The Digital Engine for Emergency Photo-analysis (DEEP) is a machine-learning application developed by the World Food Programme (WFP) and first deployed in an emergency context in Mozambique in March 2019 following Cyclone Idai. DEEP automates the analysis of high-resolution drone imagery captured during disaster response and was used in coordination with Mozambique's National Institute for Disaster Management (INGC, later INGD).
The strongest surviving source base for this case is the retained WFP feature article validated through a ReliefWeb mirror, supplemented by weaker legacy WFP references whose original URLs no longer resolve cleanly. Read conservatively, those sources support the following claims: DEEP was used in the Mozambique cyclone response to process drone-captured imagery, classify buildings as damaged or undamaged, and produce rapid damage maps for humanitarian decision support. WFP describes the system as helping reduce post-disaster assessment time and as supporting approximately 20 humanitarian partners during the response.
Cyclones Idai and Kenneth caused severe destruction across central and northern Mozambique in 2019. In that context, DEEP was applied to orthomosaic drone imagery to generate building-damage maps that could be used by humanitarian responders and government counterparts to understand where damage was concentrated. The retained source base also supports the claim that WFP had invested in drone capacity-building with Mozambican counterparts before the cyclone response and that local capacity continued to develop after the emergency phase. This matters because the AI system did not operate in isolation: it sat within a broader emergency-mapping workflow involving drone pilots, field teams, humanitarian analysts, and government disaster-management actors.
The available sources describe DEEP as a machine-learning system that can be retrained for other object-detection tasks and can operate offline on standard hardware. They also support the claim that WFP later extended related work to other emergency contexts. That combination of offline capability, retrainability, and fast image processing helps explain why WFP treated the tool as operationally useful in low-connectivity and time-sensitive emergency settings. In practical terms, the system converted large volumes of aerial imagery into a first-pass assessment product that responders could use far faster than manual image review alone.
The social-protection relevance of the Mozambique deployment is indirect but still substantive. Post-disaster response in cyclone-affected areas included emergency assistance, relief targeting, and broader humanitarian coordination that shaped where aid resources could be prioritised first. DEEP did not decide household eligibility for a cash transfer or social-assistance benefit at the individual level. Instead, it informed situational awareness and area-level damage assessment, which in turn supported decisions about where emergency assistance, logistics, and field response capacity should be directed. That places the use case within needs assessment and emergency social-assistance operations rather than automated beneficiary adjudication.
However, because several of the original WFP URLs no longer resolve, this case should be read as a well-corroborated but source-maintenance-fragile account rather than a richly documented official deployment with a fully intact primary source pack. The surviving evidence is still enough to sustain the core claims about the Mozambique deployment, its emergency purpose, and the role of machine-learning-based image classification, but it is thinner than ideal on governance and technical transparency.
The system operates in an advisory capacity within a human-in-the-loop framework. Machine-generated damage outputs were used to inform humanitarian prioritisation, not to make direct eligibility determinations for social assistance benefits. The linkage to social protection therefore remains indirect: DEEP informed broader emergency response planning rather than a documented automated beneficiary-selection pipeline. The key risks are therefore less about direct denial decisions and more about data quality, geographic coverage, and the possibility that erroneous building-damage classifications could distort where responders focus scarce relief resources.