DroughtGuard

Inspiration

As climate change intensifies droughts and food insecurity around the world, we wanted to build a tool that empowers early action helping organizations anticipate crises, allocate resources efficiently, and support vulnerable communities before conditions worsen.

What it does

DroughtGuard is an AI-powered early warning system that predicts food insecurity risks up to three months ahead. It analyzes anomalies in rainfall, temperature, vegetation, and food prices to forecast risks associated to drought. Results are displayed on an interactive map, allowing users to visualize regional vulnerability levels through clear color-coded indicators.

How we built it

We built DroughtGuard with a Flask-based backend that processes regional data, runs trained machine learning models (XGBoost) for one-, two-, and three-month forecast horizons, and serves predictions through a REST API. The frontend, built with Leaflet.js, TomSelect, and custom JavaScript, renders an interactive map and dashboard that visualize drought and food insecurity risk levels across Kenyan regions. Our models categorize regional risk dynamically using a color-coded system (green for low, yellow for moderate, and red for high). To focus on implementation and infrastructure rather than sourcing and cleaning live data, we created a synthetic dataset that mimics real-world indicators like NDVI (Normalized Difference Vegetation Index), SPI (Standardized Precipitation Index), and food price trends. We hosted our system on the cloud using DigitalOcean to train and store models, preparing a pipeline for continuous data ingestion and retraining, and Vultr to deploy the live web app connected to our .tech domain. The project is designed with a modular AI advisory layer, allowing integration with Google Gemini to generate human-readable regional summaries and recommendations without affecting the core forecasting pipeline.

Challenges we ran into

Most of our challenges involved debugging routing between the frontend and backend and aligning lagged data with the correct forecasting models for one-, two-, and three-month horizons. We also set up Google Gemini using both the google-generativeai SDK and LangChain but couldn’t fully use it due to LangChain’s sensitivity filters blocking some drought-related outputs. In the future, we plan to refine the integration and explore direct SDK use to enable reliable, real-time AI-generated explanations.

Accomplishments that we're proud of

We are proud to have developed a functional end-to-end system connecting AI predictions to an interactive visualization through GeoJSON. We also achieved stable multi-month risk forecasting using different AI models with different data. Finally, we built a clean interface that updates the map dynamically and successfully visualized complex drought data in human-readable format. So far, we only have covered the regions of Kenya, but our goal is to expand the coverage to the rest of the world.

What we learned

We learned how to integrate machine learning, geospatial data, cloud infrastructure, and interactive web visualization into a unified system. We gained experience generating synthetic and lagged datasets to train multiple predictive models for different forecast horizons and deploying them on DigitalOcean for training and storage. We also learned how to host and manage the live dashboard on Vultr, connect it to a custom domain, and dynamically update a GeoJSON-based web interface to reflect real-time regional predictions.

What's next for DroughtGuard

We plan to integrate real-time satellite and climate data to enable continuous model updates and expand DroughtGuard’s coverage beyond Kenya to a global scale. In the future, we aim to build an automated pipeline that regularly retrieves new data, retrains the models, and deploys updated forecasts automatically. We also want to add alert and notification systems for local agencies and farmers, providing early warnings when drought risk increases. Finally, we hope to enhance DroughtGuard’s analytics to highlight which factors contribute most to risk and offer actionable insights for mitigation.