pith review arXiv 2017
[18]

IEEE Transactions on Geoscience and Remote Sensing (2023)

Shan, L., Wang, W., Lv, K., Luo, B.: Boosting semantic segmentation of aerial im- ages via decoupled and multi-level compaction and dispersion. IEEE Transactions on Geoscience and Remote Sensing (2023)

2023
[19]

In: Proceedings of the 31st ACM International Conference on Multimedia (ACM MM)

Shan, L., Zhou, W., Zhao, G.: Incremental few-shot semantic segmentation via class-agnostic mask proposal and language-driven classifier. In: Proceedings of the 31st ACM International Conference on Multimedia (ACM MM). pp. 8561–8570. ACM (2023)

2023
[20]

arXiv preprint arXiv:2405.18078 (2024)

Shan, L., et al.: Edge-guided and class-balanced active learning for semantic seg- mentation of aerial images. arXiv preprint arXiv:2405.18078 (2024)

work page arXiv 2024
[21]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Siddiqui, Y., Valentin, J., Nießner, M.: Viewal: Active learning with viewpoint entropy for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9433–9443 (2020)

2020
[22]

In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2023)

Xia, J., et al.: Openearthmap: A benchmark dataset for global high-resolution land cover mapping. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2023)

2023
[23]

In: Proceedings of the Asian Conference on Computer Vision (ACCV) (2020)

Xie, S., Feng, Z., Chen, Y., Sun, S., Ma, C., Song, M.: Deal: Difficulty-aware active learning for semantic segmentation. In: Proceedings of the Asian Conference on Computer Vision (ACCV) (2020)

2020
[24]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Yang, F., He, Z., Zhang, Z., Xu, Y., Liang, X., Liang, Y.: Plug and play active learning for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22954–22963 (2024)

2024
[25]

International Journal of Computer Vision113(2), 113–127 (2015)

Yang, Y., Ma, Z., Nie, F., Chang, X., Hauptmann, A.: Multi-class active learn- ing by uncertainty sampling with diversity maximization. International Journal of Computer Vision113(2), 113–127 (2015)

2015