亚开行-利用菲律宾的卫星图像绘制贫困的空间分布图（英文）-2021.3-72页_2mb

Summary of "Mapping the Spatial Distribution of Poverty Using Satellite Imagery in the Philippines"

Core Content

This document outlines a feasibility study conducted by the Asian Development Bank (ADB) in collaboration with the Philippine Statistics Authority (PSA) and the World Data Lab to improve the granularity and accuracy of poverty estimates in the Philippines using satellite imagery and machine learning algorithms. The study is part of the ADB's Data for Development project, which aims to enhance the statistical capacities of national statistics offices (NSOs) in Asia and the Pacific to support evidence-based policymaking and the monitoring of Sustainable Development Goals (SDGs).

The primary objective is to generate more detailed and cost-effective poverty estimates at the subnational level, such as municipalities and barangays, which are not adequately covered by traditional data collection methods. The study explores the integration of geospatial data, satellite imagery, and machine learning to provide alternative sources of information for poverty mapping.

Main Views and Findings

1.1 Introduction and Background

• The Millennium Development Goals (MDGs) were significant in reducing global poverty, with the Philippines achieving a notable reduction in income poverty from 34.4% in 1991 to 16.7% in 2018.
• The SDGs require more granular data to monitor progress and ensure that no one is left behind, particularly in marginalized and disadvantaged segments of the population.
• Traditional data sources, such as the triennial Family Income and Expenditure Survey (FIES), provide reliable estimates at national and regional levels but lack sufficient detail at lower geographic levels like municipalities and barangays.

1.2 Sustainable Development Goals Data in the Philippines

• The PSA is the main data custodian for SDG monitoring in the Philippines.
• Of the 244 SDG indicators, 102 are classified as Tier 1, 55 as Tier 2, and 74 as Tier 3.
• Only 32% of Tier 1 indicators are available from PSA sources, while 67% come from other government organizations.
• The Community-Based Monitoring System (CBMS) is used to generate poverty estimates at the local level and supports SDG localization efforts.

1.3 ADB Technical Assistance

• The Data for Development project includes three main components: subnational data disaggregation, enhanced compilation of national accounts, and modernization of national statistical systems.
• The project aims to build the capacity of NSOs to use innovative data sources, such as geospatial data and satellite imagery, for SDG-related data compilation.
• Training modules include machine learning algorithms like random forest, convolutional neural networks (CNNs), and ridge regression.

1.4 Methodology Overview

• The methodology is based on the Stanford research that uses satellite imagery and machine learning to predict poverty.
• The study focuses on using daytime satellite images to predict night light intensity, which serves as a proxy for economic activity.
• The process involves training a CNN, extracting features from its output, and using ridge regression to translate image features into poverty predictions.
• Random forest estimation is used to generate grid-level poverty head count estimates.

Key Findings

• Comparing Averaged Features and Averaged Outputs: The study found that using averaged features from the CNN output provided more accurate poverty estimates than using averaged outputs.
• Validating Image-Level Estimates: Image-level estimates were validated against official poverty statistics, showing a reasonable level of accuracy.
• Comparing with Simpler Models: The CNN-based model outperformed a simpler model that relied only on night light data.
• Comparing with Published Poverty Rates: The uncalibrated machine learning poverty rates showed discrepancies with published poverty rates, highlighting the need for calibration.
• Calibrating Machine Learning Poverty Rates: Calibration improved the accuracy of the machine learning estimates, bringing them closer to official poverty data.
• Comparing with Other Poverty Metrics: Calibrated machine learning estimates were found to be consistent with other poverty metrics.
• Generating Grid-Level Poverty Head Counts: The study successfully generated grid-level poverty head count estimates, which can be used for targeted poverty reduction strategies.

Conclusion

The study demonstrates that the integration of satellite imagery and machine learning can provide more granular and accurate poverty estimates, especially at the subnational level. These estimates are crucial for the formulation of effective poverty reduction programs and for monitoring the progress of the SDGs. The ADB's efforts in this area aim to support NSOs in Asia and the Pacific to adopt innovative data sources and methods for improved data collection and analysis.

Key Techniques and Tools

• Machine Learning Algorithms: Random Forest, Convolutional Neural Networks (CNNs), and Ridge Regression.
• Satellite Imagery: Daytime and nighttime images, including those from the VIIRS sensor.
• Geospatial Data: Used to extract features and improve the accuracy of poverty predictions.
• Open Platforms and Non-Proprietary Data: Promoted for scalability and institutionalization within NSOs.

References and Appendices

• The report includes a detailed appendix on the variables used in population density estimation.
• It also references various studies and reports, including those from the UN, World Bank, and ADB, to support the methodology and findings.

This report is intended to serve as a valuable resource for the PSA and other NSOs in the Asia and Pacific region, encouraging the adoption of innovative data sources to monitor SDG progress and design targeted poverty reduction programs.