亚开行-使用泰国的地理空间数据预测贫困（英文）-2020.12-38页_5mb

Summary of "Predicting Poverty Using Geospatial Data in Thailand"

Core Content

This working paper explores the use of geospatial data to predict poverty in Thailand, offering an alternative to traditional household surveys. The study investigates whether readily available geospatial data, such as satellite imagery and crowd-sourced databases, can accurately estimate poverty levels and their spatial distribution. It also compares the predictive performance of various econometric and machine learning methods.

Main Objectives

• To assess the feasibility of using precompiled geospatial data to predict poverty in Thailand.
• To evaluate the effectiveness of different machine learning algorithms in poverty prediction.
• To contribute to the literature on poverty measurement by using structured geospatial data rather than computer vision techniques.

Key Variables and Data Sources

Satellite Data

• Rainfall: From CHIRPS (Climate Hazards Group InfraRed Precipitation with Station data).
• Land Surface Temperature (LST): Day and night data from MODIS satellites.
• Normalized Difference Vegetation Index (NDVI): From MODIS satellites.
• Night Light Intensity: From DMSP OLS and VIIRS satellites.

Crowd-Sourced Geospatial Database

• OpenStreetMap (OSM): Provides data on road count, road length, points of interest (POI), and built-up areas.
• POI Classification: Categorized into 16 types based on economic activity and official classifications from NESDB.

Poverty Data

• Income-Based Poverty: Ratio of population living below the national poverty line per total population in each tambon (subdistrict).
• Multidimensional Poverty Index: Based on census-based Basic Minimum Need data and register-based welfare card data.

Reference Period

• Income Poverty: 2015 and 2017.
• Multidimensional Poverty Index: 2017.

Methods Used

The study employs the following methods for poverty prediction:

• Generalized Least Squares (GLS): A statistical method that relaxes the assumption of homoscedasticity in OLS.
• Neural Network: A machine learning model inspired by biological neural networks.
• Random Forest: A powerful ensemble method capable of handling complex data structures.
• Support Vector Regression (SVR): A technique for regression analysis using the concept of support vectors.

Analytical Results

• Night Light Intensity: Strongly correlated with population density and poverty levels.
• Random Forest: Demonstrated the highest prediction accuracy among the methods tested, likely due to its ability to handle complex relationships in small to medium datasets.
• Variable Importance: Identified night light intensity and other population density proxies as the most significant predictors.
• Model Validation: 50% of data were used for training, and 50% for validation. Results were averaged across 100 resampled datasets.
• Comparison of Models: Random Forest outperformed other models in both income and multidimensional poverty prediction.

Key Findings

• Geospatial data can serve as a reliable alternative to traditional surveys for poverty prediction.
• Random Forest is particularly effective in this context due to its flexibility and ability to handle non-linear relationships.
• While computer vision techniques offer promising avenues, they are limited by the abstraction of features and difficulty in validation.
• Structured geospatial data provide a more transparent and verifiable approach to poverty modeling.

Implications

• The use of geospatial data can help fill data gaps in poverty mapping, especially in areas where survey data are not available or too costly to collect.
• This approach is especially useful during periods of economic disruption, such as the impact of the COVID-19 pandemic on poverty reduction.
• Future research should focus on the stability of these relationships over time and their applicability to other regions.

Conclusion

The study concludes that using precompiled geospatial data for poverty prediction is a feasible and effective method, particularly when combined with machine learning algorithms. It emphasizes the importance of validating models with structured data and recommends further research to assess the long-term reliability of these methods in poverty estimation.

References and Appendices

• References: Include studies on night light intensity, NDVI, and LST in relation to socioeconomic indicators.
• Appendices: Provide detailed lists of variables from geospatial data of 2015 and 2017.

Key Takeaways

• Geospatial data are increasingly being used as a cost-effective alternative to traditional surveys.
• Random Forest is a preferred method for poverty prediction due to its accuracy and interpretability.
• The integration of geospatial data with other socioeconomic datasets can enhance the precision and scope of poverty mapping.