Multi-Agent Based Automation of Refugee Camp Planning and Population Mapping

Manual site suitability analysis for a refugee camp involves cross-referencing paper maps, static GIS layers, and field reports—a process prone to oversight and delay. A custom workflow automates this by ingesting live satellite terrain, hydrology, and infrastructure data. Specialized agents score locations against weighted criteria (security buffers, slope, water proximity), generating ranked site options with detailed suitability reports in hours instead of weeks, enabling faster decision-making during a crisis.

Manual headcounts and rough estimates lead to severe over- or under-provisioning of food, water, and medical supplies. This workflow automates shelter detection and density estimation using high-resolution satellite or drone imagery. Computer vision agents count structures and classify types (tents, makeshift shelters), while population models apply per-structure occupancy rules. This delivers a continuously updated population map with quantified confidence intervals, directly feeding into supply chain and logistics platforms like UNHCR's COMPASS or RedRose.

Traditional operations suffer from fragmented data: planners use one set of maps, while supply chain managers use another, leading to coordination failures. A custom orchestration layer, built with frameworks like LangGraph, creates a unified data fabric. Planning agents publish approved camp layouts and population estimates as structured events, which automatically trigger procurement agents in connected ERP systems (e.g., SAP Ariba, Humanitarian Logistics Software) to initiate orders for tents, sanitation kits, and food, based on calculated per-capita needs.

Humanitarian funding requires rigorous, evidence-based reporting on camp conditions and resource utilization. Manual reporting is labor-intensive and often incomplete. This workflow embeds observability and governance by design. Every geospatial analysis, population estimate, and triggered action is logged with source imagery, model version, and agent rationale. This creates an immutable audit trail, enabling the automatic generation of donor-compliant reports that demonstrate precise impact, a critical factor for securing and renewing funding.

Reactive management leads to overcrowding and security incidents. This workflow shifts operations to a predictive posture. Population mapping agents monitor growth trends and shelter density. When thresholds are breached, they automatically trigger alerts and simulate expansion scenarios using the same suitability analysis agents, proposing optimal new sectors adjacent to the existing camp. This allows managers to pre-position infrastructure and security, preventing crises before they escalate.

Monitoring multiple, rapidly evolving camps simultaneously is impossible with manual analyst teams. This custom architecture is built for scale. The agentic workflow can be deployed per-camp or per-region, with a central orchestration hub managing priorities and resource allocation. New data sources (new satellite constellations, drone fleets) are integrated via API, and new camp areas are onboarded as configuration. This allows a single operations center to manage a portfolio of crisis responses, turning geospatial intelligence into a repeatable, leverageable capability.

This agent ingests multi-source geospatial data—satellite imagery (Sentinel-2, Planet), digital elevation models (DEMs), and vector layers for water bodies and infrastructure—to score potential camp locations. It runs a multi-criteria decision analysis (MCDA) evaluating terrain slope, flood risk, proximity to water sources, security buffers from conflict lines, and road access. The agent outputs ranked site polygons with suitability scores and constraint reports, directly feeding the planning module. Without this agent, planners manually cross-reference disparate maps, a process taking weeks and prone to oversight.

A fine-tuned computer vision model (e.g., YOLO or Segment Anything Model on high-resolution satellite/ drone imagery) automatically detects and counts individual shelters, tents, and temporary structures within a designated camp boundary. It estimates population by applying occupancy factors (e.g., 5 persons per shelter) derived from historical data or rapid surveys. This model runs on a scheduled basis (e.g., weekly) to track density changes, replacing manual photo interpretation that is slow, inconsistent, and unable to scale across multiple camps simultaneously.

The central orchestration layer, built with a framework like LangGraph, manages the state and execution flow between all specialized agents. It sequences the workflow: trigger on new crisis event -> call Site Suitability Agent -> receive ranked sites -> invoke Layout Planning Agent with top site -> integrate output with population model for capacity planning -> generate final report. It handles conditional logic, such as rerunning suitability if the top site fails a secondary check, and manages the handoff of geospatial artifacts between components, ensuring auditability and enabling complex, multi-step automation.

This system integration component translates the workflow's outputs (site plans, population estimates) into formats consumable by relief organizations' operational platforms. It exposes APIs to push camp layouts as KML files into GIS platforms (e.g., ArcGIS Online), generate population summary reports in ReliefWeb format, and create procurement lists in supply chain management systems (e.g., LOGISTA). It also pulls real-time data from these systems, such as available stock levels, to inform planning decisions. This layer is critical for moving from analytical insight to operational action without manual data re-entry.

A web-based dashboard presents agent recommendations (site rankings, layout options) to human planners for review, override, or approval. It captures rationale for any changes, which feeds back to improve the agents. This gate is essential for governance, allowing experts to apply contextual knowledge (e.g., political sensitivities, recent ground reports) that the AI lacks. The interface also displays confidence scores from the models and highlights areas of uncertainty (e.g., poor image quality), ensuring human attention is focused on the highest-risk decisions.

A monitoring subsystem tracks the performance of each AI component in production. For the shelter detection model, it logs precision/recall metrics against any ground-truthed data. It monitors data pipeline health (e.g., satellite feed latency, cloud cover obstructing imagery) and triggers alerts for manual review if input data quality falls below a threshold. This component ensures the automation remains reliable and provides the metrics needed to demonstrate ROI—such as reduction in planning cycle time and improvement in population estimate accuracy over manual methods.