Dynamic Class-Aware Active Learning for Unbiased Satellite Image Segmentation

Gadi Hemanth Kumar; Athira Nambiar; Pankaj Bodani

cs.CVApr 10, 2026

Dynamic Class-Aware Active Learning for Unbiased Satellite Image Segmentation

Authors: Gadi Hemanth Kumar, Athira Nambiar, Pankaj Bodani

Organizations: Department of Computational Intelligence, Faculty of Engineering and Technology, SRM Institute of Science and Technology, Kattankulathur 603203, Tamil Nadu, India · Space Applications Centre, Indian Space Research Organisation, Ahmedabad, India

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.

Figures & tables

Figure 1: Proposed Dynamic Class Aware Uncertainty based Active Learning (DCAU-AL) framework for semantic segmentation with a human-in-the-loop.

Method	mIoU	Background	Bareland	Rangeland	Developed Space	Road	Tree	Water	Agriculture Land	Buildings
Fully Supervised	0.670	0.703	0.217	0.694	0.638	0.603	0.786	0.858	0.704	0.749
Naive AL	0.642	0.719	0.201	0.665	0.589	0.547	0.770	0.829	0.651	0.687
DCAU-Based AL	0.664	0.833	0.248	0.661	0.604	0.582	0.758	0.834	0.664	0.714

Table 1: Comparison of mIoU and class-wise IoU scores for Fully Supervised, Naive AL, and DCAU-Based AL models. DCAU achieves near-Fully Supervised performance using only 40% labeled data and delivers significant gains in minority classes.

Cycle	mIoU	Background	Bareland	Rangeland	Developed Space	Road	Tree	Water	Agriculture Land	Building
AL-1	0.584	0.703	0.151	0.636	0.553	0.489	0.746	0.803	0.617	0.559
AL-5	0.604	0.803	0.168	0.621	0.549	0.492	0.722	0.805	0.570	0.673
AL-10	0.631	0.828	0.171	0.643	0.563	0.542	0.741	0.824	0.605	0.676
AL-15	0.652	0.831	0.228	0.667	0.601	0.563	0.732	0.820	0.669	0.695
AL-20	0.664	0.833	0.248	0.661	0.604	0.579	0.758	0.834	0.664	0.714

Table 2: DCAU-AL model mIoU and class-wise IoU scores across AL cycles on the OpenEarth dataset starting with 500 labeled samples.

Figure 2: Progression of mIoU scores across acquisition cycles for Fully Supervised, Naive AL, and DCAU-Based AL models.(Better viewed in colour)

Panel of Figure 3 — Figure 3: Segmentation results for Fully Supervised, Naive-AL, and DCAU-AL models across AL cycles, with corresponding mIoU values in parentheses.(Better Viewed in Colour)

Configuration	Learning Rate	Samples per AL Cycle	mIoU (%)
DCAU-Baseline	$1\times 10^{-3}$	50	0.654
DCAU-Baseline	$1\times 10^{-4}$	50	0.664
DCAU-Extended	$1\times 10^{-3}$	100	0.632
DCAU-Extended	$1\times 10^{-4}$	100	0.648

Table 3: Ablation study results showing mIoU performance of DCAU-AL under different annotation budgets and learning rates

Model	mIoU Score	Training Time (hrs)
Fully Supervised Model	0.670	30 hr
Naive Active Learning Model [ 11 ]	0.642	35 hr
Coreset Model [ 12 ]	0.648	38 hr
DCAU-AL (Ours)	0.664	36 hr

Table 4: Segmentation Accuracy and Training Time across different models vs Proposed DCAU-AL

Active Learning

Active learning selects the examples, labels, features, or experiments most useful for improving a model under limited budgets.

See all 288 papers

Label Less, Learn More: Resource-Efficient Active Semi-Supervised Learning for Onboard Satellite Image Annotation
Sep 25, 2026Ahmed Abdelnaby, Mohamed Elmahallawy, Marius Bernahrndt +1Data-Efficient Learning Semi-Supervised Learning
School of Engineering and Applied Science, Washington State University, Richland, WA 99354, USA · Institute for Software Technology, German Aerospace Center (DLR), Porz, 51147 Cologne, Germany
Active Learning under Simulated Class Imbalance and Label Shift in Binary Semantic Segmentation
Jan 8, 2026Julien Combes, Jean-François Coeurjolly, Alexandre DervilleActive Learning Label Shift
SVH, Michelin · SVH · Michelin
Exploring Entropy-based Active Learning for Fair Brain Segmentation
May 3, 2026Ghazal Danaee, Mélanie Gaillochet, Christian Desrosiers +2Active Learning Group Fairness
École de technologie supérieure, Montréal, QC, Canada · Polytechnique Montréal, Montréal, QC, Canada · Mila - Quebec AI Institute, Montréal, QC, Canada

Sep 25, 2026J/K move · Enter open · S save

Label Less, Learn More: Resource-Efficient Active Semi-Supervised Learning for Onboard Satellite Image Annotation

Ahmed Abdelnaby, Mohamed Elmahallawy, Marius Bernahrndt, Tobias Hecking

School of Engineering and Applied Science, Washington State University, Richland, WA 99354, USA · Institute for Software Technology, German Aerospace Center (DLR), Porz, 51147 Cologne, Germany

Data-Efficient Learning Semi-Supervised Learning Remote Sensing Image Understanding Active Learning

Open paper AskPDF arXiv

Large-scale pervasive sensing increasingly relies on high-resolution satellite imagery, yet task-specific onboard vision is constrained by costly annotation and limited computation, memory, energy, and communication resources. Existing approaches largely rely on either data-hungry supervised learning or large vision-language foundation models, limiting efficient adaptation and deployment under these constraints. We present SatLabel, a resource-aware learning framework that transforms limited satellite labels into progressively refined onboard models through adaptive sample acquisition and semi-supervised model adaptation. Rather than repeatedly training on uniformly sampled labels, SatLabel closes the loop between model uncertainty, class imbalance, and pseudo-label quality to selectively acquire informative samples while exploiting abundant unlabeled imagery. This enables a compact student to adapt to target sensing domains with reduced annotation and inference costs. We further introduce an optional Mixture-of-Experts (MoE) student with graph-based feature refinement to enhance representation capacity while retaining a lightweight footprint. We evaluate SatLabel on 11 remote-sensing datasets spanning core, extended, and unseen domains against RemoteCLIP zero-shot inference. SatLabel improves Macro-F1 on most core and extended datasets while maintaining strong cross-dataset transfer to unseen domains. More importantly, the Balanced student contains only 11.2 M parameters and occupies approximately 42.8 MB, compared with 151.3M parameters and 577 MB for RemoteCLIP, while requiring 3.65 versus 5.89 GFLOPs. Across four efficiency benchmarks, it achieves approximately 2x higher GPU-forward throughput and reduces energy per image on datasets.

Abstract

Active Learning

Explore similar work

Label Less, Learn More: Resource-Efficient Active Semi-Supervised Learning for Onboard Satellite Image Annotation

Active Learning under Simulated Class Imbalance and Label Shift in Binary Semantic Segmentation

Exploring Entropy-based Active Learning for Fair Brain Segmentation