1. What is SKAI?

SKAI is an artificial intelligence and satellite-imagery solution developed by the World Food Programme (WFP) in collaboration with Google Research.

WFP describes SKAI as an open-source tool that uses artificial intelligence and satellite imagery to automatically assess building damage at large scale.

The purpose is to provide information about disaster-affected areas quickly, helping humanitarian organisations and other decision-makers understand where damage has occurred and where response efforts may need to be prioritised.

WFP describes SKAI as supporting near-real-time damage assessment, situational awareness and analysis of affected areas.

2. What problem is WFP addressing?

A lack of reliable information at the beginning of a humanitarian crisis can make rapid response difficult.

WFP identifies the lack of on-the-ground information as a major obstacle to an effective response. This can become particularly difficult when infrastructure and communication networks have been damaged or disrupted.

Responders need to understand where affected populations are located, what infrastructure has been damaged and how severely different areas have been affected.

Large-scale assessments can require substantial time and resources, particularly after earthquakes, floods, cyclones and other disasters.

3. How does SKAI use AI?

SKAI applies machine-learning techniques to satellite imagery to assess building damage.

At a simplified level, the process can be represented as:

Satellite Imagery → AI Analysis → Building Damage Assessment → Data & Maps → Human Interpretation & Response Planning

WFP describes the technology as capable of automatically analysing satellite images and identifying changes in buildings and affected areas.

The resulting information can be combined with other information, including population density, infrastructure mapping, socio-economic vulnerability and food-security data.

SKAI therefore provides an analytical component within a broader disaster-response information process.

4. What has SKAI been used for?

WFP reports that SKAI has been used in several disaster-response contexts, including earthquakes, floods and cyclones.

• Building-damage assessment
• Monitoring of affected areas
• Situational awareness
• Identification of severely damaged structures
• Response planning
• Information for resource-allocation decisions
• Counter-verification of damage identified through other assessments

5. Myanmar earthquake: 2025

Following the March 2025 earthquake in Myanmar, WFP reports that SKAI was used for a large-scale damage assessment.

WFP analysed approximately 460,000 buildings using high-resolution satellite imagery and identified nearly 4,000 severely damaged or destroyed structures.

WFP reports that the information helped improve situational awareness and supported prioritisation of severely affected areas during the early response period.

In the Pyawbwe region, analysts created paired pre- and post-event imagery to train a supervised machine-learning model for building-damage classification.

6. Sri Lanka cyclone: 2025

Following a cyclone in Sri Lanka in 2025, WFP reports that SKAI rapidly assessed more than 300,000 buildings.

The model identified 10,476 damaged or destroyed structures, representing 3.45 percent of the buildings assessed.

WFP states that the analysis helped responders understand the scale and distribution of the impact and informed early response planning.

7. Türkiye-Syria earthquake: 2023

SKAI was also used following the 2023 Türkiye-Syria earthquake.

WFP reports that during a month-long deployment, a team of five assessed approximately 600,000 buildings with an accuracy rate exceeding 81 percent, identifying more than 28,000 heavily damaged or destroyed structures.

WFP also reports that, in less than a week, SKAI covered approximately 530 km² of urban areas with an estimated population of 1.3 million people and analysed more than 265,000 buildings in three major cities.

WFP notes that the exact impact of the data-driven approach was difficult to measure and was not extensively verified.

8. Hurricane Ian: 2022

Following Hurricane Ian in the United States, SKAI was used to support damage assessment and response planning.

WFP reports that SKAI scanned approximately 410,000 buildings across 3,000 km² within a few days and identified approximately 69,000 damaged buildings.

WFP reports a precision and recall rate of approximately 70 percent for this assessment.

The analysis supported GiveDirectly’s response. WFP reports that GiveDirectly was able to open enrollment for cash assistance in less than a week after the hurricane made landfall.

WFP reports that more than US$2.1 million in cash assistance was ultimately delivered to 4,761 low-income families in heavily affected areas.

WFP also reports that more than US$630,000 was delivered to more than 900 low-income households in one rapid-assistance phase.

9. Pakistan floods: 2022

During the 2022 Pakistan floods, WFP reports that SKAI scanned more than 850,000 households across approximately 8,000 km².

WFP states that the analysis provided situational awareness about affected areas, including urban districts and remote villages, and supported relief-assessment activities conducted by local authorities.

The resulting information was shared with authorities and used to counter-verify damage identified through ground surveys.

In Khyber Pakhtunkhwa, WFP reports that SKAI identified more than 60,000 damaged buildings, with a precision rate of approximately 80–85 percent and a recall rate of approximately 65–70 percent.

10. The role of humans

SKAI automates parts of satellite-image analysis, but the technology operates within a broader human-led response process.

Analysts use satellite imagery, machine-learning models and resulting maps to understand the situation. The resulting information can then be combined with other data and used by organisations, governments and local authorities.

In several documented cases, SKAI outputs were used alongside other assessments, including ground surveys and local information.

SKAI provides AI-assisted analysis; humanitarian actors interpret and use the resulting information within the wider response process.

11. How has SKAI evolved?

WFP reports that SKAI has expanded beyond its initial building-damage assessment applications.

In 2025, WFP reports that SKAI analysed nearly 500,000 buildings following the Myanmar earthquake and more than 300,000 following the Sri Lanka cyclone. A radar-based model was also tested following an earthquake in Afghanistan.

Under SKAI+, WFP reports that a temporary-settlement detection model performed strongly at the minimum viable product stage and was planned for pilot deployment with WFP country offices in March 2026.

12. SKAI as an open-source public good

WFP describes SKAI as an open-source tool and provides access to its technical resources and code.

WFP also describes SKAI as a collaborative solution for stakeholders involved in disaster response and humanitarian assistance.

13. What does the WFP evidence show?

The SKAI case provides documented examples of AI-assisted satellite-image analysis being used in real disaster-response contexts.

The WFP evidence includes measures of the scale of imagery and buildings analysed, numbers of damaged structures identified and, in several cases, precision and recall measurements.

WFP also describes operational uses of the information, including situational awareness, prioritisation, response planning and counter-verification of ground assessments.

These measures should not automatically be interpreted as evidence of causal programme impact. WFP itself notes that the exact impact of some data-driven applications was difficult to measure and not extensively verified.

14. What can M&E professionals learn from this case?

SKAI provides a useful example for thinking about how to evaluate an AI-assisted system that sits between data, analysis and operational decision-making.

An evaluation could examine several distinct levels:

• Data: What imagery and other data are available, and what are their limitations?
• AI performance: How accurately does the model identify damaged buildings?
• Coverage: How much geographic area can be assessed, and how quickly?
• Human validation: How are AI-generated assessments checked or interpreted?
• Decision use: How are SKAI outputs incorporated into response planning?
• Timeliness: Does access to the information enable earlier decisions?
• Equity: Are some communities or geographic areas more likely to be missed or misclassified?
• Programme effects: Does the information contribute to improved humanitarian response?

15. AI accuracy is not the same as programme impact

One important lesson from the SKAI case is the distinction between technical performance and programme-level effects.

A model may have a high precision or recall rate while the resulting information is not used effectively by decision-makers. Conversely, an imperfect model may still provide useful information when its limitations are understood and its outputs are combined with other evidence.

An M&E assessment could therefore examine the complete chain:

Satellite Data → AI Analysis → Damage Assessment → Human Interpretation → Decision → Response → Results

16. Questions for evaluating an AI-assisted disaster assessment system

• How accurate are the damage classifications?
• How do precision and recall vary across locations and disaster types?
• Which types of buildings or geographic areas are more difficult to classify?
• How quickly can usable information be produced?
• How does the AI assessment compare with ground-based assessments?
• How are disagreements between AI-generated assessments and human assessments resolved?
• How do decision-makers actually use the resulting information?
• Does the information change the timing or location of response activities?
• Are particular populations or areas systematically underrepresented or misclassified?
• What evidence exists that improved information contributes to better humanitarian outcomes?

17. What the case does not establish

The WFP Innovation page provides information about SKAI’s technology, use cases, scale and reported technical performance.

It does not provide a comprehensive causal evaluation demonstrating that SKAI itself caused improvements in humanitarian outcomes across all documented cases.

WFP explicitly notes that the exact impact of the data-driven approach in the Türkiye earthquake context was difficult to measure and not extensively verified.

Claims about assessment speed, situational awareness or support for response planning should therefore be distinguished from claims about longer-term humanitarian impact.

18. Key takeaway

SKAI is a documented example of artificial intelligence being integrated into disaster-response information systems.

WFP and Google Research developed an open-source tool that uses machine learning and satellite imagery to assess building damage at large scale and provide information that can support disaster-response activities.

The case also demonstrates why evaluating AI in humanitarian settings requires more than measuring model accuracy.

For M&E professionals, the broader question is how AI-generated information moves through the evidence and decision chain — and whether that information ultimately contributes to better and more timely humanitarian action.

The key question is not simply “How accurate is the AI?” but “What happens after the AI produces its assessment?”

Related EvalCommunity resources

• AI Tool Evaluation Protocol for M&E
  
• Principles for AI Use in Monitoring & Evaluation
  
• Use of AI in Evaluation
  
• The 7 Layers of AI for Monitoring & Evaluation Professionals