How AI and Satellites are Transforming Poverty Analysis

At the core of the study is the use of deep learning techniques applied to satellite data (Figure 1), consisting of two key components:

• Feature extraction with convolutional neural networks: Daytime (Sentinel-2) and nighttime (visible infrared imaging radiometer suite-VIIRS) satellite images were analyzed using convolutional neural networks to identify features such as infrastructure, vegetation, and urbanization, which serve as proxies for economic activity.
• Predictive modeling with machine learning: The extracted features were combined with additional geospatial variables, such as population density, air pollution levels (NO₂, SO₂, CO), land surface temperature, and points of interest (e.g., locations of facilities such as schools, hospitals, and businesses). These inputs were fed into machine learning models, such as ridge regression and random forest, to predict poverty rates at a highly granular level.

In Indonesia, the model was trained using official village- and district-level poverty data from Badan Pusat Statistik and SMERU Research Institute to generate grid level (each grid being 2.4 kilometers by 2.4 kilometers) poverty map. The trained model was then applied to Maldives to assess whether AI models developed for one country could reliably predict poverty in another^[1].

Figure 1: Visualization of the Process to Extract Features from Satellite Images

CNNs = convolutional neural networks
Source: A. Head. 2017. Can Human Development be Measured with Satellite Imagery? Association for Computing Machinery.

Indonesia: Enhanced Granularity in Poverty Estimates
The AI-powered approach successfully captured detailed variations in poverty rates across Indonesia. The random forest model outperformed traditional regression methods, explaining up to 52% of the variation in poverty rates. The improved accuracy allowed for the identification of specific regionals patterns in provinces, such as Banten and East Java, where integrating geospatial data revealed valuable insights into localized poverty dynamics.

Areas in eastern part of Indonesia showed higher poverty rates compared to western and central parts. The unique geographical and economic backgrounds of Banten and East Java mean that a standard approach might not yield optimal results. Banten has more urban and industrial profile, thus, models that depend largely on nighttime lights and built-up area indices have the potential to produce accurate outcomes.

A large portion of these grids are located at the western end of Banten, with only one grid at the center, showing potential small pockets of poverty that are not likely to be reflected using a more general approach to poverty estimation. These same models might not work as well in East Java, where rural areas and agriculture contribute more substantially, requiring the application of variables like Normalized Difference Vegetation Index or slope to correctly depict the living standards and financial disparity in the province. The results showed more pockets of poverty in East Java.

The combination of satellite imagery and machine learning techniques also showed patterns that would have been difficult to detect using traditional surveys. Areas with low nightlight intensity but significant urban infrastructure were identified as having underreported economic activity. In contrast, some rural areas with high vegetation cover and limited built-up structures were associated with higher poverty rates (Figure 2).

These high-resolution insights can be instrumental in optimizing government interventions. For instance, identifying micro-regions with high poverty rates allows for targeted social assistance programs, infrastructure investments, and localized economic development projects.

Figure 2: Poverty Rates at High-Resolution Across Indonesia

Source: Asian Development Bank. Forthcoming. Mapping the Spatial Distribution of Poverty Using Satellite Imagery and Other Geospatial Data in Indonesia.

Maldives: Testing AI Model Transferability
Maldives presented a unique challenge given its smaller size and lack of granular survey data (Figure 3). Applying the Indonesia-trained model to Maldives revealed the following key insights:

• Prediction limitations. The AI model struggled to accurately predict poverty in Maldives due to differences in data distribution and socioeconomic conditions between the two countries.
• Geospatial indicators and model adaptability. While satellite-derived visual features captured economic activity, the significant discrepancy in poverty ranges between Indonesia and Maldives affected predictive accuracy. The highest observed poverty rate in Maldives was lower than the median rate in Indonesia, suggesting that models need local calibration for improved performance.
• Alternative approaches. Incorporating local Maldivian socioeconomic indicators, such as household consumption patterns and employment distribution, could significantly improve model reliability. The study indicated the need for hybrid methodologies because the direct application of Indonesian-trained models posed limitations.

The findings showed the need for complementary datasets to refine AI model predictions. In data-scarce settings like Maldives, information from mobile phone records, transportation patterns, and livelihood economic indicators could serve as additional proxies for wealth estimation in future models.

Figure 3: Comparison of Google Maps (left), Sentinel 2 Daytime Imagery (middle) and VIIRS Nighttime Lights (right) for Male, Maldives.

Note: VIRRS = Visible Infrared Imaging Radiometer Suite. Source: Graphics generated by the study team.

The study demonstrates the transformative potential of AI in poverty mapping while also highlighting the challenges of applying models across different country contexts. Key takeaways for policymakers and researchers include:

• Better targeting interventions. High-resolution, AI-powered poverty maps enable governments and aid organizations to allocate resources more effectively, particularly in remote or underserved areas. With AI-driven insights, governments can design conditional cash transfer programs, school feeding initiatives, and microfinance opportunities tailored to specific microregions.
• Enhancing model adaptability. AI models trained in data-rich countries require adjustments when applied to low-data settings. Adaptive transfer learning approaches—where pre-trained models are finetuned with limited local data—can enhance performance. It is important to exercise caution when using poverty maps derived from remotely sensed data and AI-based approaches as the approach may underperform relative to conventional methods of poverty mapping when there are issues on the quality of remotely sensed data.
• Integrating alternative data sources. Using nontraditional data sources (e.g., mobile phone usage data, social media activity, economic indicators from fintech transactions) can enhance prediction accuracy, especially in areas with limited official statistics.
• Combining AI with local knowledge. Community-based data collection efforts, citizen surveys, and participatory mapping exercises can validate AI-driven poverty maps. Combining machine learning with local expertise improve real-world applicability.

Integrating poverty maps with other socioeconomic data provides nuanced information for policymaking and program targeting as countries move forward.

By integrating AI, remote sensing, and machine learning, poverty mapping is entering a new era—one where policy decisions can be informed through more precise, data-driven insights. Continued innovation and strategic application of these tools hold immense potential to drive more effective and targeted poverty alleviation efforts globally.

Two tools that illustrate the promise of the next phase are:

• Poverty Clock: Developed by the World Data Lab^[2], the Poverty Clock provides real-time updates on global poverty trends that offers visualization of progress toward poverty alleviation.
• Poverty Impact and Vulnerability Evaluation (PIVE): Supported by ADB's Technology Innovation Challenge and developed by LocationMind, Inc., PIVE bridges the gap between traditional statistics and modern data analytics by integrating multiple data sources, including mobile GPS data, government statistics, crowdsourced information, and remote sensing data, with machine learning-based poverty maps. The tool allows policymakers to identify areas of vulnerability, provide interventions, and optimize aid distribution.

During its pilot in Metro Manila, PIVE analyzed real-time mobility patterns and environmental risks using satellite imagery and integrated these insights with demographic data. The pilot phase highlighted how different population segments, especially marginalized communities, are affected by disasters and socioeconomic shocks. Stakeholders praised PIVE's granular data outputs, intuitive dashboard, and seamless data integration, reinforcing its potential for evidence-based policymaking and effective development strategies.

Future research could refine poverty mapping model accuracy using multi-country datasets, synthetic data augmentation, and additional socioeconomic indicators. Partnerships between governments, international organizations, and technology firms could accelerate the development of open-access AI tools for poverty prediction, democratizing data-driven decision-making globally. The application of AI in poverty mapping and its integration with other socioeconomic data is a crucial step toward more equitable and informed global development.

Note: Joseph Albert Niño Bulan, Mildred Addawe, and Christian Flora Mae Soco contributed in drafting this article while Takaaki Masaki provided valuable insights during its preparation.

^[1] The two countries, Indonesia and Maldives, were selected for the project based on their potential usefulness and alignment with country-specific strategies. Another key criterion was data availability—specifically, the presence of a household income and expenditure survey conducted reasonably close to a census year. From the perspective of using satellite imagery, the methodology should also be tested in varied contexts, such as a large archipelago like Indonesia, which exhibits diverse image features, and a smaller economy like Maldives, which presents more homogeneous image features. This contrast enables examination of how variations in image characteristics may affect the accuracy of estimates.

^[2] The World Data Lab provided technical support to ADB in developing machine learning algorithms, training the models, and leading two country-specific case studies on poverty mapping using integrated datasets. The firm also assisted ADB in conducting statistical capacity building sessions on machine learning-based poverty mapping for selected national statistical offices.

Asian Development Bank. 2021. Mapping the Spatial Distribution of Poverty Using Satellite Imagery in the Philippines. 

Asian Development Bank. 2020. Mapping Poverty Through Data Integration and Artificial Intelligence: A Special Supplement of The Key Indicators for Asia and the Pacific 2020. 

N. Jean et al. 2016. Combining Satellite Imagery and Machine Learning to Predict Poverty. Science. 353 (6301). pp.790–794. 

LocationMind, Inc. 2023. Developing “Poverty Impact and Vulnerability Assessment Tool” in the Philippines with the Asian Development Bank. 4 April.