AI for Disaster Response: Predicting Relief During Cyclones

Cyclones announce themselves with small, unsettling changes: a sudden stillness, a shift in the color of the sky, stronger gusts that arrive long before the eye makes landfall. For coastal communities, these signals are both familiar and terrifying. Families rush to secure their homes, gather essentials, and wait, radios and phones tuned to increasingly urgent warnings. It’s a tense, emotional stretch of time where hope and fear sit side by side. When the storm finally hits, everything can change in minutes.

Cyclone Amphan, which struck in May 2020, exemplifies how quickly lives can be upended. In Kolkata and across coastal Bangladesh, floodwaters swept through streets, trees were uprooted, electricity grids collapsed, and entire neighbourhoods were damaged or destroyed. Official tallies reported dozens of lives lost, but the deeper impact was felt through displaced families, broken livelihoods, and disrupted access to clean water, food, and shelter. The human reality was far greater than numbers could express.

For humanitarian responders, these first chaotic hours matter enormously. Decisions must be made even when information is incomplete. Who needs help? How urgently? Where should supplies be directed first? This is where AI can support disaster response by offering early, data-driven insights that help teams act faster and more confidently. While technology cannot stop a cyclone, it can help ensure that support reaches people in the most affected areas when they need it most.

Cyclone Amphan: scale of destruction and exposed population

The United Nations World Food Programme (WFP) mobilises aircraft, ships, warehouses and thousands of trucks in emergencies. Yet even with such capacity, delivering the right relief to the right people within the narrow window after landfall depends on an accurate early estimate of who is affected and what their needs are. That estimation is exactly what this project addressed through AI-powered methods.

The Challenge of Predicting Relief Needs

In the first hours after a cyclone, information is fragmented. Communications networks may be down, roads blocked and first-hand assessments delayed. Responders must nevertheless commit logistics, pre-position supplies and prioritise vulnerable locations. Errors in these early choices can leave families without essentials, or divert scarce resources from other crises. Predicting the scale of impact early — before full assessments are possible — requires combining meteorological data with socio-economic and demographic context in a way that captures local vulnerability.

Traditional approaches rely on exposure maps and rule-of-thumb multipliers. These are useful but often blunt. A more refined, data-driven approach could integrate storm intensity and duration with population distribution, housing quality, and socio-economic indicators to produce more accurate, actionable estimates of likely needs. That is the aim of applying AI in this context: fast, probabilistic, and operationally meaningful estimates that field teams can trust when they must act.

A Global Collaboration to Build an AI Solution

In March 2020 Omdena joined forces with the WFP Innovation Accelerator to ask a pragmatic question: can AI help estimate affected populations early enough to improve relief planning? The project convened 34 contributors — data scientists, engineers and humanitarian specialists — working in distributed, self-organising teams. Omdena’s bottom-up collaboration model emphasises rapid prototyping, shared ownership and continuous integration of humanitarian feedback.

Crucially, the team began with a deep immersion into WFP’s operational workflows. They mapped how field assessments are performed, how relief packages are assembled to meet nutritional and dignity standards, and what constraints limit operational flexibility. This practical focus ensured that any technical solution would meet real-world needs rather than academic criteria alone.

For a broader view of emerging technologies shaping emergency management, here’s an overview of top disaster management companies advancing real-time resilience.”

The data challenge was substantial: humanitarian and meteorological data are widely distributed, inconsistently formatted and often locked in scanned reports. More than 70% of the project time was spent locating, digitising, cleaning and integrating data from multiple sources so that reliable models could be trained and validated.

• Key data sources included:
• IBTrACS (cyclone tracks, wind speed, pressure)
• EmDAT (historical disaster impacts)
• World Bank socioeconomic indicators (GDP, HDI, rural %)
• Gridded Population of the World (GPW)
• WFP and WHO operational guidelines

Building the AI and Machine Learning Models

The modelling task was divided into two linked problems. The first predicted the number of people likely to be affected by a given cyclone; the second translated that estimate into quantities of food and non-food relief items using established humanitarian standards.

To predict affected population, the team engineered features that combined meteorological attributes (e.g., maximum wind speed, minimum pressure, hours over land) with demographic and socio-economic context (population density near the track, GDP per capita, cereal yields, access to services). Exploratory analysis highlighted the importance of specific variables and helped guide feature selection and preprocessing.

• Machine learning algorithms tested included:
• Random Forest Regressor
• Gradient Boosted Trees
• XGBoost
• Support Vector Regression (SVR)
• Neural Networks
• Ensemble approaches

Model evaluation showed stronger performance for moderate-scale events (affected populations up to ~150,000), which were well-represented in the historical record. For larger catastrophic events, performance suffered due to smaller sample sizes; this highlights the importance of continuous data collection and model improvement.

After producing population estimates, the pipeline applied a mathematical conversion to calculate required items. The model used WFP and WHO guidance to determine caloric needs, household sizes, non-food item requirements (e.g., blankets, hygiene kits), and adjustments for special groups such as children and pregnant women. The result was a coherent, ready-to-use relief package estimate that responders could operationalise.

From Data to Prediction: Turning Complexity Into Clarity

With the datasets in place, the team visualised how storms intersect with human settlements. Mapping cyclone tracks over population grids revealed local hotspots of exposure. Cyclone Pabuk, for instance, traced a path through multiple countries and highlighted how a single storm can affect diverse socio-economic contexts.

Cyclone Pabuk: trajectory across affected regions

A feature-importance analysis identified which variables most influenced affected-population estimates. Variables such as maximum wind speed, minimum atmospheric pressure, total hours on land, population density, GDP per capita, rural population share and HDI were among the top predictors. These findings reinforced the need to combine meteorological forecasts with socio-economic context for meaningful predictions.

Most influential features for affected-population modeling

The models were iteratively refined with cross-validation and holdout testing. Performance metrics included mean absolute error and other robust regression diagnostics. These steps ensured the model was not only accurate on historical events but also generalised reasonably to new storms when provided with timely input parameters.

Deploying the Models

Operationalising models for field use was a priority. The team deployed a Streamlit-based application that accepts cyclone parameters (e.g., current location, forecast track, maximum sustained winds, pressure, and expected duration). The front-end is intentionally minimalistic so that users in low-bandwidth environments can operate it quickly.

Streamlit’s output — an estimated affected population — feeds into an offline Python GUI that applies the relief-calculation logic. This separation ensures the tool remains functional even when connectivity is intermittent: users can run the GUI locally with preloaded socio-economic and population grids. The end-to-end workflow reduces the time needed to go from forecast to a packaged, itemised relief plan from hours to minutes in many cases.

Presented at the Omdena demo day

Documentation and open-source notebooks were prepared to enable replication and improvement by other practitioners. The team emphasised transparency: models, feature lists, and assumptions were documented so humanitarian teams could understand how outputs were generated and how to interpret confidence bounds.

A demonstration of the relief package tool can be found here.

Potential Use

The tool has clear applications beyond cyclones. Any sudden-onset event with geospatial exposure — earthquakes, tsunamis, volcanic eruptions — could benefit from a similar pipeline that combines hazard forecasts with vulnerability layers. The cyclone-focused prototype demonstrates how relatively small amounts of curated data, combined with straightforward modelling, can produce operationally useful outputs.

Imagine a cyclone approaching Madagascar. Early forecasts provide wind and pressure projections. A humanitarian planner feeds these into the Streamlit interface, which immediately returns an estimate of people likely to be affected. The offline GUI then translates the population figure into a detailed list of food rations and non-food items tailored to household composition, temperature considerations and nutrition needs. The same platform can generate scenarios — best case, likely, and worst case — to support contingency planning and pre-positioning of assets.

Because the system stores past events and outcomes, it can be re-trained as new impact data arrives. Each response therefore contributes to long-term model improvement, better preparedness, and smarter allocation of limited humanitarian resources.

Conclusion

Cyclones will remain a threat for many coastal communities worldwide. What can change is how quickly and accurately humanitarian teams prepare to respond. The Omdena–WFP collaboration shows that when technical expertise is combined with humanitarian practice, AI can fill a vital operational gap: producing timely, defensible estimates of affected populations and translating those into concrete relief plans. This advances not just speed but also the dignity and effectiveness of responses.

The work is ongoing: improving coverage for large-scale events, integrating higher-resolution population and shelter data, and embedding real-time satellite-derived damage proxies are natural next steps. But the central lesson is already clear: data-driven tools, built with empathy and operational focus, can make a measurable difference in how the world responds to sudden disasters.

Ready to strengthen your disaster response with AI powered insights? Connect with Omdena and let us see if we are a good fit.