Research map / Multi-agent systems and games / LLM agents
Earth-Agent: Unlocking the Full Landscape of Earth Observation with Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Earth-Agent, an MCP-based ReAct agent with 104 EO tools spanning spectral, product, and RGB data, plus Earth-Bench (248 expert-curated tasks, ~13.7K images) with dual-level trajectory and outcome evaluation.
Problem
Existing MLLM-based and agent-based Earth observation systems are limited to RGB imagery, shallow single-step reasoning, few images, and lack structured tool ecosystems and systematic evaluation of reasoning trajectories for complex quantitative EO analysis.
Approach
A ReAct-style agent over an MCP-based toolkit of 104 EO tools (Index, Inversion, Perception, Analysis, Statistics) that unifies RGB, spectral, and product data, evaluated with a dual-level protocol scoring both final answers and step-by-step tool trajectories against expert Python solutions.
Key results
- On Earth-Bench Auto-Planning, GPT-5 reaches 65.99% accuracy while DeepSeek-V3.1 leads tool-any-order at 78.31%.
- On Earth-Bench-Lite, Earth-Agent (GPT-5) achieves 55.83% average vs. 40.42% for GPT-Agent and 26.14% for Manus.
- On the classification/detection/grounding table, Earth-Agent reports 93.42 (AID), 60.88 (DOTA), and 60.46 (DIOR-RSVG), above listed RS MLLMs.
Surprising findings
- Instruction-following regimes improve tool-call accuracy but do not necessarily improve, and can decrease, final accuracy for advanced models.
- Open-source DeepSeek-V3.1 and Kimi-K2 outperform GPT-5 on tool-use trajectory metrics despite lower final accuracy.
Limitations
- Accuracy on RGB and Spectrum tasks is limited by underlying expert perception and spectral tools.
- Final accuracy remains modest (around 50-66%) even with SOTA LLMs, and tool-exact-match and parameter accuracy are weak.
- General-agent comparison is restricted to a 60-question Lite subset because open agents cannot process hundreds of images.
Experimental setup
Datasets: Landsat 8/9 MODIS ASTER GEE products (LST, NDVI, GPM, VIIRS, GHSL, NDWI, fire MaxFRP) AID DIOR-RSVG DOTA NWPU-VHR-10 xBD HRSC2016
Benchmarks: Earth-Bench Earth-Bench-Lite AID WHU-RS19 DOTA HRSC2016 DIOR-RSVG
Metrics: accuracy efficiency tool-any-order tool-in-order tool-exact-match parameter accuracy latency
Baselines: GPT-Agent MGX Manus Coze MiniGPT-v2 LLaVA-1.5 Sphinx GeoChat VHM LHRS-Bot
Builds on: ReAct Model Context Protocol
Closest papers by meaning
- EarthDial: Turning Multi-sensory Earth Observations to Interactive Dialogues (CVPR 2025 · similarity 0.88)
- Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.87)
- See What We Cannot See: A Geo-guided Reasoning Benchmark for Object Counting under Adverse Earth Observation Conditions (CVPR 2026 · similarity 0.87)
- GeoMMBench and GeoMMAgent: Toward Expert-Level Multimodal Intelligence in Geoscience and Remote Sensing (CVPR 2026 · similarity 0.87)
- TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation (CVPR 2026 · similarity 0.87)
- RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.