GeoAI for Humanitarians: Getting Started

Most humanitarian information management officers I work with have decent GIS skills. They can produce a partner-presence map in QGIS, build a flood vulnerability layer, run zonal statistics. What very few of them have is hands-on confidence with GeoAI — the application of machine learning techniques to geospatial data.

The hesitation is understandable. GeoAI sounds intimidating. Conference talks describe deep learning models trained on terabytes of satellite imagery to predict everything from crop yields to refugee flows. The barrier-to-entry signal is overwhelming.

In practice, most operational humanitarian GeoAI work is much simpler than the conference talks suggest. It's standard machine learning techniques applied to standard geospatial data, using tools that are mostly free and increasingly approachable. After supervising national flood, drought, avalanche, landslide, and extreme-temperature risk mapping at 4 km resolution in Afghanistan, and after rebuilding the GIS workflow at UNICEF Ethiopia and FAO Nigeria, I'm confident this is a tractable skill set for any IM officer who wants to add it.

This is the practical guide I wish I'd had when I started.

GeoAI is the application of machine learning to data that has a spatial dimension. That's the whole definition.

It's not magic. It doesn't predict the future. It doesn't replace judgment. What it does is automate pattern recognition at a scale that manual analysis can't reach — and then surface those patterns as analytical inputs that a human practitioner uses to make decisions.

Three operational categories cover most humanitarian use cases:

Classification. "Is this satellite pixel forest, agriculture, or built-up?" "Is this household at high, medium, or low risk?" Classification problems are where most GeoAI gets used in humanitarian contexts.

Regression and prediction. "How much will the NDVI in this zone drop given current rainfall trends?" "How many people are likely displaced based on the destruction signature in this Sentinel-1 image?" Estimating continuous values from spatial inputs.

Detection and segmentation. "Where in this image are the buildings, and which ones are damaged?" "What is the boundary of the flood inundation in this scene?" Pulling specific features out of imagery automatically.

If your humanitarian question fits one of those three categories and has a spatial dimension, GeoAI is in scope. If it doesn't, no model will help you.

Forget the cutting-edge research stacks for now. The tools below cover 90% of humanitarian GeoAI use cases.

Google Earth Engine. This is the gateway. Free for non-commercial use, browser-based JavaScript or Python API, and the entire planetary archive of MODIS, Landsat, Sentinel, CHIRPS, and more is at your fingertips with one-line queries. My drought analysis in Afghanistan that identified 223 of 401 districts in extreme to abnormally dry conditions used Earth Engine for VHI/VCI/TCI computation at national scale. A workflow that would have required days of raster downloads and processing took an hour.

QGIS. Open-source, infinitely extensible. The QGIS-Python integration via PyQGIS lets you script anything. The processing toolbox includes most standard GIS operations. The Semi-Automatic Classification Plugin handles supervised classification of satellite imagery. Combined with QGIS, you can do most humanitarian GeoAI without writing a single deep-learning model.

Python (Pandas, GeoPandas, Rasterio, scikit-learn). When you need to script ETL pipelines, run a classification or regression model, or build a reproducible workflow, this is the stack. GeoPandas makes spatial data feel like dataframes. Rasterio handles satellite imagery natively. Scikit-learn covers classical machine learning end-to-end.

Microsoft Planetary Computer. The newer option, similar in spirit to Earth Engine but with stronger Python integration and access to additional datasets like the Microsoft global building footprints. Worth knowing about even if Earth Engine remains your daily driver.

WorldPop. Population estimates at 100 m resolution, globally. The single most useful baseline layer for humanitarian exposure analysis, and the dataset I cite most often when scoping a new operation.

That's the working stack. Master those five and you can do most operational humanitarian GeoAI.

Use case 1: Drought monitoring with NDVI and CHIRPS. This is the canonical starter project. Pull NDVI from MODIS for the last 12 months. Pull CHIRPS rainfall for the same period. Compare both against the long-term mean for the area of operations. Generate a monthly anomaly map. Add a Vegetation Health Index layer that combines vegetation stress with temperature anomaly.

In Afghanistan, this workflow — extended with Sentinel-1 SAR backscatter for flood detection and ASTER DEM for avalanche risk modelling — became the iMMAP-OCHA Disaster Risk and Climate Outlook Mapping Methodology reference. Once you have it running for one country, replicating it for another takes hours, not weeks.

Use case 2: Flood extent mapping with Sentinel-1 SAR. SAR penetrates clouds, which makes it the only operational option for flood mapping in monsoon contexts. The classification logic is straightforward: water has very low backscatter compared to dry surfaces, so flooded areas show up as dark pixels in a SAR image. The hard part is distinguishing real water from shadows, urban reflections, and pre-existing water bodies — which is where simple thresholding gives way to supervised classification with a small training dataset.

I used variants of this workflow for flood vulnerability mapping in Borno, Yobe, and Adamawa during the rainy-season contingency planning at FAO Nigeria, and integrated GloFAS forecasts on top to produce the early-warning maps that fed the Shelter/NFI sector's preparedness plan.

Use case 3: Building footprint extraction for displacement tracking. The Microsoft global building footprints dataset has changed how rapid displacement assessment works. Combine it with pre-event and post-event satellite imagery, and you can detect new construction (informal settlements, displaced-population shelters) or destruction (conflict damage, disaster impact) at scale. The classification challenge — what counts as a "new building" vs noise — is non-trivial but tractable with simple change-detection workflows.

Here's the project I tell IM officers to commit two weeks to as their entry into operational GeoAI.

Week 1, Days 1–3. Open a Google Earth Engine account. Run the introductory tutorials. Download the Earth Engine Python API (geemap is a friendly wrapper). Pick a country you know well — your current operating context.

Week 1, Days 4–7. Compute monthly NDVI mean for the last 24 months for your country. Compare to the 2015–2024 baseline. Generate an anomaly map. Export the result as a GeoTIFF.

Week 2, Days 1–3. Bring the GeoTIFF into QGIS. Overlay it with admin-2 boundaries from OCHA Common Operational Datasets. Compute zonal statistics — mean NDVI anomaly per district. Identify the top-10 most-stressed districts.

Week 2, Days 4–5. Cross-reference the stressed districts with WorldPop population estimates. Generate a "population in vegetation-stressed districts" estimate.

Week 2, Days 6–7. Write up a one-page methodological note explaining what you did, what the data sources are, what the limitations are, and what the analysis tells you about your operational context.

You're not going to publish this. You're going to learn from it. By the end of two weeks, you'll have run an end-to-end GeoAI workflow that mirrors a real humanitarian product. Every subsequent project gets easier.

Model accuracy is not the same as operational utility. A classifier that's 95% accurate on a held-out test set can be operationally useless if the 5% errors cluster in your most consequential decisions. Always validate against ground truth from the actual operational context, not just statistical metrics.

Resolution matters more than algorithm. A simple model on 10 m Sentinel data outperforms a sophisticated model on 250 m MODIS data for most local operational questions. Get the resolution right before getting the algorithm right.

Uncertainty quantification is harder than prediction. Producing an estimate is the easy part. Producing an honest confidence band around the estimate is what makes the product trustworthy. Most humanitarian GeoAI products skip this step. They shouldn't.

Validation has a shelf life. A model trained on 2022 imagery may not generalise to 2026 conditions. Land cover changes. Building patterns shift. The displacement signature in a SAR image looks different when the underlying landscape has been transformed by drought. Re-validate periodically.

Open data has political constraints. In some operating contexts, satellite imagery analysis is sensitive. Coordinate with the country office, the cluster lead, and (where appropriate) national authorities before publishing detailed spatial products. The technical work and the political work are inseparable.

The most common mistake I see in humanitarian GeoAI is treating model output as the answer rather than as input to the answer.

A classified flood-extent map is not a needs assessment. It's a starting point for a needs assessment. The difference matters. Operational decisions are made by people, informed by evidence, accountable to affected populations. GeoAI extends what evidence is possible to assemble; it doesn't replace the judgment that turns evidence into decisions.

The IM officers who get the most out of GeoAI are the ones who treat it as another tool in the analytical kit, not as a magic answer machine. Combine the GeoAI output with field reports, partner consultations, baseline household surveys, and operational context — and you have a picture you couldn't have built any other way.

That's the skill set worth building. Start with two weeks. Pick a country. Run the workflow. The rest follows.