Geospatial AI

tags

Machine learning, LLM, Agent, Computer vision, Retrieval augmented generation

The application of AI and LLMs to geospatial data like satellite imagery, GIS databases, mapping, and Earth observation. LLMs have become production-grade tools across the geospatial industry in 2024–2026.

Enterprise GIS + LLMs

The dominant deployment pattern (as of early 2026) is agentic orchestration: LLMs serve as reasoning layers that invoke specialized GIS tools, spatial databases, and domain-specific Foundation models rather than embedding all geospatial knowledge in model weights.

Key deployments:

• Esri + Microsoft (July 2025): Azure OpenAI across ArcGIS. ArcGIS Pro Assistant generates ArcPy/SQL from natural language.
• CARTO AI Agents (early 2025): first “Agentic GIS” platform. Agents reason with spatial data in customer warehouses (BigQuery, Snowflake, Databricks). Open-source toolkit @carto/agentic-deckgl for map control via tool calls. Supports Model Context Protocol.
• NASA Earth Copilot: natural language queries of 100+ PB of Earth observation data, built on Azure OpenAI.
• Planet Labs + Anthropic (March 2025): Claude for analyzing near-daily imagery from 200+ satellites—pattern recognition, anomaly detection, NL querying.
• Satellogic: GPUs deployed directly on orbit, AI running in real-time on satellites, downlinking insights within minutes. NextGen platform (Oct 2025) features 30cm resolution with onboard AI.

Geospatial Foundation Models

See Foundation models. Spatial reasoning evaluates LLM/VLM capability gaps on geospatial benchmarks.

Key models in the geospatial domain (as of early 2026):

• AlphaEarth Foundations (Google, July 2025): integrates optical, SAR, LiDAR, climate, and text into 64-dim embeddings per 10m² cell. 1.4 trillion embedding footprints/year via Earth Engine. 24% lower error, 16× more storage-efficient. (Nature)
• Prithvi-EO-2.0 (NASA/IBM, Dec 2024): 600M params, trained on 4.2M time-series samples from Harmonized Landsat-Sentinel-2. 75.6% avg on GEO-Bench (Szwarcman et al. 2024).
• Clay Foundation Model: trained on 70M satellite/aerial images, 6400 GPU-hours on H100s. Channel-adaptive (varying bands/resolutions), 768-dim embeddings. (Hugging Face)
• TerraMind (ESA + IBM, April 2025): first “any-to-any” multimodal generative AI for Earth observation, Apache 2.0, outperforms comparable models by 8%+.
• OlmoEarth (Allen AI, 2025): ~10 TB of Earth observations, global mangrove maps at 97% accuracy updated 2× faster (OlmoEarth)

Autonomous GIS Agents

The concept of “Autonomous GIS” was formalized by Penn State’s Zhenlong Li (2023). Li’s 2025 research agenda defines five autonomy levels (routine → workflow → data → result → knowledge-aware) and three operational scales (local, centralized, infrastructure).

• GIS Copilot (2025): integrates LLM-Find (data), LLM-Geo (analysis), LLM-Cat (cartography) into QGIS. 86% success rate across 100+ multi-step tasks.
• Google Geospatial Reasoning (April 2025): Gemini 2.5 + Earth Engine + BigQuery + Maps. Scored 0.82 ± 0.02 on benchmarks (vs 0.50 baseline). (Google Research)
• Foursquare Spatial Agent (Feb 2026): multi-agent system (Planner, Data Prep, Analyst, Cartographer) using H3 hexagonal indexing, DuckDB, Kepler.gl.
• GeoAgent (Université Paris Cité, ECML PKDD 2025): code interpreter + static analysis + RAG within Monte Carlo Tree Search (Chen et al. 2024).

Frontier Lab Involvement

As of early 2026:

• Google: AlphaEarth, Earth AI, GenCast (weather, published in Nature Dec 2024, outperforms ECMWF in 97.2% of tests), Open Buildings v3 (1.8B footprints).
• OpenAI: o3 reasoning model identified overhead imagery locations without GEOINT-specific training (GEOINT Symposium, May 2025).
• Mistral AI: three-year framework with French Ministry of Armed Forces (Dec 2025). ESA “Ask ESA” platform built on Mistral LLMs. Sovereign AI positioning.
• Meta: Segment Anything Model (SAM → SAM 2 → SAM 3), segment-geospatial (SamGeo) package. Overture Maps Foundation (2.3B building footprints as of late 2025).
• Microsoft: Planetary Computer Pro (Build 2025), 1.2-1.4B building footprints (Jan 2026 count).

Geospatial Driving LLM Capabilities

Geospatial requirements push four frontier capabilities:

1. Structured output generation: GeoJSON Agents (2025) achieved 97.14% accuracy via multi-agent architecture vs 48.57% for general LLMs (Luo et al. 2025).
2. Multi-step agentic planning: complex workflows like nine-step logistics routing (geocoding → route creation → geometry extraction → intersection → reporting).
3. Multimodal processing: 6+ spectral bands, temporal sequences, resolutions from 0.1m to 30m.
4. Spatial RAG: Spatial-RAG (Feb 2025) introduced sparse-dense hybrid retrieval combining SQL + LLM semantic matching, ~86% delivery rate on geospatial QA (Yu et al. 2025).

GeoLLM (ICLR 2024) showed LLMs embed “remarkable spatial information about locations” from internet text—fine-tuning GPT-3.5 with OpenStreetMap data yielded 70% improvement in predicting population density, wealth, income. (OpenReview)

Bibliography

1. Daniela Szwarcman, Sujit Roy, Paolo Fraccaro, Þorsteinn Elí Gíslason, Benedikt Blumenstiel, Rinki Ghosal, Pedro Henrique de Oliveira, et al.. December 5, 2024. "Prithvi-EO-2.0: A Versatile Multi-Temporal Foundation Model for Earth Observation Applications". https://arxiv.org/abs/2412.02732.
2. Yuxing Chen, Weijie Wang, Sylvain Lobry, Camille Kurtz. October 24, 2024. "An LLM Agent for Automatic Geospatial Data Analysis". https://arxiv.org/abs/2410.18792.
3. Qianqian Luo, Qingming Lin, Liuchang Xu, Sensen Wu, Ruichen Mao, Chao Wang, Hailin Feng, Bo Huang, Zhenhong Du. September 11, 2025. "GeoJSON Agents: A Multi-Agent LLM Architecture for Geospatial Analysis". https://arxiv.org/abs/2509.08863.
4. Dazhou Yu, Riyang Bao, Ruiyu Ning, Jinghong Peng, Gengchen Mai, Liang Zhao. February 4, 2025. "Spatial-RAG: Spatial Retrieval Augmented Generation for Real-World Geospatial Reasoning Questions". https://arxiv.org/abs/2502.18470.