1. The Information Challenge

After cyclones Idai and Kenneth struck Mozambique in March and April 2019, WFP used drones to collect high-resolution imagery of affected areas.

Thousands of drone images were collected and stitched together into detailed maps. These maps provided information that government and relief organizations could use to support decisions about where and how to deliver assistance.

However, analysing the imagery still took days. This created an important operational challenge: how could large quantities of imagery be converted into useful information quickly enough for an emergency response?

2. The DEEP Solution

WFP developed DEEP — Digital Engine for Emergency Photo-analysis, an in-house solution designed to analyse drone imagery.

The algorithm was designed to quickly quantify buildings in a mapped area and identify buildings that were partially or completely destroyed.

WFP reports that DEEP reduced the analysis time from weeks to hours.

3. What DEEP Helped Identify

WFP describes DEEP as being used to identify and quantify buildings and distinguish between buildings that were partially and completely destroyed.

The resulting information could provide governments, donors and local communities with a faster picture of damaged infrastructure. WFP states that this information could support estimates of repair costs, emergency housing requirements and related mobilization.

4. The Reported Operational Improvement

Evidence note: “Weeks to hours” is the result reported by WFP. The source does not provide a detailed before-and-after measurement methodology or exact processing times, so no additional figures are added here.

5. From Response to Preparedness

The experience with DEEP did not end with the immediate emergency response.

WFP had trained pilots from Mozambique’s National Institute of Disaster Management (INGC) in 2017 and 2018. These pilots field-tested DEEP after cyclones Idai and Kenneth.

Three months later, they were using fixed-wing drones to map the Licungo river basin as part of preparedness activities ahead of the flood season.

6. Human Expertise Remained Part of the System

DEEP did not eliminate human involvement.

WFP reports that local pilots were trained and participated in field use of the technology. Students from Eduardo Mondlane University also helped with image annotation when the DEEP model was being trained.

The case therefore illustrates a combination of AI, human expertise, local operational capacity and data preparation.

7. Continuous Learning and Adaptation

WFP describes the initial application as training the system to detect buildings. The organization also describes possible future applications involving other objects, including bridges, roads and standing water.

WFP also mentions the possibility of using more sophisticated cameras and performing analysis during drone flights.

These are possibilities described by WFP in the source article, not results established by the Mozambique implementation described here.

8. How M&E Professionals Can Adapt the Same Approach

M&E professionals do not need to reproduce the DEEP technology to apply the underlying idea. The transferable approach is to identify a data-intensive process, use technology to accelerate a clearly defined analytical task, and connect the resulting information to a real decision.

For an M&E team, the process could look like this:

For example, an M&E team could examine whether AI could assist with a clearly defined, repetitive analytical task involving a large volume of monitoring or evaluation data. The appropriate application would depend on the type and sensitivity of the data, the consequences of errors, and the ability to validate the results.

Important: These are adaptation suggestions for M&E professionals based on the workflow illustrated by the DEEP case. They are not claims that WFP used DEEP for these M&E applications.

9. Checks and Balances

WFP also emphasizes the importance of checks and balances as the relationship between humans and AI develops, including issues related to safety, security, ethics and data protection.

For M&E applications, this is particularly important when AI-generated information may influence programme decisions or the interpretation of evidence.

10. What This Case Does Not Tell Us

The WFP article does not provide enough information to establish:

• The specific machine-learning architecture used.
• A numerical accuracy or error rate for the model.
• A detailed cost-benefit analysis.
• The number of lives or resources directly affected by DEEP.
• An independent evaluation of the model’s performance.

These details have therefore not been added to this case study.