/

Displacement in Somalia is predictable weeks before it peaks, yet no early warning system existed for the humanitarian agencies managing shelter and food across the country’s most vulnerable provinces. Omdena partnered with the UN Refugee Agency to build satellite anomaly detection and a neural network forecasting weekly vegetation conditions from 40 years of NOAA records. Delivered: province-level early warning forecasts, 99% conflict zone monitoring accuracy, and an open-source codebase.
| Outcome | Detail |
| Early warning capability | One-week ahead VHI forecast per Somali province, enabling pre-positioning before displacement peaks |
| Classifier accuracy | 99% in holdout evaluation: Random Forest for Landsat 8 land cover changes; SVM for armed conflict event zones |
| Neural network validation loss | 0.030 — province-level VHI prediction trained on 40 years of NOAA/NESDIS weekly records |
| Satellite pipeline | Unsupervised PCA + K-Means change detection; no labeled training data required |
| Province drivers confirmed | Shabeellaha Hoose: conflict-driven; Hiiraan: flood-driven; Sool: drought-driven |
| Data sources | NOAA/NESDIS VHI weekly since 1984; Landsat 8 and MODIS imagery; UNHCR displacement records 2016–2019 |
| Team | 34 AI experts and data scientists |
Somalia is one of the most displacement-intensive countries in the world. More than 2.6 million people are internally displaced at any given time, moving in response to three overlapping hazards: prolonged drought, sudden flooding, and armed conflict.
Each hazard demands a different humanitarian response, requires different pre-positioning of food, shelter, and medical supplies, and unfolds on a different timescale. By the time displacement events appeared in field reports, the acute phase of the crisis had already passed.
The UN Refugee Agency tracked weekly displacement figures by province from January 2016 through mid-2019, capturing arrivals, departures, and stated departure reasons. Drought was the most commonly cited reason across the dataset, though conflict and flooding dominated in specific provinces.
Those records were reactive, not predictive. No mechanism connected climate deterioration — a falling Vegetation Health Index, a detected flood event, a vegetation anomaly in satellite imagery — to anticipated population movement in a given province. The humanitarian response was always catch-up rather than preparation.
In 2019, the UN Refugee Agency commissioned Omdena to build a predictive intelligence system for Somalia. The challenge required linking satellite-derived climate signals to displacement patterns, a data integration problem crossing geospatial science, time-series forecasting, and humanitarian operations. A team of 34 AI experts and data scientists assembled on the Omdena platform over the project period.
The client supplied historical weekly displacement records segmented by Somali province, including counts of arrivals, departures, and stated reasons for movement. Omdena sourced satellite imagery from NASA’s EarthExplorer platform using Landsat and MODIS sensors, and vegetation health data from NOAA’s STAR/NESDIS archive, with records extending back to 1984.

The first workstream processed Landsat and MODIS satellite scenes to detect sudden changes in vegetation cover, a leading indicator of drought onset or flood damage. Computing pixel-level difference images and extracting NDVI (Normalized Difference Vegetation Index) and NDWI (Normalized Difference Water Index) channels, the team then applied Principal Component Analysis to compress the high-dimensional pixel data into a tractable feature set for clustering.
K-Means clustering with two clusters (changed and unchanged) then labeled each pixel in the difference image. The result was a change detection map showing regions where vegetation health had shifted significantly between observation windows. Areas of sudden vegetation loss flagged potential drought zones; sudden increases following precipitation events identified flood-affected areas.
The second workstream built a time-series predictor for the Vegetation Health Index (VHI), a composite indicator of climate stress. VHI combines VCI (Vegetation Condition Index, measuring greenness relative to historical norms) and TCI (Thermal Condition Index, measuring temperature stress). Readings below 15 indicate severe drought; below 35, moderate; above 65, good; above 85, very good, published weekly by NOAA’s STAR/NESDIS for each Somali province.
The team trained a neural network on weekly VHI records for each Somali province, stretching back to 1984. The model learned to predict the following week’s VHI from recent historical sequences, achieving a validation loss of 0.030 against a training loss of 0.005. Those forecasts gave the client a one-week forward view of vegetation stress before displacement pressure built in the most climate-sensitive provinces.

Two classifier workstreams ran alongside the main pipeline. A Random Forest model applied to Landsat 8 imagery of the Banadir region, Somalia’s largest displacement destination, classified land cover changes at 99% accuracy, showing how conflict and population movement reshaped the urban environment between observation windows.
An SVM classifier trained on armed conflict event records identified high-fatality violence zones by location and incident type, also reaching 99% accuracy. Together, the two classifiers let the client distinguish regions where displacement pressure was environmentally driven from those where conflict was the primary force.

Before the project, the client’s Somalia operations relied entirely on reactive displacement data — counts reported from the field after movements had already occurred. The system inverted this sequence. By forecasting VHI one week ahead and mapping vegetation anomalies continuously, field teams could identify provinces entering drought stress before displacement peaked and pre-position resources during the preparation window.
The province-level correlation analysis resolved a question the client had no prior evidence to answer: displacement drivers differ systematically by region. Banadir, the province containing Mogadishu, recorded the highest arrival volumes, with most of those arrivals departing from Shabeellaha Hoose due to drought and conflict.
Hiiraan flood events and Sool drought onsets require different preparedness windows and resource volumes. The analysis gave the client a principled basis for calibrating alert thresholds by province rather than applying a single national trigger.
The team’s first CNN+LSTM architecture for vegetation change detection required more labeled examples than Somalia’s historical satellite record provided. Switching to PCA + K-Means unsupervised clustering removed the label dependency entirely. The resulting change maps required no ground truth annotation and will apply equally to future provinces or countries where annotated satellite data does not exist.
Treating Somalia as a single prediction unit masked important regional patterns. Correlating VHI and displacement data at the province level revealed that Shabeellaha Hoose, Hiiraan, and Sool had distinct dominant hazard types: conflict, flood, and drought respectively. A national-level model would have averaged these signals out and produced alert thresholds too coarse to trigger a useful operational response.
Both the NOAA VHI archive and the UNHCR displacement dataset arrived as weekly aggregates, but with different reference periods and implicit lag assumptions. Early correlation analyses returned weak or spurious links until the team aligned the two series correctly, accounting for the typical delay between VHI deterioration and population departure.
Two classifier workstreams each reached 99% accuracy from different angles. The Random Forest analyzed Landsat 8 imagery of the Banadir region to document how conflict and population movement altered land cover over time, showing increased urban density in Mogadishu as arrivals from drought- and conflict-affected provinces accumulated.
The SVM identified high-fatality violence zones from armed conflict event records, mapping where armed incidents were most severe. Together, the classifiers gave the client both satellite-grounded environmental evidence and conflict-event evidence for displacement planning.
What makes the VHI forecaster operationally significant is the depth of signal it learned from. Forty years of freely published NOAA/NESDIS weekly records gave the model a training window that no field-collected dataset could have matched. It achieves a validation loss of 0.030 and generates province-level forecasts at effectively zero ongoing data cost, since the VHI archive is updated publicly each week.
The PCA + K-Means change detection pipeline was designed to operate without labeled training data. In humanitarian data environments, labeled ground truth for vegetation anomalies rarely exists at the scale needed for early warning. The pipeline can be pointed at any satellite scene pair and return a change map without annotation effort from field staff, lowering the barrier for future deployment to under-resourced country operations.
Omdena partnered with the UN Refugee Agency to build a climate-displacement early warning system for Somalia. A team of 34 AI experts and data scientists integrated Landsat and MODIS satellite imagery, NOAA/NESDIS vegetation health records, and the client’s weekly displacement records to develop unsupervised change detection, neural network VHI forecasting, and conflict zone classification models. The full codebase is available on OmdenaLore and GitHub.