Paper deep dive
Earth-Agent-Pro: Towards Real-World Full-Chain Earth Observation with Agents
Zhutao Lv, Chenhao Dang, Yi Feng, Yanpei Gong, Xiaolei Wang, Junyan Ye, Conghui He, Weijia Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 9/16/2026, 8:57:37 AM
Summary
The paper introduces Earth-Agent-Pro, an execution-adaptive Plan-and-Execute framework for full-chain Earth Observation (EO), and Earth-Bench-Pro, a benchmark with 744 open-ended questions across 248 task cores. The framework uses expert-authored skills to constrain planning and tool use, employs workflow-centered structured memory for dependency tracking and repair, and utilizes separate LLM adapters for planner workflow composition (SFT) and executor tool-argument grounding (GRPO). Earth-Agent-Pro significantly outperforms baselines like ReAct in LLM-as-Judge accuracy and Tools-In-Order metrics.
Entities (10)
Relation Signals (7)
Earth-Agent-Pro → evaluatedon → Earth-Bench-Pro
confidence 98% · Earth-Bench-Pro instantiates 248 expert-curated task cores... Earth-Agent-Pro achieves 66.13% LLM-as-Judge accuracy
Earth-Bench-Pro → contains → 744 questions
confidence 95% · Earth-Bench-Pro instantiates 248 expert-curated task cores as 744 questions
Earth-Agent-Pro → employstrainingmethod → Supervised Fine-Tuning
confidence 95% · sequence-level supervised fine-tuning for planner workflow composition
Earth-Agent-Pro → employstrainingmethod → Group Relative Policy Optimization
confidence 95% · node-level group relative policy optimization with locally verifiable rewards for executor tool-argument grounding
Earth-Agent-Pro → outperforms → ReACT
confidence 95% · With a shared GPT-5 backbone, Earth-Agent-Pro achieves 66.13% LLM-as-Judge accuracy, exceeding ReAct by 20.95 points in this metric
Earth-Agent-Pro → uses → Workflow-Centered Structured Memory
confidence 95% · Workflow-centered structured memory records planned steps, accepted evidence, and their dependencies, enabling repair of only the affected workflow suffix when runtime evidence invalidates a step.
Qwen3.5-9B → backbonefor → Earth-Agent-Pro
confidence 90% · Joint adapter tuning raises Qwen3.5-9B LLM-as-Judge accuracy from 38.31% to 50.00%
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Real-world Earth observation (EO) agents must translate high-level scientific questions into executable workflows to acquire observations, prepare data, perform domain computations, and derive conclusions from runtime evidence. Existing EO agents typically start from supplied observations, while benchmarks typically provide prepared inputs or candidate answers, leaving full-chain open-world EO execution largely untested. We present Earth-Agent-Pro, an execution-adaptive Plan-and-Execute framework using expert-authored skills to constrain planning and runtime tool use. Workflow-centered structured memory records planned steps, accepted evidence, and their dependencies, enabling repair of only the affected workflow suffix when runtime evidence invalidates a step. Separate large language model adapters use sequence-level supervised fine-tuning for planner workflow composition and node-level group relative policy optimization with locally verifiable rewards for executor tool-argument grounding. Earth-Bench-Pro instantiates 248 expert-curated task cores as 744 questions under three matched regimes. Its 248 Open-World Execution questions span RGB imagery, spectral observations, and remote sensing products, pairing high-level requests with runtime data requirements, executable trajectories, and open-ended answers grounded in execution evidence. With a shared GPT-5 backbone, Earth-Agent-Pro achieves 66.13% LLM-as-Judge accuracy, exceeding ReAct by 20.95 points in this metric and 24.44 points in Tools-In-Order. Joint adapter tuning raises Qwen3.5-9B LLM-as-Judge accuracy from 38.31% to 50.00%, an 11.69-point gain over the untuned configuration. Planning-only evaluation and execution with the reference workflow show that the adapters improve workflow composition and argument grounding, respectively. Code and datasets will be released soon.
Tags
Links
- Source: https://arxiv.org/abs/2609.12533v1
- Canonical: https://arxiv.org/abs/2609.12533v1
Trouble viewing inline? Open PDF directly →
Full Text
102,146 characters extracted from source content.
Expand or collapse full text
1 Earth-Agent-Pro: Towards Real-World Full-Chain Earth Observation with Agents Zhutao Lv 1,2 , Chenhao Dang 3,4* , Yi Feng 5* , Yanpei Gong 6* , Xiaolei Wang 2 , Junyan Ye 2 , Conghui He 4 , Weijia Li 1,4† 1 Tsinghua Shenzhen International Graduate School, Tsinghua University, 2 Sun Yat-Sen University, 3 Shanghai Jiao Tong University, 4 Shanghai Artificial Intelligence Laboratory, 5 Tianjin University, 6 Harbin Institute of Technology Abstract—Real-world Earth observation (EO) analysis requires agents to be capable of translating high-level scientific questions into executable workflows for acquiring observations, preparing data, performing domain computations, and deriving conclusions from runtime evidence. However, existing EO agents and bench- marks generally cover only part of this chain: agents typically start from supplied observations, whereas benchmarks typically provide prepared inputs or candidate answers, leaving full-chain open-world EO execution largely untested. We present Earth- Agent-Pro, an execution-adaptive Plan-and-Execute framework that uses expert-authored skills to constrain workflow planning and runtime tool use. It records planned steps, accepted evidence, and their dependency links in workflow-centered structured memory, then repairs only the affected workflow suffix when runtime evidence invalidates a step. Role-specific training adapts the underlying large language model (LLM) through separate planner and executor adapters: sequence-level supervised fine- tuning targets planner workflow composition, while node-level group relative policy optimization with locally verifiable rewards targets executor tool-argument grounding. To evaluate full-chain open-world EO execution, we introduce Earth-Bench-Pro, a 744- question EO agent benchmark that instantiates 248 expert- curated task cores under three matched regimes. Across RGB imagery, spectral observations, and remote sensing products, its 248 Open-World Execution questions pair high-level requests with runtime data requirements, executable trajectories, and open-ended answers grounded in execution evidence. Under a shared GPT-5 backbone, Earth-Agent-Pro reaches 66.13% LLM- as-Judge accuracy, outperforming ReAct by 20.95 points in LLM- as-Judge accuracy and 24.44 points in Tools-In-Order. Jointly tuning the Planner and Executor adapters raises Qwen3.5-9B LLM-as-Judge accuracy from 38.31% to 50.00%, an 11.69-point gain over the untuned configuration. Planning-only evaluation and execution with the reference workflow further show that the adapters improve workflow composition and argument ground- ing, respectively. In conclusion, Earth-Bench-Pro and Earth- Agent-Pro support evaluation and training of EO agents from user requests to evidence-grounded scientific answers. The code and datasets of this work will be released soon. Index Terms—Earth observation, remote sensing agents, open- world execution I. INTRODUCTION Earth observation (EO) turns satellite and aerial measure- ments into quantitative evidence for agricultural monitoring, disaster assessment, urban analysis, and environmental change research [1]–[15]. Producing such evidence from an EO ques- tion remains expert-intensive [16]–[18]. An EO analysis begins * Equal contribution. † Corresponding author. by matching the target phenomenon, region, and period to suitable observations. Analysts then choose the data source, product, bands, and sampling parameters before acquiring and preprocessing the data. The prepared observations enter a domain-specific computation pipeline that yields the requested scientific quantities. Completing this workflow requires ex- pertise in sensors, data products, geospatial processing, and domain science, which raises the entry barrier for users outside the EO community and limits access to reproducible analysis. Recent work has expanded EO capabilities beyond inter- preting supplied observations to include multistep analysis execution. Vision-language models interpret supplied images and image sequences [19]–[21], while tool-augmented agents select operations, execute code, and construct multistep work- flows [22]–[26]. Benchmarks now span fixed-input question answering, multistep geographic information system (GIS) tool use, and open-environment pipeline construction [27]– [30]. Together, these methods and benchmarks cover more of the request-to-answer chain, but evaluations differ in their scientific tasks, input assumptions, and answer formats. Mea- suring progress toward autonomy therefore requires a matched evaluation that removes expert scaffolding while keeping the scientific objective fixed. Our prior Earth-Agent and Earth-Bench provide the task and evaluation foundation for this comparison [31]. Earth-Agent connects a language model to EO tools through the Model Context Protocol, while Earth-Bench evaluates final answers and reference trajectories across red-green-blue (RGB) im- agery, spectral observations, and remote sensing products. This pairing evaluates both the scientific result and the execution path. Earth-Bench, however, begins with expert-prepared local inputs and multiple-choice questions, leaving runtime data discovery and open-ended answer generation outside its eval- uation boundary. We build on this foundation by preserving each of Earth-Bench’s 248 scientific objectives as a fixed task core and moving runtime data discovery, acquisition, and open-ended answer generation into the evaluated workflow. The resulting request-to-answer setting, which we call open- world EO execution, starts from a request that specifies a phenomenon, region, and period. The agent must locate runtime inputs, acquire suitable observations when required, prepare them, perform domain computations, and derive an answer from runtime evidence. Files, intermediate artifacts, and tool results appear only during execution, so later op- arXiv:2609.12533v1 [cs.CV] 11 Sep 2026 2 Existing EO Benchmarks Q: Calculate the distance between A and B. (A) 0m (B) 30m (C) 80m (D) 200m Multiple Choice Pick from options Easier The distance between A and B is 79.95 meters. Open-ended No reference options Much Harder Drawbacks 1. Data acquisition is skipped 2. Answers are reduced to multiple choice 3. Actual numeric/date/unit errors can be hidden 4. Has a gap from the real-world EO needs Pre-collected data Same Workflow Existing EO Agents Local Tools External Knowledge Tool A Tool BTool CAnswer Tool D Tool B Drawbacks 1. Tool selection lacks systematic constraints 2. Planning paths become redundant or indirect 3. General-purpose models lack EO-specialized training Ours (Earth-Bench-Pro & Earth-Agent-Pro) 744 expert-curated open-ended tasks Introduce Data Download Search Data Source Download Series Extra Generated Data for Model Training Skill-guided Agent ? Open EO task Planner Skill 1 Skill 2 Skill 3 Skill n ... Workflow Executor Lightweight Model Advantages 1. Open-world benchmark with open-ended EO tasks 2. Skill-guided agent for full-chain EO workflows 3. Trained lightweight agent for planning and execution Answer Unrestricted tools / information Earth-Bench-Pro Earth-Agent-Pro ! Redundant Path Fig. 1. Overview of our work. Existing EO benchmarks rely on pre-collected data and multiple-choice evaluation, while existing EO agents suffer from unconstrained tool use and redundant planning. Our work introduces an open-ended EO benchmark (Earth-Bench-Pro) and a trained skill-guided agent (Earth- Agent-Pro) for full-chain EO task planning and execution. 0 10 20 30 40 50 60 70 Ours GEMMA-4 (E4B-it) Mistralai- v03-7B InternLM3 -8B Qwen3.5- 9B Llama-33-70B -Instruct GPT-oss- 120b Score (%) Instruction Following Autonomous Planning Open-World Execution Fig. 2. LLM-as-Judge accuracy across Earth-Bench-Pro’s three regimes. Ours is our role-specialized Qwen3.5-9B. erations depend on accepted upstream evidence. The agent must track workflow dependencies, ground each operation in the evolving state, and revise affected future operations when evidence invalidates the current plan. We develop Earth-Bench-Pro and Earth-Agent-Pro as a paired benchmark and agent for open-world EO execu- tion. Earth-Bench-Pro holds each scientific objective fixed across matched regimes that progressively remove the refer- ence workflow and prepared inputs. Earth-Agent-Pro separates workflow planning from runtime operation grounding and constrains both stages with expert-authored skills. Workflow- centered structured memory links pending operations to ac- cepted evidence, allowing the agent to preserve a valid prefix and repair only an invalidated workflow suffix. The training pipeline specializes the Planner for workflow composition and the Executor for runtime argument grounding. Earth-Bench-Pro exposes a consistent difficulty gap: Open- World Execution remains the hardest regime across backbones (Figure 2). On Earth-Bench-OW, with a shared GPT-5 back- bone, tool set, and skill set, Earth-Agent-Pro reaches 66.13% accuracy under a large language model (LLM) judge, 20.95 percentage points above ReAct, and leads ReAct by 24.44 points on Tools-In-Order. Combining the Planner and Executor adapters raises Qwen3.5-9B LLM-as-Judge accuracy by 11.69 points; planning-only and oracle-plan evaluations separately measure workflow composition and argument grounding. Our contributions are: • Open-world execution task and benchmark. We for- mulate open-world EO execution and operationalize it with 248 Open-World Execution questions in Earth- Bench-Pro. Product and Spectrum tasks require runtime observation acquisition and preparation, and all questions require open-ended answers grounded in execution evi- dence. Together with 496 matched Instruction-Following and Autonomous-Planning questions, the released suite contains over 700 evaluation questions. • Open-world EO agent framework. Earth-Agent-Pro uses expert skill guidance to constrain workflow con- struction and runtime tool use, while its execution- adaptive dynamic workflow maintains the plan, accepted evidence, and their dependencies. Under a shared back- bone, tool set, and skill set, Earth-Agent-Pro outperforms ReAct, AFlow, and OpenEarthAgent on final-answer and workflow-order metrics in controlled comparisons. • Dependency-preserving data and role training. From 248 seed workflows, we construct 2,480 dependency- preserving trajectories and combine them with examples from the public OpenEarthAgent training set [25] to form a balanced mixture for Planner SFT and Executor GRPO. Combining the two adapters raises Qwen3.5-9B LLM-as- Judge accuracy by 11.69 points on Earth-Bench-OW. I. RELATED WORK A. Earth Observation Benchmarks and Datasets Earth observation (EO) datasets provide perception and multimodal evaluation on fixed inputs [32]. RSVQA [27] and VRSBench [28] assess visual question answering; GeoBench- VLM [33] and CHOICE [34] evaluate geospatial multimodal 3 Task uses 136 198019902000201020202025 Relative task density: lower - higher (b) Global Spatiotemporal&Temporal Coverage P r o d u c t S t a t s . S p e c t r a l I n d e x T h e r m a l R e t r . B T T e m p . T i m e - S e r i e s R G B P e r c e p t i o n Agriculture & Drought Urban & Infrastructure Water & Marine Vegetation & Ecosystems Cryosphere & Snow Atmosphere & Climate Hazards & Disasters Land Cover & Object Mapping Lower share Higher share Application domain (a) Benchmark Composition Fig. 3. Earth-Bench-Pro Overview. (a) Task composition across application domains and task families. (b) Global spatial distribution and temporal coverage. capabilities; VLEO-Bench [35], XLRS-Bench [36], URBench [4], DisasterM3 [37], and TerraScope [21] extend evaluation to fine-grained, ultra-high-resolution, urban, disaster, and pixel- grounded reasoning. These benchmarks mainly score final outputs rather than agent interactions, tool-use trajectories, or intermediate artifacts. Complementary EO benchmarks target specific agentic operations: RescueADI [38] evaluates se- quential disaster interpretation; GeoLLM-Engine [39] provides analyst-style API and interface interactions; GeoBenchX [29] tests multistep GIS tool calling; AEOS-Bench [40] considers constrained constellation scheduling. Agent-oriented bench- marks formalize these capabilities in their protocols: ThinkGeo [22] assesses tool use, Earth-Bench [31] evaluates trajectory- level EO tasks, and HTAM/GeoPlan-Bench [24] targets hi- erarchical task abstraction. More recently, EO-Gym [41] and TerraBench [42] introduce interactive heterogeneous data; Ge- oMMBench [43] and TerraLogic [44] target complex geosci- entific reasoning; and GeoAgentBench [45] and DORA [46] assess GIS execution verification and emergency geospatial reasoning. OpenEarth-Bench [30] evaluates open-environment EO pipelines and tool creation, whereas NASA-EO-Bench [47] focuses on data and tool discovery. Yet evaluations often start from curated cases or treat discovery separately, leaving open the joint assessment of data-source selection, acquisition, intermediate evidence, and final natural-language answers. B. Earth Observation Methods MLLM-based EO Methods. EO representation and per- ception methods, including SpectralGPT [48], RingMo-Aerial [49], REST [50], RingMoE [51], CrossEarth [52], and Hyper- SIGMA [53], have strengthened representation learning and visual interpretation across spectral, aerial, and hyperspectral observations [54]. In parallel, RSGPT [55], SkySenseGPT [56], GeoChat [19], LHRS-Bot [20], EarthGPT [57], Earth- Marker [58], EarthGPT-X [59], EarthDial [60], VHM [61], GeoPixel [62], and SkyEyeGPT [63], extend captioning and visual question answering to scene understanding, grounding, multi-sensor dialogue, and pixel-level reasoning [64], [65]. These advances strengthen understanding of provided imagery or products but generally do not discover external data, coor- dinate tool use, or maintain evidence through long-horizon EO analysis. Agent-based EO Methods. General agents establish plan- ning and tool-use paradigms through reasoning-action inter- leaving in ReAct [66], API-grounded tool use in Gorilla [67] and ToolLLM [68], reusable skills in Voyager [69], and work- flow optimization in AFlow [70]. Early EO agents, including RS-ChatGPT [71], RS-Agent [72], and Change-Agent [73], connect LLMs with visual models or expert tools for percep- tion tasks. Subsequent systems organize broader workflows through EO code generation in UnivEarth [23], tool use in ThinkGeo [22], MCP-based tool integration in Earth-Agent [31], knowledge-flow fusion in CangLing-KnowFlow [74], and hierarchical abstraction in HTAM [24]. Recent directions extend to multi-platform reasoning with RingMo-Agent [75], climate science with EarthLink [76], open geospatial environ- ments with OpenEarthAgent [25] and OpenEarth-Agent [30], workflow automation with GeoFlow [77], experience-driven cooperation with GeoEvolver [78], vague-intent handling with RemoteAgent [79], and expert-level reasoning with GeoM- MAgent [43]. However, existing systems commonly plan over broad or predefined tool spaces and optimize planning and execution as a unified policy. They rarely use expert task skills to constrain both workflow construction and runtime tool scope, or explicitly maintain accepted evidence and dependencies as execution evolves. Role-specific alignment of workflow composition and node-level argument grounding therefore remains largely underexplored. I. EARTH-BENCH-PRO BENCHMARK A. Benchmark Overview Earth-Bench-Pro builds a 744-question EO agent benchmark from 248 matched scientific task cores. Earth-Bench-Pro com- prises three matched sub-benchmarks: Earth-Bench-IF, Earth- Bench-AP, and Earth-Bench-OW. Earth-Bench-IF provides an ordered reference tool sequence and prepared inputs. Earth- Bench-AP hides the sequence while retaining prepared inputs. Earth-Bench-OW requires agents to acquire or discover data 4 TABLE I COMPARISON WITH RECENT EARTH-OBSERVATION AGENT BENCHMARKS. VALUES MARKED WITH ∗ ARE NOT EXPLICITLY REPORTED IN THE ORIGINAL PAPERS AND ARE COMPUTED FROM THE CORRESPONDING OFFICIAL OPEN-SOURCE RELEASES. BenchmarkTasks StepsAvg. Tools Multi- source EO Data Open-ended QA Executable Trajectory Real Data Acquisition Preprocessing Required Quantitative Answer Intermediate Observations Ours7445,2957.1112✓ OpenEarth- Bench [30] 596–✓×✓Part. OpenEarthAgent [25] 1,169 7,0646.0424✓Part.Part.✓ KnowFlow- Bench [74] 324–✓×✓Part.✓Part. TerraBench [42]403∼24,500∼60.8 77✓Part.✓ EO-Gym [41]1,436 5,323 ∗ 3.7 ∗ 35✓Part.Part.Part.✓ GeoPlan-Bench [24]996 10,619 ∗ 10.7 ∗ 104 ∗ ×✓× at runtime and subsequently generate an open-ended answer grounded in execution evidence. As illustrated in Figure 3, Earth-Bench-Pro covers eight application domains and six analysis types. The applica- tion domains include agriculture and drought [80], urban areas and infrastructure, water and marine systems, vegetation and ecosystems [81], cryosphere and snow, atmosphere and climate, hazards and disasters, and land cover and object mapping. The analysis types include product statistics, spec- tral indices, thermal retrieval, brightness-temperature analysis, time-series analysis, and RGB perception. The tasks also span regions across all six inhabited continents and requested observation periods from 1980 to 2025, covering a broad range of geographic settings and temporal extents. Recent EO agent benchmarks cover only parts of the end-to-end workflow. OpenEarth-Bench [30] and KnowFlow- Bench [74] include acquisition or preprocessing but omit open- ended answers. OpenEarthAgent [25], TerraBench [42], and EO-Gym [41] evaluate open-ended answers and tool trajecto- ries but provide partial coverage of acquisition, preprocessing, or quantitative analysis. GeoPlan-Bench [24] isolates workflow planning without executable reference trajectories or runtime acquisition. As shown in Table I, Earth-Bench-Pro is the only benchmark in this comparison that covers all seven dimensions. This design evaluates both final answers and the complete execution processes that produce them. B. Problem Formulation Earth-Bench-Pro represents each benchmark sample as z = (q,c,τ ⋆ ,y ⋆ ), where q is a scientific query, c encodes the input context or runtime data requirements, and y ⋆ is the reference answer. Let N z denote the number of steps in the reference trajectory τ ⋆ . The EO environment E executes these steps. Each step records tool t ⋆ i , arguments θ ⋆ i , and the resulting observation o ⋆ i : τ ⋆ = (t ⋆ i ,θ ⋆ i ,o ⋆ i ) N z i=1 , o ⋆ i =E (t ⋆ i ,θ ⋆ i ).(1) The three evaluation regimes expose different information to the agent. Let π ⋆ = (t ⋆ 1 ,...,t ⋆ N z ) denote the reference tool sequence, and let φ construct an instruction that states these tools in order alongside the scientific query. Instruction following uses the resulting instruction q IF = φ(q,π ⋆ ) and prepared inputs c prep . Autonomous planning uses the query without reference tools and the same prepared inputs. Open- world execution provides runtime input requirements c req , leaving the agent to acquire or discover data during execution: q IF = φ(q,π ⋆ ), I IF = (q IF ,c prep ), I AP = (q,c prep ), I OW = (q,c req ). (2) In the open-world execution, f denotes the tool-selection policy, g denotes the runtime argument generator, and h denotes the evidence-conditioned answer function. The tool and arguments for call i depend on the query, runtime re- quirements, and preceding accepted execution trace. The en- vironment returns a new observation, and the agent generates its answer from the completed trajectory: τ = (t i ,θ i ,o i ) L i=1 , t i = f (q,c req ,τ <i ), θ i = g(q,c req ,t i ,τ <i ), o i =E (t i ,θ i ), ˆy = h(q,c req ,τ ). (3) Execution therefore directly determines the selected data, intermediate dependencies, and ultimately the final answer. C. Data Curation Earth-Bench-Pro treats a complete executable evidence chain as the unit of annotation, linking each scientific re- quest to its answer through tool execution. We assembled an eight-member annotation team with two computer science experts, three remote sensing specialists, and three Earth science specialists. The team jointly reviews tool executability, remote sensing data requirements, scientific interpretation, and answer correctness, bringing technical constraints and domain judgment into the same annotation process. 5 Earth-Agent-Pro (a) Skill-Guided Workflow Planning (b) Workflow-Centered Structured Memory (c) Evidence-Grounded Execution & Repair Open-World EO Task& Input User Request On 12 Jul 2021, what percentage of the Chengdu Plain had TVDI > 0.75? Download Context Download-first Initial files: none MOD11A1 LST | scale 0.02 MOD13Q1 NDVI | scale 1e-4 2021-07-12 | GSD 30 m Task Constraints AOI bbox: [103.800, 30.300, 104.500, 30.501] Criterion: TVDI > 0.75 Required output: area ratio Shared by Planner and Executor Expert Skills 풜ₛ Tool Catalog 풯 EO Environment SkillRouting 4-role Sequential Debate Planner Catalog Validator Workflow Skeleton π: Dynamic workflow Wᵢ Accepted evidence Mᵢ accepted call Reuseable artifacts& provenance Executor: ReAct loop constrained to current node Ground& Execute current node Hybrid Verification Accept If Terminal: Update Memory-> Answer Controller If Non-terminal: Update Memory-> Next Node In-nodecorrection fix arguments -> retry current node AnswerController: StructuralRepair Preserve Prefix -> Revise Suffix Verified EO Answer accepted results only√ evidence complete√ traceable provenance√ Scientific response grounded in verified execution error retry Successful Prefix Current Node Vᵢ Revisable Suffix ->Active Skill： earth-acquire- thermal-retrieval Download×2 -> Scale×2 -> TVDI -> Threshold Download×2 : LST, NDVI Scale×2 : 0.02 / 0.0001 Compute TVDI Calculatethreshold ratio: TVDI > 0.75 TVDI > 0.75 Affected cropland 28.00% Node Vᵢ: compute_tvdi(lst_path, ndvi_path) Mᵢ -> NDVI / LST paths； Execute -> TVDI candidate Download LST Download NDVI Scale LST Scale NDVI Derived grounding view: Scaled LST: MOD11A1 2021-07-12 Scaled NDVI: MOD13Q1 2021-07-12 ⚫lst_path<-Scaled LST ⚫ndvi_path<-Scaled NDVI Evidence gap Retries exhausted specified by user query and task context Fine- tuning EO workflow specialist -> Tool schema specialist -> Benchmark minimalist -> Skeptical moderator -> Final candidate workflow Tool exists? Skill-allowed? Valid sequence ? valid -> accept error -> retry •Deterministic Gates •LLM Checker Task evidence sufficient? Yes -> Answer No -> Resume workflow 풮 푖 =푊 푖 ,푀 푖 Fig. 4. Earth-Agent-Pro Framework. (a) Skill-guided workflow planning uses a shared expert specification to constrain both workflow construction and runtime argument grounding. (b) Workflow-centered structured memory couples the evolving workflow with provenance-aware accepted evidence, preventing failed attempts from contaminating downstream execution. (c) Evidence-grounded execution and repair performs node-local correction and, upon persistent failure or an evidence gap, preserves the successful prefix while revising only the affected workflow suffix. Together, these mechanisms enable traceable and bounded adaptation without discarding valid execution progress. Experts transform 248 scientific task cores into open-world workflows for which execution determines the inputs, inter- mediate dependencies, and answers. These tasks retain the scientific objectives of Earth-Bench [31]. Experts re-execute the original solutions in a shared EO execution environment and correct inconsistencies among the question, tool selection, call arguments, intermediate dependencies, and answer. For 188 Product and Spectrum tasks, experts replace fixed local inputs with online acquisition and required preprocessing, authoring 323 task-level acquisition specifications. Each spec- ification encodes the required source, product, bands, spatial extent, observation period, resolution, sampling interval, and scaling information. The remaining 60 RGB tasks retain user- provided visual inputs, while concrete files still require runtime discovery. After reconstruction, file paths and intermediate artifacts become execution outcomes, and downstream calls must bind the outputs returned by upstream tools. Each reconstructed task pairs an open-ended query with reference evidence regenerated through execution. Experts remove all answer choices and rewrite each question as a self-contained request specifying the scientific objective, study region, observation period, and task constraints. They rerun the workflow in the same environment, record every call and returned observation, and recompute the answer from the outputs. Each sample retains the representation introduced in Section I: z = (q,c,τ ⋆ ,y ⋆ ), τ ⋆ = (t ⋆ i ,θ ⋆ i ,o ⋆ i ) N z i=1 .(4) Here, context c records runtime input requirements, and each reference step records tool t ⋆ i , arguments θ ⋆ i , and obser- vation o ⋆ i . This representation preserves artifact provenance, data dependencies, and answer evidence along the complete execution path. Quality control checks three relations along the evidence chain: queries against input requirements, data dependencies across tool calls, and final answers against execution evidence. Experts retain and inspect intermediate rasters, files, numerical outputs, detections, masks, and statistics. They also verify answer values, units, precision, temporal references, and trend descriptions. The 248 verified Open-World Execution work- flows contain 1,765 tool calls, average 7.1 calls per task, and use 84 distinct tools. We derive matched Instruction-Following and Autonomous-Planning variants from each verified task core, yielding 744 questions and 5,295 reference tool calls. Every answer retains a traceable execution basis, and every intermediate result records its upstream source. IV. EARTH-AGENT-PRO: AN EXECUTION-ADAPTIVE AGENT FRAMEWORK A. Framework Overview Earth-Agent-Pro is a Plan-and-Execute-style agent that sep- arates scientific workflow planning from runtime operation grounding. Before concrete files and intermediate results be- come available, the Planner constructs an ordered workflow skeleton. As data and intermediate artifacts emerge, the Ex- ecutor grounds each node, invokes its tool, and updates the state. The dynamic workflow and accepted execution memory form the workflow-centered structured memory shared by both roles throughout the entire execution process. The Planner and workflow controller operationalize the abstract tool-selection policy f in Eq. (3), the argument generator G exec implements g, and the Answer Controller implements h. Figure 4 presents the complete control process. Let R skill and R tool denote skill and tool routing, letT denote the global tool catalog, and let A s denote the tools allowed by skill s. The two routing stages reduce the catalog to the candidate set 6 P s , after which the Planner generates the tool sequence π plan through sequential debate. The initialization operator converts this sequence into dynamic workflow W 0 and creates the initial state S 0 with empty memory M 0 : s = R skill (q,c), P s = R tool (q,c,s,T ∩A s ), π plan = Plan(q,c,s,P s ), S 0 = (W 0 ,M 0 ) = Init(π plan ), ∅ . (5) The Executor advances node by node over this structured state and exposes only verified results to subsequent operations. When a node fails or the final evidence remains incomplete, the Planner proposes a local replacement suffix and the Ex- ecutor installs it while retaining the successful prefix. B. Skill-Guided Workflow Planning Expert skills encode EO procedure knowledge as task-level specifications shared by the Planner and Executor. We author six skills covering spectral analysis, product-based computa- tion, and RGB perception. Each skill describes its tool scope, operation order, argument constraints, and repetition rules across acquisition, preprocessing, domain computation, and aggregation. As illustrated in Figure 5, skill guidance removes irrelevant tools before workflow generation and continues to constrain node realization during execution. The Skill Router and Tool Router implement the two constraints in Eq. (5). The full catalogT contains 112 callable tools and their schemas. The Skill Router selects a skill from the query and input context, then excludes inapplicable tools through A s . The Tool Router retrieves P s for the task objective, leaving the Planner to compose a workflow over legal and task-relevant tools for subsequent execution. The candidate tools must form a scientifically valid and interface-compatible operation sequence. The Planner uses a four-role sequential debate that examines the EO procedure, tool schemas, sequence parsimony, and final executability in order. Each role revises the candidate sequence using the preceding discussion. A catalog validator then checks tool ex- istence and active-skill membership. It returns validation errors with the candidate sequence for another planning attempt until the Planner obtains π plan . The selected skill remains active after initial planning. For the current operation, the Executor combines the skill instruc- tions with the tool schema, query constraints, and accepted upstream results to generate arguments. When deterministic rules cannot resolve a tool result, the language-model checker uses the same skill to assess whether the result follows the task procedure. The Planner and Executor therefore share one task-level specification for operation selection and runtime argument grounding. C. Workflow-Centered Structured Memory Earth-Agent-Pro represents runtime step k using the struc- tured state S k = (W k ,M k ). The dynamic workflow W k contains the active node sequence V k and the revision archive R k . The execution memory M k is a chronologically ordered (a) Without Skills Unscoped Toolscape 117 Tools Data Index Inversion Perception Statistics Analysis download_product calculate_ndvi extract_modis_bt compute_tvdi RemoteSAM Ambiguous Workflow (b) With Skills Skill-Scoped Toolscape 20~30 Allowed Tools Shortlist download product apply_scale _offset apply_cloud _mask calculate _ndvi calc_image _mean Skill-Guided Workflow Statistics AcquirePreprocess Spectral Index PreprocessAcquire Thermal Retrieval Acquire Thermal Retrieval Spectral Index Statistics AcquirePreprocessPerceptionStatistics Preprocess Spectral Index Acquire 112 Tools ✓ ✓ download product apply_scale _offset calculate _ndvi ... Spectral Index BT Temperature Thermal Retrieval Time-Series Trend Thermal Retrieval Skill-guided Skills enable focused tool selection and reliable workflows. RGB Perception Acquire Preprocess Spectral Index Spectral Index Fig. 5. With vs. Without Skills. The top shows an unscoped tool space and ambiguous workflows without skills, while the bottom shows focused tool selection and a coherent workflow enabled by skill guidance. collection of evidence records. Each record e u stores invoked tool ̄ t u , arguments θ u , observation o u , and provenance ν u : S k = (W k ,M k ), W k = (V k ,R k ), V k = (v k 1 ,...,v k n k ), M k = (e 1 ,...,e m k ), e u = ( ̄ t u ,θ u ,o u ,ν u ). (6) Each node records its planned tool, invoked tool, execution status, grounded arguments, result summary, failure diagnosis, and revision identifier. Execution memory accumulates only verified calls. Provenance for file artifacts also records vari- ables, time, region, bands, and batches, allowing the Executor to resolve concrete files, numerical series, and cross-input correspondence for the later nodes. The structured state limits the information exposed to sub- sequent operations at each step. Argument generation uses the query, current node and schema, active skill, and previously accepted call history. Rejected attempts provide retry feedback for the current node and leave the accepted history unchanged. The Executor therefore avoids repeatedly recovering the valid evidence, failed attempts, and pending operations from an expanding free-form interaction transcript. D. Evidence-Grounded Execution and Workflow Repair The Executor advances the workflow through a ReAct loop constrained to the current node. For node v k i in stateS k , let t k i denote its planned tool, let a index in-node attempts, and let e t k i,a denote the invoked tool. The first attempt sets e t k i,1 = t k i . σ(t) returns the schema of tool t. Argument generator G exec combines this schema, the current node, and accepted memory to generate arguments. Execution environment E returns an 7 observation, after which hybrid verification gate V hyb produces decision δ k i,a and feedback b k i,a : ˆ θ k i,a = G exec (q,c,s,σ( e t k i,a ),v k i ,W k ,M k ), ˆo k i,a =E ( e t k i,a , ˆ θ k i,a ), (δ k i,a ,b k i,a ) = V hyb (q,s,W k ,M k , e t k i,a , ˆ θ k i,a , ˆo k i,a ), δ k i,a ∈accept, retry, replace. (7) The hybrid gate first applies deterministic checks for struc- tural errors, error markers, and batch coverage. The language- model checker evaluates cases that these rules cannot resolve using the query, grounded arguments, active skill, and remain- ing workflow. Retry updates the arguments and feedback for the current tool. Replace selects another tool from P s and updates the feedback. Both actions remain local to the current attempt and leave persistent state S k unchanged. Accept in- stead commits the node result to the dynamic workflow and appends its evidence record to execution memory. Let ν k i,a denote call provenance, let U acc denote node commitment, and let append add an item to an ordered collection. After all in- node attempts fail, the Executor writes the final diagnosis to the failed node and then triggers localized workflow repair: e k i,a = ( e t k i,a , ˆ θ k i,a , ˆo k i,a ,ν k i,a ), S k+1 = U acc (W k ,i,e k i,a ), append(M k ,e k i,a ) , if δ k i,a = accept, S k , if δ k i,a ∈retry, replace. (8) The Planner generates candidate suffix e V i: from failed node i, where d k i denotes the final failure diagnosis stored by that node. The Executor validates candidate operations against the active skill and tool catalog. It rejects candidates that repeat an earlier repair, show no execution progress, have abnormal length, or contain excessive operation repetition. After a candidate passes these checks, the Executor uses ⊕ to concatenate the successful prefix and new suffix in execution order, then records the old suffix with its diagnosis in R k+1 . The repair leaves M k unchanged and re-executes only the failed position and subsequent operations: e V i: = Plan repair (q,c,s,W k ,M k ,i,d k i ), V k+1 =V k [1:i− 1]⊕ e V i: , R k+1 = append R k , (V k [i:], e V i: ,d k i ) , W k+1 = (V k+1 ,R k+1 ), M k+1 = M k . (9) The Answer Controller extends the same repair mechanism to workflows whose tools succeed but whose evidence remains incomplete. It checks required computations, spatiotemporal coverage, aggregation, units, numerical scale, and non-finite results. With sufficient evidence, the controller generates the final answer only from M k . Incomplete evidence produces an actionable diagnosis, after which the Executor reopens the specified node or defaults to the final node when the diagnosis identifies none. Earth-Agent-Pro therefore uses a node as the boundary for runtime correction and a workflow suffix as the boundary for structural revision, respectively. V. ROLE-SPECIALIZED TRAINING A. Training Overview The Planner and Executor expose different prediction in- terfaces. The Planner composes an ordered tool sequence from expert skill guidance and candidate tool schemas. The Executor grounds a prescribed node from accepted runtime evidence. A monolithic trajectory objective would mix work- flow errors with argument errors and spread credit across long tool chains. Role-specialized training isolates these errors at their respective prediction interfaces. Thetrainingpipelineexpandsinstancesthrough dependency-aware trajectory instantiation and then extracts role-specific supervision from complete trajectories. Let D P and D E denote the Planner and Executor records. Parameters ψ P and ψ E are the trainable low-rank adaptation (LoRA) parameters of the two role-specific adapters. Planner context x P contains the query, inputs, skill guidance, and candidate tool schemas. Executor context x E i contains the prescribed node and accepted reference-trajectory prefix. For a Planner record,ω P tokenizes reference tool sequence π ⋆ . For Executor sample j at node i,ω E i,j is the response token sequence. Supervised fine-tuning (SFT) optimizes workflow compo- sition, while on-policy group relative policy optimization (GRPO) [82] optimizes runtime argument grounding. Let ˆ A i,j denote group-relative advantage, and let ρ i,j,ℓ and ̄ρ i,j,ℓ denote the raw and clipped token-level importance ratios. The frozen reference distribution is p ref . Coefficients β and η weight the Kullback–Leibler (KL) regularizer and entropy bonus, respectively. The total objective combines the two role losses, and we optimize a separate adapter for each role: L plan (ψ P ) =−E D P X ℓ logp ψ P (ω P ℓ | x P ,ω P <ℓ ), L exec (ψ E ) =−E i,j,ℓ h min ρ i,j,ℓ ˆ A i,j , ̄ρ i,j,ℓ ˆ A i,j i + β b D KL (p ψ E ∥p ref )− ηH(p ψ E ), L train (ψ P ,ψ E ) =L plan (ψ P ) +L exec (ψ E ), (ψ ⋆ P ,ψ ⋆ E )∈ arg min ψ P ,ψ E L train . (10) B. Dependency-Preserving Training Data Construction Training data follow the executable sample representation of Earth-Bench-Pro. A seed sample is z = (q,c,τ ⋆ ,y ⋆ ) with τ ⋆ = ((t ⋆ i ,θ ⋆ i ,o ⋆ i )) N z i=1 . We decompose it into an immutable execution skeleton κ(z) and a mutable instance stateu(z). The skeleton retains the reference tool sequence π ⋆ , the corresponding tool schemas, and dependency structure G z , a mathematical abstraction of the dependencies encoded by slot groups and cross-node references. The instance state contains locations, spatial extents, temporal ranges, sample identifiers, file references, and replaceable intermediate observations. For each seed sample, a language model represents the mutable state as typed slots, groups repeated references and dependent fields, and generates a sample-specific instantiation program F z . Given root-slot values e u root , the program prop- agates replacements along these dependencies to produce ez while preserving κ(ez) = κ(z). It therefore instantiates coupled 8 LLM Replaceable-Slot Discovery for per Source Question [P] Place [I] Interim Parameter [T] Time [F] File References Slot definitions are reasonable Experts Review Quetion: Use [T]January VIIRS nighttime-light data over [P] Leon from [T] 2013-2024, calculate annual mean intensity, and fit a linear trend. download_product base_save_name: [P] Leon bbox: [I] (-101.84, 20.98, -101.54, 21.23) time_range: [T] 2013-01-01 ... 2024-01-01 source: VIIRS nighttime lights [F] Leon_avg_rad_ 2024-01-01.tif ... [F] Leon_avg_rad_ 2013-01-01.tif Generate Python scripts to make training data through substitution Workflow remains the same Constraint relationships are correct Joint Slot Substitution Training data generation Model Training Data [P] Leon · [T] 2013-2024 Paris 2015-2026 [I] (-101.84, 20.98, -101.54, 21.23) (31.12, 29.95, 31.40, 30.16) [T]+[F] Leon_avg_rad_202 4-01-01.tif Leon_RAD_AVG_2026 -01-01.tif [P]+[T]+[F] Leon_avg_rad_2024- 01-01.tif CARIO_RAD_AVG_ 20050101.tif Fig. 6. Synthetic training data generation. An LLM identifies replaceable slots for each seed sample and writes a sample-specific Python program. Human experts verify that the program preserves workflow structure and dependency constraints before it generates consistent training instances. fields such as a place and its bounding box together and propagates changed node outputs to all downstream arguments and observations. Human reviewers inspect the slot specifica- tion and program before batch synthesis. Figure 6 shows the process from a seed sample through a reviewed instantiation program to dependency-consistent training instances. Let R denote the number of mutable slots. Operator Compose re- constructs a sample from its execution skeleton and instance state. We formalize this construction as follows: κ(z) = π ⋆ ,σ(t ⋆ i ) N z i=1 ,G z , u(z) = (u 1 ,...,u R ), z = Compose κ(z),u(z) , ez = F z κ(z), e u root , κ(ez) = κ(z). (11) The reviewed programs fix the sensor or data product, selected bands, ordered tools, argument schemas, processing operations, and dependency structure. They then instantiate samples under a fixed random seed and apply quality filtering. This process yields 2,480 instantiated trajectories derived from 248 seed workflows across RGB imagery, spectral observa- tions, and remote sensing products. We add examples from the OpenEarthAgent training set [25] to cover GIS analysis, object-centric visual perception, and geometric reasoning, then carefully balance the two sources to form the final trajectory set Z train . The Planner extractor creates one sequence-level record per trajectory with target π ⋆ . The Executor extractor creates one argument record per node with target θ ⋆ i : ( D P =(x P (z),π ⋆ (z))| z ∈Z train , D E =(x E i (z),θ ⋆ i (z))| z ∈Z train , 1≤ i≤ N z . (12) C. Planner SFT for Workflow Composition Planner SFT maps domain-constrained context x P to refer- ence workflow π ⋆ . In Eq. (10),L plan computes autoregressive negative log-likelihood over the assistant response serializing π ⋆ . This response contains no tool arguments or observations, so the objective matches the Planner’s inference interface. Each Planner example uses a single-turn conversation for- mat. The system and user messages form x P and provide the scientific query, input context, skill guidance, and candidate tools. The assistant target serializes π ⋆ as a JSON object containing only the tool_sequence field. We mask prompt tokens, so L plan supervises only the assistant response and focuses adaptation on tool selection and ordering. Parameter-efficient adaptation focuses the Planner’s learning signal on its role-specific workflow decisions. We freeze the pretrained backbone throughout the entire optimization process and update only a Planner-specific low-rank adaptation (LoRA) module [83] to learn EO tool-selection and ordering patterns. At inference, the sequential debate and catalog vali- dation described in Section IV enforce procedure, schema, and active-skill constraints beyond the learned workflow prior. D. Executor GRPO for Runtime Argument Grounding Executor training treats the current workflow node as a local decision interface. Given node context x E i and the schema of prescribed tool t ⋆ i , policy p ψ E generates argument dictionary ˆ θ i at inference and sampled dictionaries ˆ θ i,j during GRPO. The reward evaluates only argument structure and values because the workflow already fixes the tool identity. This node- level interface isolates feedback for argument grounding from workflow decisions at other nodes. The Executor reward compares each parsed response di- rectly with the reference argument dictionary θ ⋆ i and requires no learned reward model. Let parse denote the JSON parser, let ⊥ denote parse failure, let J denote the set of valid JSON values, and let D + ⊂ J denote the set of nonempty dictionaries. For candidate j at node i,ω E i,j denotes the raw token sequence, ξ i,j = parse(ω E i,j ) denotes its parsed value, and we set ˆ θ i,j = ξ i,j when ξ i,j ∈ D + . For a nonempty dictionary, K ⋆ and ˆ K denote the reference and predicted key sets, K c = K ⋆ ∩ ˆ K and K d = K ⋆ △ ˆ K denote the 9 common and discrepant keys, and m = |K ⋆ | > 0. Value- agreement term A val counts type-aware matches over common keys under relation ≡, which recursively compares nested values and their types with a tolerance of 10 −6 for floating- point values. The following equation combines key coverage, a symmetric-difference penalty, and value agreement in S arg , then defines complete reward r i,j . A missing key incurs a larger penalty than a superfluous key because it reduces coverage and increases the symmetric difference. Runtime verification separately determines execution validity: A val = X k∈K c I[ ˆ θ i,j,k ≡ θ ⋆ i,k ], S arg ( ˆ θ i,j ,θ ⋆ i ) = max −1, |K c |−|K d | + A val m , r i,j = −2,ξ i,j =⊥, −1,ξ i,j ∈ J\ D + , S arg ( ˆ θ i,j ,θ ⋆ i ), ξ i,j ∈ D + . (13) For each node input, we sample G = 4 candidate argument dictionaries from the current policy and standardize their rewards within the group. Let r i,j denote the reward for candidate j, let p old denote the policy that generated it, and let ε c denote the GRPO clipping radius. Relative advantage ˆ A i,j supplies a quality signal under the same node context without a separate critic. In Eq. (10), the KL regularizer limits deviation from the reference model, and the entropy bonus discourages premature collapse. During training, we freeze the pretrained backbone and update only the Executor-specific LoRA adapter. The trained policy grounds arguments for the current node, while explicit runtime mechanisms control environment inter- action, observation verification, retries, and workflow repair. The group-relative advantage, token-level importance ratio, and its clipped form define the GRPO surrogate term: ˆ A i,j = r i,j − mean j ′ (r i,j ′ ) std j ′ (r i,j ′ ) + 10 −6 , ρ i,j,ℓ = p ψ E (ω E i,j,ℓ | x E i ,ω E i,j,<ℓ ) p old (ω E i,j,ℓ | x E i ,ω E i,j,<ℓ ) , ̄ρ i,j,ℓ = clip [1−ε c , 1+ε c ] (ρ i,j,ℓ ). (14) VI. EXPERIMENTS We evaluate Earth-Agent-Pro from seven perspectives. We first examine its sensitivity to LLM backbones and com- pare it with remote-sensing MLLMs. We then assess cross- benchmark transfer on the OpenEarthAgent benchmark [25] and compare Earth-Agent-Pro with general-purpose agents. Next, we isolate the effect of the agent framework through controlled comparisons on Earth-Bench-Pro. Finally, we ablate domain skills and disentangle the contributions of Planner SFT and Executor GRPO [82] through adapter composition, planning-only evaluation, and oracle-plan execution. Together, these experiments evaluate the effectiveness and transferability of Earth-Agent-Pro and identify the mechanisms. A. Experimental Setup a) Benchmarks and Comparison Groups: Our main eval- uation is conducted on Earth-Bench-Pro (Section I), which converts Earth-Bench into open-ended, end-to-end Earth- observation tasks that require data acquisition, preprocess- ing, analysis, and evidence-grounded answer generation. We additionally evaluate transfer to the OpenEarthAgent bench- mark [25], compare against remote-sensing MLLMs on the categories used by TerraScope [21], and compare with general- purpose agents across Spectrum, Products, and RGB tasks. b) Baselines and Models: We first evaluate Earth-Agent- Pro with three closed-weight LLM backbones, GPT-5, Gemini- 3.5-flash, and Claude-4.6-sonnet, and open-weight backbones from the GLM, Llama, InternLM, GPT-oss, Qwen, Gemma, and Mistral families. We also include fine-tuned Qwen3.5- 9B and Qwen3.5-4B variants. We conduct this training and all open-weight model evaluations on eight NVIDIA H200 graphics processing units (GPUs), use vLLM for model serving, and query the three closed-weight models through their official APIs. The remote-sensing comparison includes GeoChat [19], TeoChat [84], EarthDial [60], LHRS-Bot [20], EarthMind [85], and TerraScope [21]. For framework-level evaluation on the OpenEarthAgent benchmark [25], we com- pare ReAct [66], AFlow [70], OpenEarthAgent [25], and Earth-Agent-Pro under a shared GPT-5 backbone. We next compare against the general-purpose agents Codex and Manus. Finally, we repeat the framework comparison on Earth-Bench- Pro using the same GPT-5 backbone to more clearly isolate the specific effect of the agent framework itself. c) Metrics: We adopt benchmark-specific evaluation pro- tocols. On Earth-Bench-Pro, we mainly follow the dual-level evaluation protocol of Earth-Agent [31]. Final-answer quality is evaluated using LLM-as-Judge accuracy, ROUGE-L [86], and embedding-based semantic matching (EM), which respec- tively measure judged correctness, lexical overlap, and seman- tic similarity to the reference answer. Execution trajectories are evaluated using Efficiency, Tools-Any-Order (TAO), Tools-In- Order (TIO), Tool-Exact-Match (TEM), and Parameter Ac- curacy. These trajectory metrics collectively characterize tool coverage, tool ordering, strict agreement with the reference workflow, and the correctness of tool arguments. For the OpenEarthAgent benchmark [25], we retain its offi- cial evaluation protocol, including category-wise tool-selection F1, tool-order fidelity, and task-level accuracy. For the final- result comparisons in Tables I and V, we report accuracy only. A non-numerical prediction is counted as correct when it matches the reference answer, while a numerical prediction is considered correct when its relative error with respect to the ground truth is below 15%. All reported accuracy values are computed as the percentage of correct predictions over the evaluated samples for each corresponding comparison. B. Main Results on Earth-Bench-Pro a) Backbone and Fine-Tuning Results: We first exam- ine how the policy backbone affects Earth-Agent-Pro on the open-ended Earth-Bench-Pro tasks. We take the off-the- shelf Qwen3.5-4B and Qwen3.5-9B models as our baselines, enabling direct comparison with our fine-tuned variants at the same model scales. Table I reports final answer quality, tra- jectory length, and step-level agreement with expert workflows under the same unified evaluation protocol. 10 TABLE I PERFORMANCE OF DIFFERENT LLM BACKBONES WITH EARTH-AGENT-PRO ON EARTH-BENCH-PRO. ALL QUALITY METRICS ARE PERCENTAGES. EFFICIENCY IS A RATIO FOR WHICH VALUES CLOSER TO 1 ARE BETTER. AMONG OPEN-SOURCE MODELS, BEST RESULTS ARE BOLDED AND RUNNER-UPS ARE UNDERLINED. Model LLM-as- Judge ROUGE- L EMEfficiencyTAOTIOTEMParam. GPT-566.1321.6813.311.091187.6171.9760.1427.28 Gemini-3.5-flash63.3125.4320.561.075289.2873.9064.4932.19 Claude-4.6-sonnet56.8529.0929.031.191190.1776.5963.5228.02 GLM-4.5-Air48.3920.4520.161.006385.4672.7662.3027.01 Llama-3.1-8B-Instruct13.715.152.020.981175.3460.8144.9522.04 Llama-3.2-3B-Instruct4.442.252.420.687351.0944.2636.6213.96 Llama-3.3-70B-Instruct33.878.696.451.052685.1773.5156.6927.39 Llama-4-Scout32.669.917.660.933581.8170.4757.5725.88 InternLM3-8B19.356.865.240.761769.3358.4948.8520.73 Intern-S140.7315.4913.711.0022 85.4473.6260.2431.26 GPT-oss-20B38.7114.1010.481.151085.9471.0859.5928.62 GPT-oss-120B48.3913.548.871.041586.1671.6261.6129.84 Qwen3.5-27B49.60 19.1215.321.001787.6775.2768.7129.86 Gemma-4 (E4B-it)31.4512.499.680.784169.3361.3655.1323.26 Gemma-4 (26B-A4B-it)47.5818.4914.521.076684.1972.6261.8031.23 MistralAI-v0.3-7B14.116.324.440.587654.3146.1842.1815.61 Qwen3.5-4B32.2616.6114.520.944779.2166.8757.9921.58 Qwen3.5-9B38.3116.3315.731.012985.5974.5767.2925.05 Ours (4B, fine-tuned)43.1522.4221.771.111787.93 77.5268.9628.98 Ours (9B, fine-tuned)50.0020.65 19.761.157191.2881.1171.0132.77 Observation 1: answer and trajectory metrics reveal different model strengths. GPT-5 obtains the highest LLM- as-Judge score (66.13%), whereas the fine-tuned 4B model leads on ROUGE-L and EM. In contrast, the fine-tuned 9B model achieves the strongest TAO, TIO, TEM, and parameter accuracy, showing that final-answer quality and workflow fidelity do not necessarily improve together. Observation 2: role-specialized training consistently strengthens small backbones. Relative to Qwen3.5-9B, the fine-tuned 9B variant improves LLM-as-Judge accuracy from 38.31% to 50.00% and raises parameter accuracy by 7.72 points. The 4B model exhibits similarly broad gains, including increases of 10.89 points in Judge accuracy and 10.97 points in TEM. These results indicate that the training procedure im- proves both answer generation and structured tool execution. C. Comparisons with Existing Methods and Agents a) Remote-Sensing MLLM Comparison: Table I re- ports the performance of Earth-Agent-Pro and representative remote-sensing MLLMs on five TerraScope task categories, including boundary detection, comparative ranking, coverage estimation, area measurement, and distance measurement. The evaluation uses open-ended questions instead of multiple- choice options, requiring models to directly produce semantic or numerical answers without candidate guidance. Observation 1. Remote-sensing MLLMs remain weak on quantitative reasoning. Across boundary detection and comparative ranking, existing remote-sensing MLLMs achieve moderate accuracy, which suggests that these models can often recognize coarse land-cover semantics and compare visible regions. However, their performance drops sharply on coverage estimation, area measurement, and distance measure- ment, indicating that remote-sensing MLLMs still lack reliable metric grounding when exact numerical answers are required. Observation 2. Earth-Agent-Pro achieves the highest overall accuracy. Earth-Agent-Pro obtains the best aver- age accuracy among compared methods, reaching 50.58% with Qwen3.5-9B and 50.41% with GPT-5. Compared with the strongest baseline, TerraScope (42.60%), these results correspond to improvements of 7.98 and 7.81 percentage points, respectively. The advantage is most pronounced in area measurement, where both Earth-Agent-Pro variants achieve 31.58%, compared with at most 4.65% among the remote- sensing MLLMs. Earth-Agent-Pro maintains strong semantic performance, achieving 66.43% on boundary detection and up to 71.35% on comparative ranking. These results demonstrate its effectiveness across both semantic understanding and quan- titative reasoning tasks in open-ended evaluation settings. b) Cross-Benchmark Comparison on OpenEarthAgent: We further evaluate the cross-benchmark transferability of Earth-Agent-Pro on the OpenEarthAgent benchmark [25]. To isolate the effect of the agent framework, we first compare ReAct, AFlow, OpenEarthAgent, and Earth-Agent-Pro using the same GPT-5 backbone and the official OpenEarthAgent benchmark [25] evaluation protocol. We additionally evaluate the role-specialized 9B and 4B variants to examine whether 11 TABLE I COMPARISON WITH GENERAL RS MLLMS ACROSS THE TERRASCOPE TASK CATEGORIES. A NUMERICAL PREDICTION IS COUNTED AS CORRECT IF ITS RELATIVE ERROR FROM THE GROUND-TRUTH VALUE IS LESS THAN 15%. THE AVG. COLUMN REPORTS SAMPLE-WEIGHTED ACCURACY ACROSS THE FIVE TASK CATEGORIES. BEST RESULTS ARE IN BOLD, AND SECOND-BEST ARE UNDERLINED. Model Boundary Detection (855) Comparative Ranking (855) Coverage Estimation (855) Area Measurement (855) Distance Measurement (129) Avg. GeoChat [19]57.5458.8312.160.120.7831.02 TeoChat [84]56.1465.857.250.120.0031.16 EarthDial [60]55.0952.987.020.000.0027.73 LHRS-Bot [20]55.2056.738.071.550.0028.97 EarthMind [85]54.8559.7712.284.650.0030.74 TerraScope [21]66.2069.9440.700.000.0042.60 Earth-Agent-Pro (GPT-5)66.4370.6440.3531.581.5550.41 Earth-Agent-Pro (Qwen3.5-9B) 66.4371.3540.3531.581.5550.58 TABLE IV FRAMEWORK COMPARISON ON THE OPENEARTHAGENT BENCHMARK. WE REPORT TOOL-SELECTION F1 SCORES FOR PERCEPTION (PER.), OPERATION (OP.), LOGIC, AND GIS; TOOL-ORDER FIDELITY UNDER ANYORDER, SAMEORDER, AND UNIQUE; AND TASK-LEVEL ACCURACY FOR NONGENERATIVE ANSWERING (ANS.) AND IMAGE GENERATION (GEN.). BEST RESULTS ARE IN BOLD, AND SECOND-BEST RESULTS ARE UNDERLINED. Framework Tool Selection F1Tool-Order FidelityTask Accuracy Per.Op.LogicGISAnyOr.SameOr.Uni.Ans.Gen. ReAct (GPT-5)40.8766.9017.1888.4146.7146.7148.5939.4580.69 AFlow (GPT-5)32.5467.6936.3086.7551.1150.5654.1745.3865.12 OpenEarthAgent (GPT-5)16.8847.006.2691.5046.9646.7947.8143.8846.21 OpenEarthAgent-4B57.4949.1244.8094.5161.5060.6567.3243.2571.55 Earth-Agent-Pro (GPT-5)62.3055.6835.1595.5964.2364.8070.9756.5474.01 Ours (4B, fine-tuned)56.1663.2342.9295.1270.1369.7775.8746.8672.26 Ours (9B, fine-tuned)57.7044.7944.8096.1270.5868.7073.9547.0672.99 backbones can retain competitive workflow-planning ability. Observation1:Earth-Agent-Proimprovescross- benchmark workflow fidelity. Under the shared GPT-5 back- bone and evaluation protocol, Earth-Agent-Pro achieves the strongest Perception, GIS, and all three tool-order scores. It also improves answering accuracy by 11.16 points over the strongest GPT-5 baseline, indicating that the framework advantage transfers beyond Earth-Bench-Pro. Observation 2: compact fine-tuned models retain strong procedural capabilities. The fine-tuned 9B model leads on GIS and AnyOrder, while the 4B model achieves the best SameOrder and Unique scores. These results suggest that role- specialized training enables small backbones to learn transfer- able workflow patterns across distinct benchmark settings. c) General-Purpose Agent Comparison: We compare Earth-Agent-Pro with Codex and Manus across Spectrum, Products, and RGB tasks. Table V reports both task accuracy and mean wall-clock latency, exposing the trade-off between accuracy and execution cost. Observation 1: a domain-specific workflow improves overall completion. Earth-Agent-Pro (GPT-5) obtains 75.00% average accuracy, 28.33 points above Codex, while also re- ducing average latency from 268.73 s to 208.64 s. Its largest advantage appears on RGB tasks, where it completes all evalu- ated examples. The simultaneous gains in accuracy and latency indicate that improved task completion does not necessarily require higher execution cost. d) Framework Comparison on Earth-Bench-Pro: We now directly isolate the effect of the Earth-Agent-Pro frame- work in a strictly controlled setting by comparing it with ReAct, AFlow, and OpenEarthAgent under the same GPT-5 backbone. Table VI provides evidence for the framework’s effectiveness on our target benchmark. Observation 1: Earth-Agent-Pro leads seven of eight framework-comparison metrics. It improves LLM-as-Judge accuracy by 20.95 points over the strongest baseline and raises TAO from 70.47% to 87.61%. The corresponding gains in TIO, TEM, and parameter accuracy show that the improvement extends beyond selecting a plausible tool set to preserving more of the expert workflow structure. D. Ablation Studies a) Domain-Skill Ablation: We remove the retrieved do- main skill while keeping the framework and model fixed within each pair. Table VII reports the result for ReAct, 12 TABLE V COMPARISON WITH GENERAL-PURPOSE AGENTS. ACCURACY VALUES ARE REPORTED AS PERCENTAGES, AND AVERAGE LATENCY IS MEASURED IN SECONDS. BEST RESULTS ARE BOLDED AND RUNNER-UPS ARE UNDERLINED. Framework Accuracy (%) Avg. Latency (s) SpectrumProductsRGBAvg. Codex40.0075.0025.0046.67268.73 Manus40.0045.0045.0043.33291.00 Earth-Agent-Pro (GPT-5)70.0055.00100.0075.00208.64 Earth-Agent-Pro (Qwen3.5-9B)45.00 40.0065.0050.0067.83 Earth-Agent-Pro (Qwen3.5-9B, fine-tuned)70.0040.0065.0058.33111.79 TABLE VI FRAMEWORK COMPARISON ON EARTH-BENCH-PRO USING A SHARED GPT-5 BACKBONE. ALL QUALITY METRICS ARE REPORTED AS PERCENTAGES. EFFICIENCY IS MEASURED AS A RATIO, WHERE VALUES CLOSER TO 1 INDICATE BETTER PERFORMANCE. BEST RESULTS ARE BOLDED AND RUNNER-UPS ARE UNDERLINED. Framework LLM-as- Judge ROUGE-LEMEfficiencyTAOTIOTEMParam. ReAct45.1819.7526.600.863657.8347.5343.3324.33 AFlow40.1610.8036.700.851170.4745.4637.1820.33 OpenEarthAgent41.5313.8325.000.810160.0642.6239.3222.67 Earth-Agent-Pro66.1321.6813.311.091187.6171.9760.1427.28 TABLE VII ABLATION STUDY OF DOMAIN SKILLS. ALL QUALITY METRICS ARE REPORTED AS PERCENTAGES, WHILE EFFICIENCY IS MEASURED AS A RATIO. BOLD VALUES INDICATE THE BETTER RESULT WITHIN EACH ADJACENT PAIR. FrameworkModelSkills LLM-as- Judge ROUGE-LEMEfficiencyTAOTIOTEMParam. ReActQwen3.5-9B ×34.2713.9710.082.154956.4631.3921.5510.44 ✓39.9214.0011.692.214557.1337.6629.7612.14 Earth-Agent-Pro Qwen3.5-9B ×35.0815.9915.320.995575.6161.9451.7124.84 ✓38.3116.3315.731.012985.5974.5767.2925.05 Ours (9B, fine-tuned) ×37.1015.1114.921.140582.4170.9355.9230.00 ✓50.0020.6519.761.157191.2881.1171.0132.77 Earth-Agent-Pro with the base 9B model, and Earth-Agent- Pro with the fine-tuned 9B model. Bold values indicate the better result within each with/without-skill pair; for Efficiency, better means closer to 1. Observation 1: domain skills improve reported quality metric. The improvement is particularly strong for trajectory fidelity. For the base Earth-Agent-Pro model, adding skills raises TIO from 61.94% to 74.57% and TEM from 51.71% to 67.29%. This pattern is consistent with skills constraining the workflow search space toward expert processing procedures. Observation 2: skills and role-specialized training are complementary. For the fine-tuned model, adding skills im- proves LLM-as-Judge accuracy by 12.90 points, ROUGE-L by 5.54 points, and EM by 4.84 points. It also raises TAO, TIO, TEM, and parameter accuracy by 8.87, 10.18, 15.09, and 2.77 points, respectively. These simultaneous gains across all metrics show that the external skill prior remains effective after role-specialized training. The two components operate at different levels: training strengthens the learned planning and execution policies, whereas skills provide explicit domain procedures that constrain workflow construction. Their com- bination consequently achieves the strongest answer-quality and trajectory-fidelity results, confirming that training does not make external skill guidance redundant. b) Role-Specialized Training Ablation: Finally, we dis- entangle the effects of Planner SFT and Executor GRPO through adapter composition, planning-only evaluation, and oracle-plan execution under carefully controlled component substitutions. Base and Tuned denote the pretrained and role- specialized variants, respectively. Adapter composition. Table VIII evaluates all four com- binations of the Planner and Executor adapters. Bold values indicate the best result and underlined values indicate the runner-up; for Efficiency, values closer to 1 are better. Observation 1: Planner and Executor tuning provide complementary benefits. Planner tuning produces the larger 13 TABLE VIII ADAPTER COMPOSITION FOR ROLE-SPECIALIZED TRAINING ON EARTH-BENCH-PRO. ALL QUALITY METRICS ARE PERCENTAGES, WHILE EFFICIENCY IS A RATIO. PlannerExecutor LLM-as- Judge ROUGE-LEMEfficiencyTAOTIOTEMParam. BaseBase38.3116.3315.731.012985.5974.5767.2925.05 BaseTuned43.1517.3118.150.9865 85.4472.2465.6428.34 TunedBase44.7620.6718.551.183989.4778.7870.0930.33 TunedTuned50.0020.6519.761.157191.2881.1171.0132.77 TABLE IX PLANNING-ONLY EVALUATION ON EARTH-BENCH-PRO. BOLD VALUES INDICATE THE BETTER RESULT WITHIN THE BASE/TUNED PAIR. PlannerTAOTIOTEM GPT-588.8676.3870.03 InternLM3-8B72.1555.3846.87 GEMMA-4 (E4B-it) 83.7571.3461.80 MiniMax-M2.787.9071.8967.66 Base86.7071.8766.96 Tuned92.5183.3576.14 TABLE X ORACLE-PLAN EXECUTOR EVALUATION ON EARTH-BENCH-PRO. EFF. DENOTES EFFICIENCY. BOLD VALUES INDICATE THE BETTER RESULT WITHIN THE BASE/TUNED PAIR. Executor LLM-as- Judge ROUGE-LEMEff.Parm. GPT-568.9524.3118.150.987239.58 InternLM3-8B38.7112.0410.480.862031.13 GEMMA-4 (E4B-it) 33.8712.7410.890.710933.66 MiniMax-M2.751.2119.4418.150.830935.72 Base46.3721.9220.160.908134.19 Tuned52.8225.9223.390.927437.20 gains in workflow metrics, whereas Executor tuning primarily improves final-answer quality and parameter grounding. Their combination achieves the strongest overall Judge, EM, TAO, TIO, TEM, and parameter scores. Observation 2: end-to-end parameter accuracy also depends on planning. The tuned-Planner/base-Executor con- figuration has higher parameter accuracy than the base- Planner/tuned-Executor configuration. This occurs because the end-to-end metric also depends on workflow alignment and execution progress, motivating the isolated evaluations below. Planning-only isolates workflow composition. The planning-only evaluation isolates the Planner’s pre- diction of workflow structure. We run only the four-role sequential debate and compare its output tool sequence directly with the reference workflow. This setting omits tool execu- tion and Executor-side tool replacement, early termination, same-node retries, suffix replanning, and workflow editing. The resulting metrics measure static workflow composition because the planning-only trace retains proposed tools without execution, while the complete system records only verified tool calls. Planning-only scores can therefore exceed end-to-end trajectory scores when execution failures shorten the accepted trace. Observation 3: Planner SFT improves workflow com- position. The tuned Planner reaches 92.51% TAO, 83.35% TIO, and 76.14% TEM, improving over the base Planner by 5.81, 11.48, and 9.18 percentage points, respectively. It also exceeds GPT-5 by 3.65, 6.97, and 6.11 points on the same metrics. These gains isolate improved prediction of the reference tool set and order; end-to-end completion remains subject to subsequent execution. Oracle-plan execution isolates argument grounding. The oracle-plan evaluation isolates the Executor by fixing the workflow structure. We replace the predicted plan with the reference tool order and disable tool replacement, early termination, suffix replanning, and workflow editing. Each node still generates arguments, invokes the specified tool, verifies the returned observation, and retries the same tool when needed. This protocol fixes tool selection and ordering while preserving runtime argument grounding and environ- ment interaction throughout the complete execution process. Observation 4: Executor GRPO improves argument grounding. Relative to the base Executor, the tuned Execu- tor raises LLM-as-Judge accuracy from 46.37% to 52.82%, ROUGE-L from 21.92% to 25.92%, EM from 20.16% to 23.39%, and parameter accuracy from 34.19% to 37.20%. Efficiency also moves from 0.9081 to 0.9274, closer to the ideal value of 1. GPT-5 still reaches 68.95% LLM-as-Judge accuracy and 39.58% parameter accuracy, indicating a remain- ing gap associated with backbone capability. Together with the adapter-composition results in Table VIII, these isolated evaluations show that Planner SFT improves workflow composition and Executor GRPO improves argu- ment grounding under a fixed workflow. The combined system benefits from these distinct interface-specific gains. VII. DISCUSSION A. Qualitative Case Studies We further analyze three representative execution traces covering spectral observations, RGB imagery, and remote- sensing products, as shown in Figure 7. These examples complement the aggregate results by illustrating three factors that determine successful open-world EO execution: workflow completeness, preservation of intermediate dependencies, and correct grounding of tools and arguments. 14 Task1: Urban Heat Island Index Estimation Query: Calculate the Urban Heat Island Index (UHII) over Paris on June 28, 2020, using ASTER TTM-derived LST (Bands 10–12) by comparing mean urban LST (>295K) and rural LST (≤290K). Annotation GT Execution Trace · Ours with skills Step 1 Step 1 · Acquire ASTER Products aster_l1t = search( dataset="ASTER L1T", roi=paris_polygon, date="2020-06-28", bands=[10,11,12] ) Output: paired ASTER thermal bands (B10–B12) Step 2 Step 2 · Band Extraction And Radiometric Calibration bt10 = extract_bt(band10) bt11 = extract_bt(band11) bt12 = extract_bt(band12) Step 3 Step 3 · Compute TTM-derived LST lst = ttm_lst( bt10, bt11, bt12 ) Step 4 Step 4 · Aggregate & calculate UHII urban_mean = mean( lst > 295 K ) rural_mean = mean( lst <= 290 K ) uhii = urban_mean - rural_mean Step 5 Final Urban LST : 295.4807 K Rural LST : 252.1676 K UHII = 43.3131 K Output: land surface temperature (LST) map Output: calibrated brightness temperature (BT10–BT12) Output: urban mean, rural mean, and UHII GT UHII = 43.3131 K Step-by-Step End-to-End TAO 1 TIO 1TEM 1 Parm 0.4 EM 0.7394Efficiency 1 Rouge-L 0.6667 extract TTM mean ×2 43.3131 K Ours Qwen-9B Download ASTER Products Extract TTM bt10 bt11 bt12 LST Urban LST Rural LST UHII GPT-5 Agent Comparison Tool & Parm are correct Correct Tool & Incorrect Parm Incorrect Tool difference download download extract TTM mean ×2 difference 43.3131 K download extract TTM mean ×2 difference 0 K download mean ×2 difference 24.25 K Task2: Tennis Court Area Estimation Query: As part of a city's sports infrastructure audit, estimate the total area occupied by tennis courts. Use aerial imagery and a ground resolution of 0.13 meters per pixel to identify all courts and report the total built-up area. Annotation GT Execution Trace · Ours with skills Step 1 Step 1 · Acquire image files image_files = get_filelist( "benchmark/data/question213" ) Output: Image list obtained (P0377.png). Step 2 Step 2 · Detect tennis courts detections = SM3Det(prompt="tennis court") Step 3 Step 3 · Generate segmentation mask segmentation = SAM2(image, detected_boxes) Step 4 Step 4 · Calculate built-up area area = calculate_area(segmented_image, gsd=0.13) Output: 72.17% (2021-04-23 → 2021-05-09) Step 5 Result: 879.3746 m² Output: Segmented tennis court mask image. Output: Detected tennis court bounding boxes. Output: Total tennis court area = 879.3746 m² About 880 m² Ours Qwen-9B P0377.png GPT-5 Agent Comparison Final SM3Det SAM2 About 880 m² GT = About 880 m² get_filelist SM3Det SAM2 calculate_area 879.3746 m² get_filelist SM3Det calculate_area 1584.74 m² calculate_area tennis court mask Step-by-Step End-to-End TAO 1 TIO 1 TEM 1Parm 0.5 EM 0.3103Efficiency 1 Rouge-L 0 Tool & Parm are correct Correct Tool & Incorrect Parm Incorrect Tool get_filelist SM3Det SAM2 calculate_area Task3: Vegetation Coverage Change Analysis Query: Using the vegetation coverage data of the Northeast Plain from April 1 to August 31, 2021, calculate the percentage change in vegetation coverage between consecutive dates, and report the maximum percentage increase. Annotation GT Execution Trace · GPT-5 with skills Step 1 Step 1 · Acquire MODIS NDVI products download_product(collection="MODIS/061/MOD13Q1", band="NDVI", region="Northeast Plain", time="2021-04-01 ~ 2021-08-29") Output: NDVI time series (10 images) Step 2 Step 2 · Radiometric Calibration apply_scale_offset(scale=0.0001) Step 3 Step 3 · Compute regional mean calc_batch_image_mean(images=NDVI_series) Step 4 Step 4 · Compute percentage changes percentage_change(values=mean_series) Output: 72.17% (2021-04-23 → 2021-05-09) Step 5 max_value_and_index(values=change_series) Output: mean vegetation coverage for each date Output: calibrated NDVI images Output: consecutive percentage changes GT = 72.1% 72.1% Ours Qwen-9B Download MODIS-NDVI products percentage_change GPT-5 Agent Comparison Step 5 · Find maximum increase Output: 72.17% (2021-04-23 → 2021-05-09) calibrate mean find maximum increase 72.1% Step-by-Step End-to-End TAO 1 TIO 1 TEM 1 Parm 0.2 EM 0.2483Efficiency 1 Rouge-L 0 download calibrate mean difference max_value Tool & Parm are correct Correct Tool & Incorrect Parm Incorrect Tool download calibrate mean difference max_value download calibrate mean percentage max_value download calibrate mean percentage max_value 72.17% 14.1% 5.49% TTM get_filelist SM3Det calculate_area 1584.74 m² Fig. 7. Representative execution cases on Earth-Bench-Pro. From top to bottom: Spectrum-based UHII estimation, RGB-based tennis-court area estimation, and Product-based vegetation-coverage change analysis. The first two cases show that our agent preserves the required tool dependencies and matches the reference answers, whereas the third illustrates errors caused by incorrect downstream tool selection. In the Spectrum case, the agent must acquire ASTER prod- ucts, extract Bands 10–12, retrieve land-surface temperature using TTM, aggregate the urban and rural temperatures, and calculate their difference. GPT-5 skips the band extraction stage and applies the operations to incorrectly prepared in- puts, producing 24.25 K. Qwen-9B includes extraction but incorrectly grounds the thermal-processing and aggregation operations, resulting in 0 K. In contrast, our agent preserves the executable dependency chain and recovers the reference UHII of 43.3131 K. This case shows that selecting plausible tools is insufficient when their inputs and arguments are not grounded in the preceding observations. The RGB case further demonstrates the importance of intermediate representations. GPT-5 and Qwen-9B detect the tennis courts but proceed directly from the detections to area calculation. Without converting the coarse detections into a pixel-level mask, both agents overestimate the occupied area as 1584.74 m 2 . Our agent inserts SAM2 segmentation between detection and measurement and calculates the area from the resulting mask using the specified ground resolution. It there- fore obtains 879.3746 m 2 , consistent with the reference answer of approximately 880 m 2 . The difference arises not from the final measurement tool alone, but from whether the correct intermediate artifact is constructed and passed downstream. The Product case exposes a limitation. All three agents perform MODIS acquisition, radiometric calibration, and regional-mean calculation. GPT-5 follows the reference change-analysis and maximum-value operations, obtaining 72.17%, close to the reference value of 72.1%. Qwen-9B and our agent instead select incorrect downstream operations, pro- ducing 14.1% and 5.49%, respectively. Thus, correct upstream data preparation does not guarantee a correct answer when the scientific meaning of a downstream operation is mismatched. Taken together, these cases show that open-world EO execution requires more than individually reasonable tool calls. Reliable quantitative answers depend on maintaining the complete workflow, constructing the intended intermediate representations, and grounding every downstream operation in both the accepted execution evidence and the scientific objective. The Product failure also indicates that skill guidance improves workflow reliability but does not fully eliminate semantic tool-selection errors. 15 B. Limitations and Trade-offs Open-world EO execution remains limited by exact work- flow fidelity and numerical grounding. Even the strongest TEM and parameter scores in Table I are 71.01% and 32.77%, respectively. The gap between reliable tool-set coverage and much lower parameter accuracy suggests that selecting rele- vant tools is easier than instantiating a fully faithful executable workflow. Open-ended QA removes answer candidates and requires the model or agent to directly produce a calibrated numerical value, which more closely reflects real EO requests for measurements. It also requires target localization, metric reasoning, unit handling, and correct answer formatting. The zero and near-zero Distance accuracy across most methods in Table I shows that precise open-ended spatial measurement remains an unresolved challenge for both remote-sensing MLLMs and current agent systems. Workflow fidelity and final-task accuracy remain distinct, and capability profiles vary across models and modalities. In Table IV, AFlow leads Operation F1 and ReAct leads image-generation accuracy. The fine-tuned 4B and 9B models achieve higher tool-order fidelity than Earth-Agent-Pro (GPT- 5) but remain below it in answering accuracy, suggesting that final-answer synthesis remains more dependent on backbone capability. On Product tasks in Table V, Codex reaches 75.00% accuracy, compared with 55.00% for the strongest Earth- Agent-Pro variant. This modality-specific result limits any claim of uniform superiority across all task families. Accuracy improvements also incur additional execution cost. Table V shows that fine-tuning the 9B variant raises average accuracy from 50.00% to 58.33% and latency from 67.83 s to 111.79 s. The fine-tuned small model remains faster than Earth-Agent-Pro (GPT-5), whose latency is 208.64 s, while closing part of the accuracy gap. In Table VI, Earth- Agent-Pro achieves an Efficiency ratio of 1.0911, whose distance to the ideal value of 1 is smaller than that of ReAct, AFlow, or OpenEarthAgent. Its trajectories are slightly longer than the reference on average, whereas the baselines tend to terminate with fewer calls. Table VII further shows that adding skills moves Efficiency farther from 1 in all three ablations, increasing the absolute deviation by 0.0596 for ReAct, 0.0084 for the base Earth-Agent-Pro model, and 0.0166 for the fine- tuned model. This modest length overhead accompanies con- sistent improvements in answer quality and trajectory fidelity and may reflect additional domain-recommended processing or verification steps, exposing a quality–length trade-off. C. Error Analysis Figure 8 shows how training, skill guidance, and framework design affect operational reliability. In Panel A1, role-specific training reduces file-hallucination events from nine to seven, while our fine-tuned Qwen3.5-9B model matches GPT-5 at nine logged invalid-parameter events. Panel A2 shows the backbone gap: Mistral-7B and Llama-3.2-3B retain 138 and 280 errors under WS, dominated by tool hallucinations and invalid parameters. Panel B shows that skills eliminate logged tool-hallucination events in both Qwen3.5-9B configurations and reduce total errors by 36.0% for the fine-tuned model and Error Analysis Unaware of termination conditionsTool hallucination File hallucinationInvalid parameters 0 20 40 60 0 100 200 300 A1 · MODELS WITH SKILLS · SCALE 0–60A2 · SCALE 0–300 9 GPT-5 16 Ours 9B 18 Qwen3.5-9B (baseline) 34 Llama-4-Scout 44 GPT-oss-20B 138 Mistral-7B 280 Llama-3.2-3B 0 15 30 45 60 B · SKILLS EFFECT 25 NS 16 WS 46 NS 18 WS Ours 9BQwen3.5-9B baseline 0 100 200 300 400 500 C · FRAMEWORK × SKILLS 46 NS 18 WS 405 NS 349 WS Earth-Agent-ProReAct Fig. 8. Skills reduce operational error counts. Each run contains 248 trajectories. Bars count logged operational errors, and one trajectory may contribute multiple events. Panel A compares with-skills (WS) counts across models, while Panel B compares no-skills (NS) with WS for the two 9B configurations. Panel C compares Earth-Agent-Pro and ReAct under the shared Qwen3.5-9B backbone. Colors denote repeated no-progress calls until the step limit, calls to nonexistent tools, accesses to nonexistent files or paths, and parameter schema, format, or value violations. 60.9% for the base model. The larger base-model gain shows that skills transfer across configurations and remain effective without role-specific training. Panel C holds Qwen3.5-9B fixed and shows that Earth-Agent-Pro records fewer events than ReAct in every error category under NS and WS, lowering the total by 88.6% and 94.8%, respectively. VIII. CONCLUSION This work studies open-world EO execution, where an agent must locate runtime inputs, acquire observations when required, prepare data, perform domain computations, and generate an open-ended answer from runtime evidence given a high-level scientific question. Earth-Bench-Pro instantiates 248 expert-curated task cores as 744 matched questions, supporting comparisons across instruction following, autonomous plan- ning, and open-world execution through executable evidence chains. Earth-Agent-Pro connects workflow planning with runtime operation grounding through expert skill guidance, workflow-centered structured memory, and localized suffix repair. Role-specialized training uses sequence-level SFT for Planner workflow composition and node-level GRPO for Executor argument grounding. On Earth-Bench-OW, Earth- Agent-Pro leads seven of eight framework-comparison metrics under a shared GPT-5 backbone and reaches 66.13% LLM- as-Judge accuracy. Adapter composition and isolated evalua- tions record higher workflow-composition scores for the tuned Planner and higher answer and parameter scores for the tuned Executor under a fixed workflow. The experiments also expose remaining constraints for open-world EO agents. Even the strongest configurations achieve higher tool coverage than exact workflow matching and parameter accuracy, while final-answer quality remains sensitive to backbone capability. Future work should improve argument grounding under runtime uncertainty, reduce the trajectory overhead introduced by skill-guided execution, and strengthen evidence synthesis in compact models. 16 REFERENCES [1] J. Transon, R. d’Andrimont, A. Maugnard, and P. Defourny, “Survey of hyperspectral earth observation applications from space in the sentinel-2 context,”RemoteSensing, vol. 10, no. 2, p. 157, 2018. [2] J. Ye, J. He, X. Zhang, Y. Lin, H. Lin, C. He, and W. Li, “Satellite image synthesis from street view with fine-grained spatial textual guidance: A novel framework,”IEEEGeoscienceandRemoteSensingMagazine, vol. 13, no. 3, p. 395–414, 2025. [3] K. Anderson, B. Ryan, W. Sonntag, A. Kavvada, and L. Friedl, “Earth observation in service of the 2030 agenda for sustainable development,” Geo-spatialInformationScience, vol. 20, no. 2, p. 77–96, 2017. [4] B. Zhou, H. Yang, D. Chen, J. Ye, T. Bai, J. Yu, S. Zhang, D. Lin, C. He, and W. Li, “Urbench: A comprehensive benchmark for evaluating large multimodal models in multi-view urban scenarios,” inProceedingsof theAAAIConferenceonArtificialIntelligence, vol. 39, no. 10, 2025, p. 10 707–10 715. [5] I. P. Kokkoris, B. Smets, L. Hein, G. Mallinis, M. Buchhorn, S. Balbi, J. ˇ Cerneck ` y, M. Paganini, and P. Dimopoulos, “The role of earth observation in ecosystem accounting: A review of advances, challenges and future directions,”EcosystemServices, vol. 70, p. 101659, 2024. [6] W. Li, J. He, J. Ye, H. Zhong, Z. Zheng, Z. Huang, D. Lin, and C. He, “Crossviewdiff: A cross-view diffusion model for satellite-to-street view synthesis,” arXivpreprintarXiv:2408.14765, 2024. [7] J. Li, J. He, W. Li, J. Chen, and J. Yu, “Roadcorrector: A structure-aware road extraction method for road connectivity and topology correction,” IEEETransactionsonGeoscienceandRemoteSensing, vol. 62, p. 1–18, 2024. [8] C. F. Brown, M. R. Kazmierski, V. J. Pasquarella, W. J. Ruck- lidge, M. Samsikova, C. Zhang, E. Shelhamer, E. Lahera, O. Wiles, S. Ilyushchenkoetal., “Alphaearth foundations: An embedding field model for accurate and efficient global mapping from sparse label data,” arXivpreprintarXiv:2507.22291, 2025. [9] J. Ye, J. He, W. Li, Z. Lv, Y. Lin, J. Yu, H. Yang, and C. He, “Leveraging BEV paradigm for ground-to-aerial image synthesis,” in ProceedingsoftheIEEE/CVFInternationalConferenceon ComputerVision, 2025, p. 28 451–28 461. [Online]. Available: https://openaccess.thecvf.com/content/ICCV2025/html/YeLeveraging BEVParadigmforGround-to-AerialImageSynthesisICCV2025 paper.html [10] X. Liu, H. Zhai, Y. Shen, B. Lou, C. Jiang, T. Li, S. B. Hussain, and G. Shen, “Large-scale crop mapping from multisource remote sensing images in google earth engine,”IEEEJournalofSelectedTopicsin AppliedEarthObservationsandRemoteSensing, vol. 13, p. 414–427, 2020. [11] J. Ye, Q. Luo, J. Yu, H. Zhong, Z. Zheng, C. He, and W. Li, “Sg- bev: Satellite-guided bev fusion for cross-view semantic segmenta- tion,” in2024IEEE/CVFConferenceonComputerVisionandPattern Recognition(CVPR). IEEE, 2024, p. 27 748–27 757. [12] W. Li, H. Yang, Z. Hu, J. Zheng, G.-S. Xia, and C. He, “3D building reconstruction from monocular remote sensing images with multi-level supervisions,” inProceedingsoftheIEEE/CVF ConferenceonComputerVisionandPatternRecognition(CVPR). IEEE, Jun. 2024, p. 27 728–27 737. [Online]. Available: https: //doi.org/10.1109/CVPR52733.2024.02619 [13] W. Li, W. Zhao, J. Yu, J. Zheng, C. He, H. Fu, and D. Lin, “Joint semantic–geometric learning for polygonal building segmentation from high-resolution remote sensing images,” ISPRSJournalof PhotogrammetryandRemoteSensing, vol. 201, p. 26–37, Jul. 2023. [Online]. Available: https://doi.org/10.1016/j.isprsjprs.2023.05.010 [14] W. Li, R. Dong, H. Fu, J. Wang, L. Yu, and P. Gong, “Integrating Google Earth imagery with Landsat data to improve 30-m resolution land cover mapping,” RemoteSensingofEnvironment, vol. 237, p. 111563, Feb. 2020. [Online]. Available: https://doi.org/10.1016/j.rse.2019.111563 [15] W. Li, Y. Lai, L. Xu, Y. Xiangli, J. Yu, C. He, G.-S. Xia, and D. Lin, “OmniCity: Omnipotent city understanding with multi- level and multi-view images,” in ProceedingsoftheIEEE/CVF ConferenceonComputerVisionandPatternRecognition(CVPR). IEEE, Jun. 2023, p. 17 397–17 407. [Online]. Available: https: //doi.org/10.1109/CVPR52729.2023.01669 [16] M. Sudmanns, D. Tiede, S. Lang, and A. Baraldi, “Semantic and syntactic interoperability in online processing of big earth observation data,” InternationalJournalofDigitalEarth, vol. 11, no. 1, p. 95–112, 2018, pMID: 29387171. [Online]. Available: https://doi.org/10.1080/17538947.2017.1332112 [17] K. Weise, R. H ̈ ofer, J. Franke, A. Guelmami, W. Simonson, J. Muro, B. O’Connor, A. Strauch, S. Flink, J. Eberle, E. Mino, S. Thulin, P. Philipson, E. van Valkengoed, J. Truckenbrodt, F. Zander, A. S ́ anchez, C. Schr ̈ oder, F. Thonfeld, E. Fitoka, E. Scott, M. Ling, M. Schwarz, I. Kunz, G. Th ̈ urmer, A. Plasmeijer, and L. Hilarides, “Wetland extent tools for sdg 6.6.1 reporting from the satellite-based wetland observation service (swos),”RemoteSensing ofEnvironment, vol. 247, p. 111892, 2020. [Online]. Available: https://w.sciencedirect.com/science/article/pii/S0034425720302625 [18] Y. Feng, P. Zhang, G. Xiao, L. Ding, and L. Meng, “Towards a barrier-free geoqa portal: Natural language interaction with geospatial data using multi-agent llms and semantic search,”International JournalofAppliedEarthObservationandGeoinformation, vol. 144, p. 104825, 2025. [Online]. Available: https://w.sciencedirect.com/ science/article/pii/S1569843225004728 [19] K. Kuckreja, M. S. Danish, M. Naseer, A. Das, S. Khan, and F. S. Khan, “Geochat: Grounded large vision-language model for remote sensing,” inProceedingsoftheIEEE/CVFConferenceonComputerVisionand PatternRecognition, 2024, p. 27 831–27 840. [20] D. Muhtar, Z. Li, F. Gu, X. Zhang, and P. Xiao, “Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model,” inEuropeanConferenceonComputerVision. Springer, 2024, p. 440– 457. [21] Y. Shu, B. Ren, Z. Xiong, X. X. Zhu, B. Demir, N. Sebe, and P. Rota, “Terrascope: Pixel-grounded visual reasoning for earth observation,” in ProceedingsoftheIEEE/CVFConferenceonComputerVisionand PatternRecognition, 2026, p. 16 712–16 722. [22] A. Shabbir, M. A. Munir, A. Dudhane, M. U. Sheikh, M. H. Khan, P. Fraccaro, J. B. Moreno, F. S. Khan, and S. Khan, “Thinkgeo: Evaluating tool-augmented agents for remote sensing tasks,”arXiv preprintarXiv:2505.23752, 2025. [23] C. H. Kao, W. Zhao, C. Lam, A. Umap, S. Revankar, S. Speas, S. Bhagat, R. Datta, C. P. Phoo, U. Malletal., “Towards llm agents for earth observation,” inFindingsoftheAssociationforComputational Linguistics:ACL2026, 2026, p. 2597–2611. [24] K. Li, J. Wang, Z. Wang, H. Qiao, W. Zhang, D. Meng, and X. Cao, “Designing domain-specific agents via hierarchical task abstraction mechanism,”arXivpreprintarXiv:2511.17198, 2025. [25] A. Shabbir, M. U. Sheikh, M. A. Munir, H. Debary, M. Fiaz, M. Z. Zaheer, P. Fraccaro, F. S. Khan, M. H. Khan, X. X. Zhuetal., “Openearthagent: A unified framework for tool-augmented geospatial agents,”arXivpreprintarXiv:2602.17665, 2026. [26] C. Dang, S. Xiong, C. He, and W. Li, “Skiller: Language-level reinforce- ment learning for reusable skill extraction in small language models,” arXivpreprintarXiv:2608.10538, 2026. [27] S. Lobry, D. Marcos, J. Murray, and D. Tuia, “Rsvqa: Visual question answering for remote sensing data,”IEEETransactionsonGeoscience andRemoteSensing, vol. 58, no. 12, p. 8555–8566, 2020. [28] X. Li, J. Ding, and M. Elhoseiny, “Vrsbench: A versatile vision-language benchmark dataset for remote sensing image understanding,”Advances inNeuralInformationProcessingSystems, vol. 37, p. 3229–3242, 2024. [29] V. Krechetova and D. Kochedykov, “Geobenchx: Benchmarking llms in agent solving multistep geospatial tasks,” inProceedingsofthe1st ACMSIGSPATIALInternationalWorkshoponGenerativeandAgentic AIforMulti-ModalitySpace-TimeIntelligence, 2025, p. 27–35. [30] S. Zhao, F. Liu, X. Zhang, H. Chen, X. Gu, Z. Jiang, F. Ling, B. Fei, W. Zhang, J. Wang etal., “Openearth-agent: From tool calling to tool creation for open-environment earth observation,”arXivpreprint arXiv:2603.22148, 2026. [31] P. Feng, Z. Lv, J. Ye, X. Wang, X. Huo, J. Yu, W. Xu, W. Zhang, L. Bai, C. He, and W. Li, “Earth-Agent: Unlocking the full landscape of earth observation with agents,” in International ConferenceonLearningRepresentations, 2026. [Online]. Available: https://openreview.net/forum?id=dkIXAbWuxO [32] F. Wang, M. Chen, X. He, Y.-F. Zhang, Y. Li, F. Liu, Z. Guo, Z. Hu, J. Wang, J. Xu etal., “Omniearth-bench: Towards holistic evaluation of earth’s six spheres and cross-spheres interactions with multimodal observational earth data,” arXivpreprintarXiv:2505.23522, 2025. [33] M. Danish, M. A. Munir, S. R. A. Shah, K. Kuckreja, F. S. Khan, P. Fraccaro, A. Lacoste, and S. Khan, “Geobench-vlm: Benchmarking vision-language models for geospatial tasks,” in Proceedingsofthe IEEE/CVFInternationalConferenceonComputerVision, 2025, p. 7132–7142. [34] X. An, J. Sun, Z. Gui, and W. He, “Choice: benchmarking the remote sensing capabilities of large vision-language models,” Advancesin NeuralInformationProcessingSystems, vol. 38, 2026. 17 [35] C. Zhang and S. Wang, “Good at captioning bad at counting: Bench- marking gpt-4v on earth observation data,” inProceedingsofthe IEEE/CVFConferenceonComputerVisionandPatternRecognition, 2024, p. 7839–7849. [36] F. Wang, H. Wang, Z. Guo, D. Wang, Y. Wang, M. Chen, Q. Ma, L. Lan, W. Yang, J. Zhangetal., “Xlrs-bench: Could your multimodal llms understand extremely large ultra-high-resolution remote sensing imagery?” inProceedingsoftheIEEE/CVFConferenceonComputer VisionandPatternRecognition, 2025, p. 14 325–14 336. [37] J. Wang, W. Xuan, H. Qi, Z. Liu, K. Liu, Y. Wu, H. Chen, J. Song, J. Xia, Z. Zheng etal., “Disasterm3: A remote sensing vision-language dataset for disaster damage assessment and response,” Advancesin NeuralInformationProcessingSystems, vol. 38, 2026. [38] Z. Liu, D. Zhao, B. Yuan, and Z. Jiang, “Rescueadi: Adaptive disaster interpretation in remote sensing images with autonomous agents,”IEEE TransactionsonGeoscienceandRemoteSensing, vol. 63, p. 1–14, 2025. [39] S. Singh, M. Fore, and D. Stamoulis, “Geollm-engine: A realistic environment for building geospatial copilots,” in2024IEEE/CVF ConferenceonComputerVisionandPatternRecognitionWorkshops (CVPRW). IEEE, 2024, p. 585–594. [40] L. Wang, Y. Xiang, H. Huang, D. Li, C. Gao, and S. Liu, “To- wards realistic earth-observation constellation scheduling: Benchmark and methodology,” AdvancesinNeuralInformationProcessingSystems, vol. 38, p. 85 923–85 944, 2026. [41] S. Ma, Z. Li, S. Li, X. Xu, R. Zhu, T. Boston, and J. A. Taylor, “Eo-gym: A multimodal, interactive environment for earth observation agents,” arXivpreprintarXiv:2605.01250, 2026. [42] D. T. Nguyen, T. Nguyen, F. A. Maani, H. M. Le, M. U. Sheikh, N. Saeed, M. H. Khan, and S. Khan, “Terrabench: Can agents reason over heterogeneous earth-system data?”arXivpreprint arXiv:2606.13148, 2026. [43] A. Xiao, S. Cheng, Y. Xu, Y. Ren, H. Chen, and N. Yokoya, “Geomm- bench and geommagent: Toward expert-level multimodal intelligence in geoscience and remote sensing,” inProceedingsoftheIEEE/CVF ConferenceonComputerVisionandPatternRecognition, 2026, p. 34 843–34 853. [44] Y. Yan, L. Mou, B. Yang, and Q. Li, “Terralogic: A benchmark for hierarchical geospatial reasoning in earth observation,”arXivpreprint arXiv:2607.12497, 2026. [45] B. Yu, C. Yang, D. Hou, C. Liu, J. Liu, C. Wang, Z. Zhang, H. Li, and W. Yang, “Geoagentbench: A dynamic execution benchmark for tool- augmented agents in spatial analysis,”arXivpreprintarXiv:2604.13888, 2026. [46] J. Wang, W. Xuan, H. Qi, P. Dai, K. Liu, H. Chen, Z. Zheng, J. Xia, S. Ermon, and N. Yokoya, “Can LLM Agents Respond to Disasters? Benchmarking Heterogeneous Geospatial Reasoning in Emergency Op- erations,”arXivpreprintarXiv:2605.11633, 2026. [47] M. Yu, Y. Sun, C. Yi, Y. Wen, and H. Yang, “Bringing agentic search to earth observation data discovery,”arXivpreprintarXiv:2607.02387, 2026. [48] D. Hong, B. Zhang, X. Li, Y. Li, C. Li, J. Yao, N. Yokoya, H. Li, P. Ghamisi, X. Jia etal., “Spectralgpt: Spectral remote sensing foun- dation model,”IEEETransactionsonPatternAnalysisandMachine Intelligence, vol. 46, no. 8, p. 5227–5244, 2024. [49] W. Diao, H. Yu, K. Kang, T. Ling, D. Liu, Y. Feng, H. Bi, L. Ren, X. Li, Y. Mao etal., “Ringmo-aerial: An aerial remote sensing foun- dation model with affine transformation contrastive learning,”IEEE TransactionsonPatternAnalysisandMachineIntelligence, 2025. [50] W. Chen, L. Bruzzone, B. Dang, Y. Gao, Y. Deng, J.-G. Yu, L. Yuan, and Y. Li, “Rest: Holistic learning for end-to-end semantic segmentation of whole-scene remote sensing imagery,”IEEETransactionsonPattern AnalysisandMachineIntelligence, 2025. [51] H. Bi, Y. Feng, B. Tong, M. Wang, H. Yu, Y. Mao, H. Chang, W. Diao, P. Wang, Y. Yuetal., “Ringmoe: Mixture-of-modality-experts multi-modal foundation models for universal remote sensing image interpretation,” IEEETransactionsonPatternAnalysisandMachine Intelligence, 2025. [52] Z. Gong, Z. Wei, D. Wang, X. Hu, X. Ma, H. Chen, Y. Jia, Y. Deng, Z. Ji, X. Zhuetal., “Crossearth: Geospatial vision foundation model for domain generalizable remote sensing semantic segmentation,”IEEE TransactionsonPatternAnalysisandMachineIntelligence, 2025. [53] D. Wang, M. Hu, Y. Jin, Y. Miao, J. Yang, Y. Xu, X. Qin, J. Ma, L. Sun, C. Lietal., “Hypersigma: Hyperspectral intelligence comprehension foundation model,”IEEETransactionsonPatternAnalysisandMachine Intelligence, 2025. [54] L. Zhang, Y. Zhao, R. Dong, J. Zhang, S. Yuan, S. Cao, M. Chen, J. Zheng, W. Li, W. Zhangetal., “A 2-mae: A spatial-temporal-spectral unified remote sensing pre-training method based on anchor-aware masked autoencoder,”IEEETransactionsonGeoscienceandRemote Sensing, 2026. [55] Y. Hu, J. Yuan, C. Wen, X. Lu, Y. Liu, and X. Li, “Rsgpt: A remote sensing vision language model and benchmark,”ISPRSJournalof PhotogrammetryandRemoteSensing, vol. 224, p. 272–286, 2025. [56] J. Luo, Z. Pang, Y. Zhang, T. Wang, L. Wang, B. Dang, J. Lao, J. Wang, J. Chen, Y. Tan etal., “Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding,” arXivpreprintarXiv:2406.10100, 2024. [57] W. Zhang, M. Cai, T. Zhang, Y. Zhuang, and X. Mao, “Earthgpt: A universal multimodal large language model for multisensor im- age comprehension in remote sensing domain,” IEEETransactionson GeoscienceandRemoteSensing, vol. 62, p. 1–20, 2024. [58] W. Zhang, M. Cai, T. Zhang, Y. Zhuang, J. Li, and X. Mao, “Earth- marker: A visual prompting multimodal large language model for remote sensing,”IEEETransactionsonGeoscienceandRemoteSensing, vol. 63, p. 1–19, 2024. [59] W. Zhang, M. Cai, Y. Ning, T. Zhang, Y. Zhuang, S. Lu, H. Chen, J. Li, and X. Mao, “EarthGPT-X: A spatial MLLM for multilevel multisource remote sensing imagery understanding with visual prompting,” IEEE TransactionsonGeoscienceandRemoteSensing, vol. 63, p. 1–21, 2025. [Online]. Available: https://doi.org/10.1109/TGRS.2025.3626941 [60] S. Soni, A. Dudhane, H. Debary, M. Fiaz, M. A. Munir, M. S. Danish, P. Fraccaro, C. D. Watson, L. J. Klein, F. S. Khan etal., “Earthdial: Turning multi-sensory earth observations to interactive dia- logues,” inProceedingsoftheComputerVisionandPatternRecognition Conference, 2025, p. 14 303–14 313. [61] C. Pang, X. Weng, J. Wu, J. Li, Y. Liu, J. Sun, W. Li, S. Wang, L. Feng, G.-S. Xiaetal., “Vhm: Versatile and honest vision language model for remote sensing image analysis,” inProceedingsoftheAAAIConference onArtificialIntelligence, vol. 39, no. 6, 2025, p. 6381–6388. [62] A. Shabbir, M. Zumri, M. Bennamoun, F. S. Khan, and S. Khan, “Geopixel: Pixel grounding large multimodal model in remote sensing,” ProceedingsoftheInternationalConferenceonMachineLearning, 2025. [63] Y. Zhan, Z. Xiong, and Y. Yuan, “Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model,” ISPRSJournalofPhotogrammetryandRemoteSensing, vol. 221, p. 64–77, 2025. [64] J. Ye, H. Lin, L. Ou, D. Chen, Z. Wang, Q. Zhu, C. He, and W. Li, “Where am i? cross-view geo-localization with natural language descriptions,” in2025IEEE/CVFInternationalConferenceonComputer Vision(ICCV). IEEE, 2025, p. 5890–5900. [65] J. Ye, Z. Lv, W. Li, J. Yu, H. Yang, H. Zhong, and C. He, “Cross- view image geo-localization with panorama-bev co-retrieval network,” inEuropeanConferenceonComputerVision. Springer, 2024, p. 74– 90. [66] S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, “ReAct: Synergizing reasoning and acting in language models,” in InternationalConferenceonLearningRepresentations, 2023. [Online]. Available: https://openreview.net/forum?id=WE vluYUL-X [67] S. G. Patil, T. Zhang, X. Wang, and J. E. Gonzalez, “Gorilla: Large language model connected with massive apis,”AdvancesinNeural InformationProcessingSystems, vol. 37, p. 126 544–126 565, 2024. [68] Y. Qin, S. Liang, Y. Ye, K. Zhu, L. Yan, Y. Lu, Y. Lin, X. Cong, X. Tang, B. Qianetal., “Toolllm: Facilitating large language models to master 16000+ real-world apis,” in InternationalConferenceonLearning Representations, vol. 2024, 2024, p. 9695–9717. [69] G. Wang, Y. Xie, Y. Jiang, A. Mandlekar, C. Xiao, Y. Zhu, L. Fan, and A. Anandkumar, “Voyager: An open-ended embodied agent with large language models,” TransactionsonMachineLearningResearch, 2024. [Online]. Available: https://openreview.net/forum?id=ehfRiF0R3a [70] J. Zhang, J. Xiang, Z. Yu, F. Teng, X. Chen, J. Chen, M. Zhuge, X. Cheng, S. Hong, J. Wangetal., “Aflow: Automating agentic workflow generation,” in InternationalConferenceonLearningRepresentations, vol. 2025, 2025, p. 34 040–34 077. [71] H. Guo, X. Su, C. Wu, B. Du, L. Zhang, and D. Li, “Remote sensing chatgpt: Solving remote sensing tasks with chatgpt and visual models,” in IGARSS2024–2024IEEEInternationalGeoscienceandRemote SensingSymposium. IEEE, 2024, p. 11 474–11 478. [72] W. Xu, Z. Yu, B. Mu, J. Wang, Z. Wei, and M. Peng, “RS-Agent: Automating remote sensing tasks through intelligent agent,”Science ChinaInformationSciences, vol. 69, no. 8, p. 180302, 2026. [Online]. Available: https://doi.org/10.1007/s11432-026-5026-5 18 [73] C. Liu, K. Chen, H. Zhang, Z. Qi, Z. Zou, and Z. Shi, “Change-agent: Toward interactive comprehensive remote sensing change interpretation and analysis,”IEEETransactionsonGeoscienceandRemoteSensing, vol. 62, p. 1–16, 2024. [74] Z. Chen, H. Wang, J. Yao, J. Zhang, P. Ghamisi, J. Zhou, P. M. Atkinson, and B. Zhang, “Cangling-knowflow: A unified knowledge- and-flow-fused agent for comprehensive remote sensing applications,” arXivpreprintarXiv:2512.15231, 2025. [75] H. Hu, P. Wang, Y. Feng, K. Wei, W. Yin, W. Diao, M. Wang, H. Bi, K. Kang, T. Ling, K. Fu, and X. Sun, “RingMo-Agent: A unified remote sensing foundation model for multi-platform and multi-modal reasoning,”IEEETransactionsonPatternAnalysis andMachineIntelligence, 2026, early Access. [Online]. Available: https://doi.org/10.1109/TPAMI.2026.3718699 [76] Z. Guo, J. Wang, X. Yue, W. Wei, Z. Jiang, W. Xu, B. Fei, W. Zhang, X. Gu, L. Chengetal., “Earthlink: A self-evolving ai agent for climate science,”arXivpreprintarXiv:2507.17311, 2025. [77] A. Bhattaram, J. Chung, S. Chung, R. Gupta, J. Ramamoorthy, K. Gul- lapalli, D. Marculescu, and D. Stamoulis, “Geoflow: Agentic work- flow automation for geospatial tasks,” in Proceedingsofthe33rd ACMInternationalConferenceonAdvancesinGeographicInformation Systems, 2025, p. 1150–1153. [78] P. Dai, W. Xuan, J. Wang, H. Chen, J. Song, Y. Ou, and N. Yokoya, “Experience-driven multi-agent systems are training-free context-aware earth observers,”arXivpreprintarXiv:2602.02559, 2026. [79] L. Yao, S. Xu, F. Liu, C. Zhang, B. Yao, R. Min, Y. Li, C. Ouyang, S. Di, and M.-L. Zhang, “Remoteagent: Bridging vague human intents and earth observation with rl-based agentic mllms,” arXivpreprint arXiv:2604.07765, 2026. [80] Q. Li, Y. Zhang, Z. Mai, Y. Chen, H. Huang, J. Zhang, Z. Zhang, Y. Wen, W. Li, H. Fuetal., “Can large multimodal models understand agricultural scenes? benchmarking with agromind,”AdvancesinNeural InformationProcessingSystems, vol. 38, 2026. [81] J. Zheng, S. Yuan, W. Li, H. Fu, L. Yu, and J. Huang, “A review of individual tree crown detection and delineation from optical remote sensing images: Current progress and future,”IEEEGeoscienceand RemoteSensingMagazine, vol. 13, no. 1, p. 209–236, 2025. [Online]. Available: https://doi.org/10.1109/MGRS.2024.3479871 [82] Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. K. Li, Y. Wu, and D. Guo, “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,” 2024. [Online]. Available: https://arxiv.org/abs/2402.03300 [83] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” in InternationalConferenceonLearningRepresentations, 2022. [Online]. Available: https://openreview.net/forum?id=nZeVKeeFYf9 [84] J. Irvin, E. Liu, J. Chen, I. Dormoy, J. Kim, S. Khanna, Z. Zheng, and S. Ermon, “Teochat: A large vision-language assistant for tem- poral earth observation data,” inInternationalConferenceonLearning Representations, vol. 2025, 2025, p. 68 883–68 911. [85] Y. Shu, B. Ren, Z. Xiong, D. Pani Paudel, L. Van Gool, B. Demir, N. Sebe, and P. Rota, “Earthmind: Towards multi-granular and multi- sensor earth observation with large multimodal models,”arXive-prints, p. arXiv–2506, 2025. [86] C.-Y. Lin, “Rouge: A package for automatic evaluation of summaries,” in Textsummarizationbranchesout, 2004, p. 74–81.