Agentic Large Language Model Workflows for Earth Observation: Tool Orchestration, Foundation-Model Adaptation, and the Evaluation Gap
Abstract
Large language model (LLM) agents have moved rapidly from general-purpose assistants to domain systems that orchestrate specialised remote-sensing models. Within roughly two years, geospatial agents have demonstrated reliable task planning across dozens of Earth Observation (EO) tasks, benchmark environments containing hundreds of tools, and fine-tuned open-weight controllers that outperform far larger proprietary models at generating geospatial tool-use chains. This review synthesises that literature and argues that it has converged on a common architecture with a common limit. Every system surveyed orchestrates models that human researchers built, over data that human analysts prepared; none designs a model, and none constructs its own analysis-ready data. We identify three gaps that follow from this. First, a reliability wall: reported agent success falls from above 95% on single-tool tasks to below 70% beyond eight tool calls, and scaling benchmark volume tenfold does not expose or improve this, so the binding variable is composition depth rather than data quantity. Second, a preprocessing gap: benchmark datasets arrive co-registered, temporally aligned and quality-flagged, so agents are never evaluated on the multi-source data-cube construction that dominates real EO practice. Third, an evaluation gap: automated LLM-as-judge scoring is widely reported as ground truth without calibration against human experts. We survey the open geospatial foundation models that now form the practical substrate for EO modelling, argue that adaptation strategy rather than architecture is the under-searched design variable, and set out a research agenda in which domain constraint — verified templates, physically grounded data contracts, explicit compute budgets and the capacity to refuse — is treated as the mechanism by which agentic autonomy becomes verifiable rather than merely fluent.
Related
- Limits and promises of earth observation foundation models in predicting multi-trophic soil biodiversity
- Parameter-Efficient Adaptation of Pre-Trained Vision Foundation Models for Active and Passive Seismic Data Denoising
- Cost-aware evaluation of time-series foundation models for urban air-quality and temperature forecasting
- Observation modes of EarthCARE/CPR with different Doppler measurement accuracy: evaluation of their applicability
- A generic framework to compare Local Climate Zone mapping workflows: application to the Paris Region
BibTeX
@misc{alheejawi2026agentic,
title = {{Agentic Large Language Model Workflows for Earth Observation: Tool Orchestration, Foundation-Model Adaptation, and the Evaluation Gap}},
author = {salah Mohammed alheejawi and Valentas Gruzauskas},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x5tn59},
url = {https://eartharxiv.org/repository/view/14838/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Agentic Large Language Model Workflows for Earth Observation: Tool Orchestration, Foundation-Model Adaptation, and the Evaluation Gap
AU - salah Mohammed alheejawi
AU - Valentas Gruzauskas
PY - 2026
DA - 2026/09/05
PB - EarthArXiv
DO - 10.31223/x5tn59
UR - https://eartharxiv.org/repository/view/14838/
AB - Large language model (LLM) agents have moved rapidly from general-purpose assistants to domain systems that orchestrate specialised remote-sensing models. Within roughly two years, geospatial agents have demonstrated reliable task planning across dozens of Earth Observation (EO) tasks, benchmark environments containing hundreds of tools, and fine-tuned open-weight controllers that outperform far larger proprietary models at generating geospatial tool-use chains. This review synthesises that literature and argues that it has converged on a common architecture with a common limit. Every system surveyed orchestrates models that human researchers built, over data that human analysts prepared; none designs a model, and none constructs its own analysis-ready data. We identify three gaps that follow from this. First, a reliability wall: reported agent success falls from above 95% on single-tool tasks to below 70% beyond eight tool calls, and scaling benchmark volume tenfold does not expose or improve this, so the binding variable is composition depth rather than data quantity. Second, a preprocessing gap: benchmark datasets arrive co-registered, temporally aligned and quality-flagged, so agents are never evaluated on the multi-source data-cube construction that dominates real EO practice. Third, an evaluation gap: automated LLM-as-judge scoring is widely reported as ground truth without calibration against human experts. We survey the open geospatial foundation models that now form the practical substrate for EO modelling, argue that adaptation strategy rather than architecture is the under-searched design variable, and set out a research agenda in which domain constraint — verified templates, physically grounded data contracts, explicit compute budgets and the capacity to refuse — is treated as the mechanism by which agentic autonomy becomes verifiable rather than merely fluent.
ER -