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

Source: arXiv - 2602.17605v1

Overview

The paper presents a new framework for geospatial discovery that can actively and adaptively sample locations to uncover hidden targets (e.g., environmental contaminants) while operating under tight resource constraints. By marrying active learning, online meta‑learning, and concept‑guided reasoning, the authors show how to make smarter sampling decisions even when ground‑truth data are sparse and biased.

Key Contributions

• Concept‑Weighted Uncertainty Sampling – a novel active‑learning query strategy that scales classic uncertainty sampling with concept relevance derived from readily available domain attributes (land‑cover type, proximity to known sources, etc.).
• Relevance‑Aware Meta‑Batch Formation – an online meta‑learning routine that builds training batches emphasizing semantic diversity, improving the model’s ability to generalize across shifting geospatial distributions.
• Unified Geospatial Discovery Pipeline – integrates the above components into a single end‑to‑end system that can be deployed in dynamic, data‑scarce environments.
• Real‑World Validation on PFAS Contamination – demonstrates the approach on a large‑scale dataset of per‑ and polyfluoroalkyl substances (PFAS) contamination, achieving higher detection rates than baseline active‑learning and reinforcement‑learning methods.

Methodology

1. Concept Representation – Each candidate location is described by a vector of concepts (e.g., land‑use class, distance to industrial sites). A lightweight relevance model learns how strongly each concept correlates with the presence of the target.
2. Uncertainty Sampling with Relevance Weighting – Traditional uncertainty (e.g., entropy of a classifier’s prediction) is multiplied by the learned relevance score, biasing the sampler toward uncertain points that are also conceptually promising.
3. Online Meta‑Learning Loop – As new samples are collected, the model performs a meta‑update: it forms a meta‑batch consisting of diverse concepts (ensuring coverage of the latent concept space) and updates its parameters to quickly adapt to the latest data.
4. Iterative Sampling Cycle – The system repeatedly (a) selects the next location via the relevance‑weighted query, (b) obtains the true label (e.g., lab test), (c) updates the relevance model and the classifier via the meta‑learning step, and (d) repeats until the budget is exhausted.

The entire pipeline runs online, requiring only modest compute (a few gradient steps per iteration) and can be wrapped around any standard classifier (logistic regression, shallow neural nets, etc.).

Results & Findings

• The relevance‑weighted sampler discovers ~30 % more contaminated sites than the next‑best method when only 10 % of the sampling budget is used.
• Ablation studies show that removing concept weighting drops performance to near classic uncertainty sampling, confirming the importance of the relevance model.
• The meta‑batch diversity mechanism reduces catastrophic forgetting when the underlying contamination pattern shifts (e.g., after a new industrial discharge), maintaining stable performance over time.

Practical Implications

• Environmental Monitoring Platforms – Agencies can plug the framework into existing GIS tools to prioritize field sampling, dramatically cutting lab costs while improving detection of hazardous substances.
• Disaster Response & Public Health – Rapidly evolving hazards (e.g., chemical spills, disease vectors) can be tracked with fewer on‑ground measurements, enabling faster, data‑driven decision making.
• Edge‑Deployable Solutions – Because the algorithm relies on lightweight gradient updates and inexpensive concept features, it can run on edge devices (drones, handheld sensors) that collect and process data in the field.
• Generalizable to Other Domains – Any problem where locations are described by auxiliary attributes (e.g., retail site selection, wildlife habitat surveys) can benefit from the same relevance‑guided active sampling loop.

Limitations & Future Work

• Concept Quality Dependency – The approach assumes that the supplied domain concepts are informative; poor or missing concept data can degrade relevance estimation.
• Scalability of Meta‑Updates – While lightweight, the meta‑learning step still incurs overhead that may become noticeable for ultra‑high‑resolution grids (e.g., sub‑meter satellite imagery).
• Limited Exploration of Deep Models – Experiments focused on shallow classifiers; extending to deep convolutional or transformer‑based models could unlock richer feature representations.
• Future Directions – The authors suggest incorporating automated concept discovery (e.g., using unsupervised clustering of satellite imagery) and testing the framework in multi‑agent settings where several sampling robots coordinate their queries.

Authors

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

Paper Information

• arXiv ID: 2602.17605v1
• Categories: cs.CV, cs.AI, cs.CY, cs.LG
• Published: February 19, 2026
• PDF: Download PDF