---

🌍 GeoDisaster

Benchmarking Orchestrated Agents for Operational Disaster Geo-Intelligence

🎉 Accepted at BMVC 2026

📄 Paper · 💻 Code & Documentation · 🤗 Hugging Face Dataset

Grounded, multi-step disaster geo-intelligence from real Earth-observation and GIS evidence.

---

✨ Benchmark at a glance

2,921 verified tasks	43 question types	5 disaster families	2 agent settings
2,629 train · 292 test	55,648 turn-level SFT rows	18 geospatial tools	Optical · SAR · Raster · Vector · Graph

GeoDisaster evaluates whether tool-using agents can transform heterogeneous Earth-observation (EO) and Geographic Information System (GIS) inputs into structured, evidence-grounded disaster decisions. Each task connects a real geospatial question to declared input artifacts, an executable reference toolchain, and a verifiable answer target.

The dataset provides matched records for single-agent reasoning and orchestrated multi-agent execution. The same 2,921 underlying tasks are available as reference task specifications, full conversations, and turn-level training examples; these are different representations of the same benchmark, not additional independent tasks.

For evaluation: use the gold-free evaluation_inputs configuration / data/core/geodisaster_2921_test_inputs.jsonl. Reference answers and test trajectories must remain on the evaluator side. See Evaluation without answer leakage.

🌋 Five disaster task families

Task family	Dataset key	What it evaluates
🌲 Deforestation Monitoring	deforestation	Forest-loss detection, multi-year attribution, and change measurement from EO layers
🌪️ Multi-Hazard Analysis	multi_hazard	Hazard statistics, cross-layer comparison, spatial exposure, and region prioritization
🏚️ Building-Damage Assessment	xbd_damage	Damage assessment and inventory from building footprints and disaster evidence
🛣️ Flood-Safe Routing	routing	Network reasoning, hazard-aware route planning, and route comparison
🌊 Sentinel-1 SAR Flood Monitoring	sen1_flood	SAR-driven flood mapping, flood measurement, and structured flood reporting

Evidence modalities: optical and SAR imagery; forest-loss masks; hazard and flood rasters; building-damage polygons; vector geometries; road networks; exposure layers; and structured geospatial artifacts.

📦 What is included?

Representation	Purpose	Contents
Core benchmark	Evaluation and reference	Task/question specification, available artifact handles, canonical ground-truth toolchain, structured expected answer
Full conversations	Trajectory analysis and supervised examples	Complete agent episodes: question → tool calls / observations → terminal answer, for single- and multi-agent settings
Turn-level SFT	Training and role-specific analysis	One assistant target per row, with the required context and metadata

Release statistics	Count
Unique benchmark tasks	2,921
Training tasks (per full agent form)	2,629
Test tasks (per full agent form)	292
Single-agent turn-level rows	20,453
Multi-agent role-level rows	35,195
Total turn-level rows	55,648
Terminal targets across both full agent forms	5,842

Training and test partitions are disjoint by task ID, source sample, declared path, normalized routing content, and complete model-visible single-agent trajectory. Within a split, exact repeated episodes are weighted by inverse complete-trajectory multiplicity via training_weight.

🚀 Quick start

Load the released configurations with the Hugging Face datasets library:

from datasets import load_dataset

REPO = "vimageiitb/GeoDisaster"

core        = load_dataset(REPO, "core")
eval_inputs = load_dataset(REPO, "evaluation_inputs")

single      = load_dataset(REPO, "single_agent")
multi       = load_dataset(REPO, "multi_agent")

single_sft  = load_dataset(REPO, "single_agent_sft")
ga_sft      = load_dataset(REPO, "multi_agent_sft_ga")

Other specialist role-SFT files are available in the dataset repository under data/multi_agent/sft_train/ and data/multi_agent/sft_test/. You can also read any released .jsonl file directly, one record per line.

Choosing a view: use core for reference specifications, single_agent / multi_agent for full episodes, SFT views for assistant-turn supervision, and evaluation_inputs for agent-visible held-out tasks. Never give reference-bearing test files to the evaluated agent.

🗂️ Repository layout

data/data/
├── core/
│   ├── geodisaster_2921_{train,test}_source.jsonl   # evaluator/reference tasks
│   ├── geodisaster_2921_test_inputs.jsonl          # gold-free public test inputs
│   └── scoring_policy_43.json                     # active structured-answer policy
├── single_agent/
│   ├── {train_conversations,test_reference_conversations}.jsonl
│   ├── {train_turns,test_turns}.jsonl
│   └── single_agent_system_prompt.md
└── multi_agent/
    ├── {train,test}_conversations.jsonl
    ├── sft_train/{orc,ga,pa,vra}_{train,eval}.jsonl
    ├── sft_test/{orc,ga,pa,vra}_{train,eval}.jsonl
    └── tool_ownership_and_prompt_context.md

source_assets/                                  # bundled permissible EO/GIS inputs
release_tools/
├── fetch_sen1floods11_subset.py                # pinned SAR input installer
└── build_isolated_test_bundle.py               # held-out evaluation bundle
user-guide-final.md                             # authoritative benchmark/runtime guide
ATTRIBUTIONS.md                                # upstream licensing and provenance
SHA256SUMS                                     # distributable archive checksums

🔍 Record IDs, shared fields, and dataset schema

Record identity

Every representation shares three identity fields:

• id: <family>/<sample>/<question_type> — the unique linking key across benchmark forms.
• sample_id: source sample identifier.
• question_type: one of the 43 question types.

For example, deforestation/000086/multi_year_loss_attribution identifies the same task across its core specification, conversations, and SFT turns. Source input files use repository-relative paths such as source_assets/deforestation/000086/masks/lossyear.tif. Non-distributed teacher replay outputs, if referenced, use portable generated-artifact:// logical locators.

Multi-agent conversation fields

Files: data/multi_agent/{train,test}_conversations.jsonl

Field	Type	Description
raw_question	string	Plain question used for active orchestrator/specialist prompts
question	string	Stored evaluation/display context: question, input handles, and answer schema; active prompts use raw_question with sanitized handle cards
initial_artifacts	object	Declared input handle cards with path, label, existence, and optionality
available_initial_handles	list	Handles available to the task
plan	object	ORC plan with goal and {id, agent, subtask, needs} steps
multiagentic_gt_chain	list	Full USER ↔ ORC ↔ specialist trajectory, actions, and observations
executions	list	Per-step {action, tool_output} execution log
expected_final_answer	object	Audit copy of the reference answer; the core record is authoritative
final_answer_ref	string	Terminal transport token (ANSWER_JSON:answer_1)
resolved_final_answer	object	Schema-valid answer projected from worker evidence
resolved_answer_sha256	string	Canonical-JSON checksum linking derived records
resolved_answer_provenance	string	Export provenance: full evidence or validated compact-trace cache
answer_support	object	Supporting evidence metadata
pure_toolchain_sim_status	string	Toolchain replay status

Other released record types

Path	Main fields
data/core/*_source.jsonl	question, initial_artifacts, available_initial_handles, gt_chain, expected_final_answer, source_replay, path_summary, split_seed
data/single_agent/{train_conversations,test_reference_conversations}.jsonl	messages (role, content), tool_sequence, final_answer_ref, resolved_answer_sha256
data/single_agent/{train_turns,test_turns}.jsonl	messages, kind, assistant_action, assistant_turn_index
data/multi_agent/sft_{train,test}/*.jsonl	messages, role, step_id, turn_no, kind

Each role-SFT row contains the exact system-prompt snapshot associated with its target. The data package also includes the standalone single-agent prompt and the multi-agent role/tool prompt context. Executable inference-time prompt builders are authoritative in the code repository.

🛡️ Evaluation without answer leakage

Only data/core/geodisaster_2921_test_inputs.jsonl should be exposed as test records to an evaluated agent. The canonical reference data, test_source, multi-agent test conversations, and single-agent test-reference conversations contain evaluator-only targets.

For a filesystem-isolated evaluation, create a separate root containing just the 292 held-out test tasks and their permitted inputs:

python3 release_tools/build_isolated_test_bundle.py \
  --output /path/outside/repository/eval_bundle

export GEODISASTER_DATASET_ROOT=/path/outside/repository/eval_bundle
export GEODISASTER_STRICT_DATA_ROOT=1

The bundle builder rejects private path declarations; omits gt/, ground_truth/, hidden/, and trajectory products when a declared input is a directory; and writes a SHA-256 manifest. Strict runtime mode prevents fallback searches through the development repository. Keep gold scoring in a separate process or mount; do not mount the full dataset/development repository into the agent sandbox.

The active structured-answer scoring policy is data/core/scoring_policy_43.json (geodisaster_structured_answer_v2). Older v1 policy text inside correction_provenance.publication_contract is historical publication provenance, not the active evaluation policy. Values in correction_provenance.evidence are opaque build labels, not runtime paths or agent inputs.

🧑‍🤝‍🧑 Agent settings and tool ownership

The released trajectories cover two settings:

• Single-agent: one agent plans, invokes tools, and finishes the task.
• Multi-agent: an orchestrator (ORC) assigns work to GA, PA, and VRA specialists.

Role	Tools / responsibility
ORC — Orchestrator	Planning and delegation; owns no geospatial tools
GA — Geospatial Agent	ReadData, PrepareScene, MeasureArea, ExtractObjects, MapFlood, RenderMap
PA — Planning/Analysis Agent	ComputeStats, CompareSources, DetectChange, ClassifyChange, ScoreHazard, EstimateExposure, RankRegions, PlanRoutes, CompareRoutes
VRA — Visual Reasoning Agent	AssessDamage, DescribeFlood, CompareScenes

Terminal-answer contract: all 5,842 full-episode terminal targets use the reference-only token ANSWER_JSON:answer_1. The runtime creates answer_1 only from a schema-valid projection of worker evidence, resolves it at termination, and passes the resulting JSON to the evaluator. The token is transport, not a second answer format; agents do not emit the large resolved JSON themselves. Unknown or missing references fail explicitly, with no ground-truth fallback in live execution.

• Multi-agent ORC ends with Terminate; the single-agent baseline ends with FinalAnswer. Both carry the same answer reference.
• TaskComplete closes a specialist subgoal; it does not transmit the terminal answer.
• Single-agent training support covers SFT and deterministic/live evaluation; GRPO/RCEA training is supported only for the four multi-agent roles.

🧪 Example task and trajectory records

Core task specification (abridged for readability):

{
  "id": "deforestation/000086/multi_year_loss_attribution",
  "sample_id": "000086",
  "question_type": "multi_year_loss_attribution",
  "question": {"handle": "question_1", "text": "… identify the year with the largest forest loss …"},
  "gt_chain": [
    {"tool": "ReadData", "args": {"input": "forest_loss_history_1"}},
    {"tool": "ComputeStats", "args": {}}
  ],
  "expected_final_answer": {"year": 2016, "loss_ha": 201.039833}
}

This example's initial_artifacts include a declared forest_loss_history_1 handle pointing to source_assets/deforestation/000086/masks/lossyear.tif.

Single-agent conversation (abridged; the stored record includes full messages):

{
  "id": "deforestation/000086/multi_year_loss_attribution",
  "tool_sequence": ["ReadData", "PrepareScene", "ComputeStats", "DetectChange"],
  "messages": [
    {"role": "user", "content": "QUESTION … AVAILABLE INPUT HANDLES …"},
    {"role": "assistant", "content": "{\"action\":{\"tool\":\"ReadData\",\"args\":{\"input\":\"forest_loss_history_1\"}}"},
    {"role": "assistant", "content": "{\"action\":{\"tool\":\"FinalAnswer\",\"answer_ref\":\"ANSWER_JSON:answer_1\"}}"}
  ]
}

The corresponding multi-agent example uses the same task id, and additionally includes plan, multiagentic_gt_chain, executions, resolved_final_answer, and reference-answer metadata.

🛰️ Source assets and Sen1Floods11 setup

Bundled files in source_assets/ cover redistributable EO/GIS inputs, while the selected Sen1Floods11 S1/JRC payload is intentionally not included in the thin public release.

To download the pinned 100-chip subset from the official upstream bucket and verify the required files:

python3 release_tools/fetch_sen1floods11_subset.py
python3 release_tools/fetch_sen1floods11_subset.py --verify-only

Important reproducibility details:

• The 500 sen1_flood task records describe a hydrated execution contract; exists: true does not mean every source file is already present in the downloaded archive.
• The installer fetches and verifies 200 upstream S1/JRC source paths and produces 100 portable execution rasters (300 hydrated files in total).
• SHA256SUMS covers the distributed thin archive but excludes locally hydrated files under source_assets/sen1_flood/; the installer verifies those separately using its pinned upstream manifest.
• Dataset input paths are repository-relative. Follow user-guide-final.md for task-specific requirements and correct execution semantics.

⚖️ Licensing and attribution

This is a mixed-license dataset. Its Hugging Face license field is therefore other, not a blanket CC-BY-4.0 declaration.

Component	Applicable terms
Project-authored questions, annotations, and tool-use chains	CC BY 4.0
Redistributed source_assets/	Retain their respective upstream licenses and conditions
xBD / xView2-derived assets	Non-commercial restrictions from their original sources
Selected Sen1Floods11 subset	Retrieved from the official source; subject to upstream terms

Read ATTRIBUTIONS.md for upstream attributions and component-specific licenses before reusing or redistributing source imagery and geospatial files.

📖 Documentation and paper

The dataset package includes user-guide-final.md, the authoritative guide covering the benchmark contract, all 43 question types, runtime/evaluation behavior, and release setup. For executable inference-time code and additional project documentation, see the GeoDisaster GitHub repository.

Paper: GeoDisaster: Benchmarking Orchestrated Agents for Operational Disaster Geo-Intelligence — accepted at BMVC 2026.

Authors: Maram Hasan, Aman Verma, Savitra Roy, Hariseetharam Gunduboina, Daksh Jain, Muhammad Haris Khan, Subhasis Chaudhuri, and Biplab Banerjee.

📄 Read PDF · 🔎 arXiv abstract · 💻 Code

Citation

If you use GeoDisaster tasks, reference outputs, or trajectories, please cite the accompanying paper:

@misc{hasan2026geodisaster,
  title         = {GeoDisaster: Benchmarking Orchestrated Agents for Operational Disaster Geo-Intelligence},
  author        = {Maram Hasan and Aman Verma and Savitra Roy and Hariseetharam Gunduboina and Daksh Jain and Muhammad Haris Khan and Subhasis Chaudhuri and Biplab Banerjee},
  year          = {2026},
  eprint        = {2606.17246},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2606.17246},
  note          = {Accepted at BMVC 2026}
}