Open Source AI Tools for Crisis Mapping: A Field Guide (2026)

Open Source AI Tools for Crisis Mapping

Open source AI tools for crisis mapping matured significantly between 2023 and 2026, and the field now has a recognisable stack that volunteers, NGOs, and UN agencies share. The advantage of open tools is transparency: methods can be inspected, biases can be discussed openly, and smaller organisations can adopt the same techniques as larger ones without licensing barriers.

This guide focuses on tools that are openly documented, actively maintained, and used in real humanitarian operations rather than only in academic settings. For the broader 2026 tooling landscape, see The Best AI Tools for Humanitarian Data Analysis.

The Base Map Layer

OpenStreetMap remains the foundation. The Humanitarian OpenStreetMap Team (HOT) runs the Tasking Manager, coordinates volunteer mapping for active crises, and maintains tools such as fAIr that combine machine learning with human validation to speed building and road detection in under-mapped areas. Mapillary and KartaView provide street-level imagery that supports validation work.

Satellite Imagery and Segmentation

For satellite imagery, the open stack in 2026 includes Sentinel Hub and Copernicus Open Access Hub for free imagery, QGIS for desktop analysis, and Google Earth Engine and Microsoft Planetary Computer for cloud-scale processing. On the AI side, Segment Anything from Meta and torchgeo from PyTorch have become standard building blocks for segmentation tasks such as identifying settlements, damaged buildings, or flooded areas.

UNOSAT, the UN satellite analysis service, publishes methodology notes that show how these components fit together for operational damage assessment. For the underlying technique, see How AI Reads Satellite Images to Count Displaced People.

Population and Settlement Datasets

Open population datasets underpin most crisis mapping. WorldPop, Meta's High Resolution Settlement Layer, and the EU Global Human Settlement Layer all publish openly. Each has different strengths and known limitations, and the standard 2026 practice is to compare at least two sources before using a population estimate in operational planning.

Conflict and Event Data

ACLED publishes geocoded conflict events under an academic and humanitarian licence. The Uppsala Conflict Data Program publishes complementary datasets. Both are widely used as inputs to AI models that forecast displacement risk or map active conflict zones in near real time.

Validation and Coordination

Open tools only work when humans validate them. The Humanitarian OpenStreetMap Team's validation workflows, the ITHACA emergency mapping service, and the Standby Task Force volunteer network all maintain processes that catch the kinds of errors AI models reliably make: misclassified vegetation as buildings, shadows read as damage, and old imagery treated as current.

For the wider question of where AI can fail in humanitarian work, see Can AI Be Neutral? The Problem of Bias in Humanitarian Data.

Sources and Further Reading

• Humanitarian OpenStreetMap Team: https://www.hotosm.org/
• HOT fAIr machine learning project: https://hotosm.github.io/fAIr/
• UNOSAT: https://unosat.org/
• Copernicus Open Access Hub: https://scihub.copernicus.eu/
• Microsoft Planetary Computer: https://planetarycomputer.microsoft.com/
• WorldPop: https://www.worldpop.org/
• EU Global Human Settlement Layer: https://ghsl.jrc.ec.europa.eu/
• ACLED: https://acleddata.com/
• Uppsala Conflict Data Program: https://ucdp.uu.se/

Tool descriptions reflect publicly available documentation through mid 2026.