Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery

Jowaria Khan; Anindya Sarkar; Yevgeniy Vorobeychik; Elizabeth Bondi-Kelly

cs.CVFeb 19, 2026

Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery

Authors: Jowaria Khan, Anindya Sarkar, Yevgeniy Vorobeychik, Elizabeth Bondi-Kelly

Organizations: University of Michigan, Ann Arbor, MI, USA · Washington University in St. Louis, St. Louis, MO, USA

Abstract

In environmental monitoring, data collection is often costly, sparse, and shaped by urgent public-health needs. This is particularly true for cancer-causing PFAS (Per- and polyfluoroalkyl substances) contamination, where discussions with domain experts and environmental organizations highlight the need to strategically identify high-risk, under-observed regions under tight sampling budgets. More broadly, similar challenges arise in disaster response and public health settings, where dynamic environments make it essential to efficiently uncover hidden targets from limited ground truth. Yet sparse and biased geospatial labels limit the applicability of existing learning-based methods, such as reinforcement learning. To address this, we propose a unified geospatial discovery framework that integrates active learning, online meta-learning, and concept-guided reasoning. Our approach introduces two key innovations built on a shared notion of concept relevance, capturing how domain-specific factors influence target presence: a concept-weighted uncertainty sampling strategy, where uncertainty is modulated by learned relevance from readily available concepts such as land cover and source proximity; and a relevance-aware meta-batch formation strategy that promotes semantic diversity during online-meta updates, improving generalization in dynamic environments. We evaluate our framework on PFAS contamination discovery as a real-world inspired environmental monitoring task, demonstrating robust target discovery under limited data and changing conditions.

Figures & tables

Figure 1: Learning under limited sampling budget: selecting the most informative locations.

Baseline	Description
GA	G reedy A pproach: A non-adaptive, single-pass method that queries the highest-confidence regions after supervised training.
DIG	D omain- I nformed G reedy: A single-pass baseline that leverages domain-informed simulation to estimate contamination and rank regions without learning or sequential adaptation.
Prithvi	A standard supervised geospatial model that iteratively selects high-confidence samples with online updates during inference. ( Blumenfeld, 2023 ) .
UCB	OWL-GPS with a bandit-based exploration and exploitation sampling policy using U pper C onfidence B ounds ( Auer et al., 2002 ) .
AL	A ctive L earning: Exploration-focused, pure uncertainty based sampling during training and inference ( Cacciarelli and Kulahci, 2023 ) .
AML	A ctive M eta L earning: OWL-GPS with meta-batch updates driven by latent representations of previously queried samples ( Kaddour et al., 2020 ) .

Table 1: Summary of baseline methods used for comparison.

Year	Method	Acc.	F-score	Prec.	Recall	SR
LC	GA	46 $\pm$ 0.8	35 $\pm$ 1.0	48 $\pm$ 0.9	40 $\pm$ 1.3	47 $\pm$ 1.4
	AL	58 $\pm$ 1.8	41 $\pm$ 1.5	50 $\pm$ 1.2	47 $\pm$ 1.9	60 $\pm$ 1.0
	Prithvi	61 $\pm$ 1.2	40 $\pm$ 1.4	51 $\pm$ 1.5	56 $\pm$ 1.7	61 $\pm$ 1.8
	UCB	55 $\pm$ 2.1	45 $\pm$ 1.7	53 $\pm$ 1.3	58 $\pm$ 1.5	55 $\pm$ 1.3
	AML	91 $\pm$ 1.3	48 $\pm$ 1.3	46 $\pm$ 1.6	50 $\pm$ 1.8	90 $\pm$ 1.1
	OML	92 $\pm$ 1.7	48 $\pm$ 1.6	47 $\pm$ 1.8	50 $\pm$ 1.9	92 $\pm$ 1.5

Table 2: Comparison with baselines (mean

\pm

std. over 3 trials). DIG is a deterministic domain-informed static baseline for PFAS only.

Panel of Figure 4 — Figure 4: Performance and exploration behavior on 2019 PFAS data.

Baseline	Key Limitations in OWL-GPS setting
UCB	Assumes independent arms; fails under spatial correlations, leading to inefficient exploration.
AL	Uncertainty-based sampling without relevance or feedback-driven correction; often selects informative samples yet fails to discover target-rich regions.
DIG	Domain-informed approach based on pollutant transport simulation; lacks adaptation to newly observed samples, limiting sequential performance.
GA	Exploits high-confidence regions early with no exploration; resulting in suboptimal performance.
Prithvi	Statically trained with online updates at inference, but not designed for strategic exploration.
AML	Relies on task identity and few-shot setup; fails in geospatial streams without task boundaries.

Table 3: Baseline weaknesses in the OWL-GPS setting.

Variant	Accuracy	F-score	Precision	Recall	SR ( $\mathcal{C}=50$ )
No Relevance Encoder (No-RE)	86 $\pm$ 1.8	56 $\pm$ 1.7	50 $\pm$ 1.5	74 $\pm$ 1.8	95 $\pm$ 2.0
Random Meta-Training Set	67 $\pm$ 2.0	56 $\pm$ 2.1	57 $\pm$ 1.6	46 $\pm$ 1.7	69 $\pm$ 1.4
No Relevance-Guided Sampling (No-RG)	72 $\pm$ 1.9	55 $\pm$ 1.7	55 $\pm$ 2.0	50 $\pm$ 1.7	77 $\pm$ 1.5
Ours (Full Framework)	88 $\pm$ 2.0	65 $\pm$ 2.0	57 $\pm$ 2.0	83 $\pm$ 2.0	98 $\pm$ 1.0

Table 4: Ablation Study: Importance of Method Components.

Panel of Figure 5 — Figure 5: Interpreting decision-making through concept relevance.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

SR Performance across varying $\alpha$ with $\mathcal{C}=50$
$\alpha=0.2$	$\alpha=1.0$	$\alpha=5.0$
94% $\pm$ 1.0	98% $\pm$ 1.0	98% $\pm$ 1.9

Appendix

Table 5: Effect of

\kappa(\mathcal{C})

Search Budget ( $\mathcal{C}$ )	Accuracy (%)	F-score (%)	SR (%)
10	80 $\pm$ 1.5	50 $\pm$ 1.1	43 $\pm$ 1.3
30	81 $\pm$ 1.3	56 $\pm$ 1.2	58 $\pm$ 1.6
50	88 $\pm$ 2.0	65 $\pm$ 2.0	98 $\pm$ 1.0

Appendix

Table 6: Search performance across varying search budgets

\mathcal{C}

Buffer configuration	Accuracy (%)	F1-score (%)	SR (%)
Core $>$ Reservoir	77 $\pm$ 1.9	50 $\pm$ 1.1	88 $\pm$ 0.5
Core = Reservoir	87 $\pm$ 2.2	56 $\pm$ 1.0	98 $\pm$ 1.1
Core $<$ Reservoir (Ours)	88 $\pm$ 2.0	65 $\pm$ 2.0	98 $\pm$ 1.0

Appendix

Table 7: Search performance across varying buffer configurations.

Clustering approach	Accuracy (%)	F-score (%)	SR (%) ( $\mathcal{C}=50$ )
DBSCAN	81 $\pm$ 2.2	57 $\pm$ 1.6	94 $\pm$ 1.1
Ours	88 $\pm$ 2.0	65 $\pm$ 2.0	98 $\pm$ 1.0

Appendix

Table 8: Effectiveness of the clustering approach.

Method	Accuracy (%)	F-score (%)	SR (%)
Ours (RG $\rightarrow$ MLP)	85 $\pm$ 1.7	56 $\pm$ 1.8	94 $\pm$ 2.3
Ours	88 $\pm$ 2.0	65 $\pm$ 2.0	98 $\pm$ 1.0

Appendix

Table 9: Analyzing the impact of relevance-guided sampling.

Orthogonalization	Accuracy (%)	F-score (%)	SR (%) ( $\mathcal{C}=50$ )
No	80 $\pm$ 1.0	56 $\pm$ 1.5	86 $\pm$ 0.7
Yes (our framework)	88 $\pm$ 2.0	65 $\pm$ 2.0	98 $\pm$ 1.0

Appendix

Table 10: Importance of concept orthogonalization.

Panel of Figure 10 — Figure 10: PFAS dataset splits. Green denotes training samples and purple denotes test samples. Splits are geographically disjoint; apparent overlap arises from visualization at this scale.

Figure 12: Example binary output from the pollutant transport simulation: 0 represents low contamination, and 1 represents high contamination. The simulation approximates the distribution of PFAS contamination based on hydrological and environmental parameters, providing a practical alternative to full-scale SWAT simulations.

Adaptive Sampling

Adaptive sampling allocates limited data, measurements, or computation where they are expected to be most useful.

See all 202 papers

DT-GOL: Dual-Track Geometric Online Learning in Nonstationary Environment with Label Delay
Jun 22, 2026Yulin Wang, Yi He, Dianlong You +1Concept Drift
College of Computer and Information Science, Southwest University Chongqing, China · Department of Data Science, College of William and Mary Williamsburg, VA, USA · School of Artificial Intelligence (School of Software), Yanshan University Qinhuangdao, Hebei, China
Towards Intelligent Geospatial Data Discovery: a knowledge graph-driven multi-agent framework powered by large language models
Mar 21, 2026Ruixiang Liu, Zhenlong Li, Ali Khosravi KazaziGraph-Based Retrieval LLM Agents
Geoinformation and Big Data Research Laboratory, Department of Geography, The Pennsylvania State University, University Park, PA, USA
DynCur-Geo: Dynamic Curiosity Reward Shaping for Multimodal Active Geo-Localization
Aug 19, 2026Yiming Sun, Yang Zhang, Pengfei ZhuUAV Navigation Reward Shaping
School of Automation, Southeast University, Nanjing 210096, China

Jun 22, 2026J/K move · Enter open · S save

DT-GOL: Dual-Track Geometric Online Learning in Nonstationary Environment with Label Delay

Yulin Wang, Yi He, Dianlong You, Di Wu

College of Computer and Information Science, Southwest University Chongqing, China · Department of Data Science, College of William and Mary Williamsburg, VA, USA · School of Artificial Intelligence (School of Software), Yanshan University Qinhuangdao, Hebei, China

Concept Drift

Open paper AskPDF arXiv

Online learning is crucial for handling complex data streams in big data applications. Recent research has begun to focus on dynamic scenarios, i.e., non-stationary environments. However, a crucial yet often overlooked aspect is label latency, where new data may not receive labels in time due to the slow and expensive labeling process, thus hindering rapid adaptation to dynamic environments. To resolve this impasse, we propose Dual-Track Geometry Online Learning (DT-GOL), a novel framework that shifts from temporal compensation to spatial reasoning to bridge the supervised latency gap. By modeling the delay challenge as a semi-supervised task, we leverage real-time topological evolution of features as a reliable geometric surrogate for unobservable conceptual changes to achieve proactive supervised adaptation within the delay window. Unlike rigid self-training, we introduce a dynamic evidence calibration mechanism that distills geometric information into soft labels that perceive uncertainty, effectively mitigating the confirmation bias inherent in hard pseudo-labels. Furthermore, to resolve the stability-plasticity dilemma, we design a decoupled dual-track architecture in which a master learner serves as a stable anchor, updated strictly from delayed ground truth, while a transient branch leverages soft geometric knowledge for low-risk forward adaptation. Extensive experiments on real and synthetic datasets demonstrate that DT-GOL significantly outperforms existing state-of-the-art baseline methods, especially in scenarios with concept drift.

Abstract

Adaptive Sampling

Explore similar work

DT-GOL: Dual-Track Geometric Online Learning in Nonstationary Environment with Label Delay

Towards Intelligent Geospatial Data Discovery: a knowledge graph-driven multi-agent framework powered by large language models

DynCur-Geo: Dynamic Curiosity Reward Shaping for Multimodal Active Geo-Localization