GRSS Special Stream on Tool-Using Agents for Remote Sensing: Autonomous Geospatial Planning, Reliable Tool Orchestration, and Decision-Level Evaluation
Many consequential remote-sensing tasks are not single inference problems. They require a sequence of actions: discovering and retrieving observations, checking metadata and coordinate reference systems, preprocessing sensor data, running image or signal algorithms, querying geospatial databases, invoking GIS operations or physical models, comparing alternative results, and translating evidence into a decision. Today these workflows are largely hand-crafted and brittle. Tool-using agents offer a different computational paradigm: an agent interprets a task, constructs an executable plan, selects and parameterizes appropriate tools, observes intermediate outputs, verifies whether each action succeeded, and replans when evidence or operating conditions change. For example, a flood-mapping agent could inspect cloud cover, acquisition time, and data quality; choose usable optical imagery or switch to SAR under cloud; invoke preprocessing and GIS overlay tools; compare the derived extent with terrain and hydrographic constraints; and reject, revise, or rerun the inundation map when verification checks fail.
The central scientific problem is therefore not merely whether a foundation model can recognize a scene or answer a question, but how an agent can control a remote-sensing workflow reliably. Remote-sensing agents must reason about spatial scale, time, sensor physics, units, geometry, topology, cloud and data quality, and coordinate systems while interacting with heterogeneous software and data services. Research is needed on agent controllers, task and tool representations, action selection, argument grounding, state and memory, execution monitoring, intermediate-result verification, error recovery, multi-agent coordination, and provenance. These components require identifiable algorithmic innovation and evaluation beyond prompt engineering or the simple chaining of existing models.
List of Topics:
- Remote-sensing-specific agent architectures and closed-loop decision-and-action controllers
- Autonomous task decomposition, planning, replanning, and termination under geospatial constraints
- Discovery, selection, routing, and orchestration of GIS operations, remote-sensing processors, code, retrieval systems, geospatial databases, physical models, and sensor services
- Reliable tool invocation: interface learning, schema alignment, argument grounding, parameter generation, units, scale, and coordinate-system handling
- Observation-feedback mechanisms for intermediate-result validation, execution monitoring, error diagnosis, recovery, reflection, and self-correction
- State, memory, provenance, and traceable action histories for long-horizon remote-sensing workflows
- Physics- and geography-constrained reasoning for selecting tool sequences and validating outputs
- Tool-mediated multimodal agents for optical, SAR, hyperspectral, LiDAR, thermal, map, text, and in-situ data
- Retrieval and geospatial knowledge access as agent actions, including source attribution, temporal validity, and evidence grounding
- Multi-agent division of labor, communication, negotiation, consensus, conflict resolution, and collaborative tool use
- Autonomous sensor tasking, active acquisition, adaptive observation, and closed-loop processing
- Mixed-initiative human-agent workflows, approval gates, safe fallback, and intervention-aware planning
- Remote-sensing agent benchmarks and metrics for task success, tool-selection accuracy, argument validity, efficiency, reliability, calibration, robustness, and failure recovery
- End-to-end system validation in real Earth-observation tasks where the agentic and tool-use contribution
is explicitly isolated and quantified
Guest Editors:
- Guoxiong Zhou, Central South University of Forestry and Technology, China
- Weiwei Cai, Jiangnan University, China
- Johnny (Liujun) Li, University of Idaho, USA
- Jialei Zhan, University of Defense Technology, China
 Schedule:
- Submission System Open: November 10, 2024
- Submission System Close: April 30, 2025
All submissions must be formatted using the IEEE standard format (double-column, single-spaced) typically 4-5 pages. Please use the following link to download the article template. Please see GRSL’s scope and the Checklist for Authors for more information. All submissions will be peer-reviewed according to the GRSS guidelines. Submitted articles must not have been published or be under review elsewhere. Submit your manuscript to ieee.atyponrex.com/journal/grsl , using the Author Portal interface, and select our specific stream. Accepted papers are subject to GRSL’s usual publication charges.
