Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery
Jowaria Khan, Anindya Sarkar, Yevgeniy Vorobeychik, and Elizabeth Bondi-Kelly
TL;DR
This paper introduces a novel geospatial discovery framework that combines active learning, online meta-learning, and concept-guided reasoning to efficiently identify hidden targets in dynamic environments with limited data.
Contribution
It presents a unified approach leveraging concept relevance for adaptive sampling and meta-learning, improving discovery efficiency in resource-constrained, real-world geospatial tasks.
Findings
Effective in uncovering targets with limited data
Improves generalization in dynamic environments
Demonstrates success on real-world contamination dataset
Abstract
In many real-world settings, such as environmental monitoring, disaster response, or public health, with costly and difficult data collection and dynamic environments, strategically sampling from unobserved regions is essential for efficiently uncovering hidden targets under tight resource constraints. Yet, sparse and biased geospatial ground truth limits 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*, which captures how domain-specific factors influence target presence: a *concept-weighted uncertainty sampling strategy*, where uncertainty is modulated by learned relevance based on readily-available…
Peer Reviews
Decision·ICLR 2026 Conference Desk Rejected Submission
* The Greedy Intersection Algorithm for meta-batch diversity is clever and attempts to maintain semantic coverage without task boundaries. * Ablation studies cover a range of hyperparameters and show some resilience of the method. * The integration of domain-specific concepts into relevance encoding via CVAE is a principled and interpretable modeling choice.
* The experiments use only two datasets. Only one run per configuration is reported with no variance bars. SR gains over some baselines are margin. * Equation (6) uses a handcrafted combination of latent relevance distance and decoder uncertainty. The sampling objective is ad-hoc and insufficiently validated. * The framework assumes domain experts can predefine concept variables that are always available. This is rarely feasible in real domains or multi-modal geospatial datasets. * The SR fo
1. The problem of Open-World Learning for Geospatial Prediction and Sampling (OWL-GPS) is important and relevant to many real-world applications. 2. The method design addresses important challenges in the OWL-GPS problem, including strict sampling budgets and evolving data distributions, and non-revisitable inputs. 3. The experimental results show the improvements over the baselines.
1. The contribution of the components are not well explained. Is the relevance uncertainty from domain-specific concepts like spectral channels trying to use selected features based on domain knowledge to do the sampling? This feels like a reasonable domain application choice but the novelty is not apparent. 2. Although the paper provides additional details on the datasets in the supplementary, it would be helpful to clarify how the characteristics of each dataset correspond to the OWL-GPS pro
1. To the best of my knowledge, the problem formulation is novel. Also to the best of my knowledge, the problem formulation captures real-world challenges for online data collection with limited budgets, seen in many environmental monitoring settings. 2. This paper has several contributions: a novel problem formulation, a novel method to address the problem, and a new benchmark for this problem. 3. The proposed method performs better on average than the baselines (none of which are designed for
1. There are so many design choices in the proposed methods, but not enough space to explain them, and not enough explain experiments to undertsand what matters. There are ablations on the relecance encoder and meta-training set, but only for one dataset (2019) - why not all three datasets? Does the GS orthogonalization matter? Forgive me if I missed it but I don't think that's tested. 2. Related work is sparse: the part on geospatial foundation models only cites 2 papers? There are a ton. Eithe
The paper tackles a timely and challenging problem of online adaptation for geospatial tasks, proposing a unified framework that integrates concept encoding and meta-learning.
- Regarding interpretability, the study would benefit from visualizing which concepts or POI dimensions are most influential, along with strategy heatmaps or trajectory plots explaining the model’s sampling behavior. Qualitative analysis of failure cases could clarify which factors or uncertainty scores led to misdetections. - Regarding experiments, significance testing is missing; an ablation that removes the Gram–Schmidt orthogonalization step is needed to validate the contribution of concept
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsDomain Adaptation and Few-Shot Learning · Machine Learning and Data Classification · Machine Learning and Algorithms