Spatial Statistical Analysis for Poverty Mapping Using Machine Learning:

Abstract

Poverty is a multidimensional problem influenced not only by economic factors but also by spatial dimensions such as geographic location, accessibility, and environmental characteristics. This study aims to analyze spatial patterns of poverty and develop a poverty prediction model using a geospatial data-based machine learning approach. The data used comes from a combination of open sources such as the Central Statistics Agency (BPS), Landsat satellite imagery, and regional infrastructure data. The methods used include spatial autocorrelation analysis (Moran's I) to identify poverty clustering patterns, Local Indicators of Spatial Association (LISA) to detect poverty hotspots, and Random Forest and Gradient Boosting models to predict poverty levels based on environmental, social, and economic variables. The results show that poverty has a significant spatial pattern, where areas with high poverty rates tend to cluster in areas with low infrastructure access and high population density. The machine learning model demonstrated better prediction accuracy than the traditional linear regression approach, with an R² value reaching 0.87 and a lower prediction error rate (RMSE). These findings emphasize the importance of integrating spatial analysis and machine learning technology in understanding the dynamics of poverty geographically. This research contributes to the development of spatial data analysis methods in the context of public policy, particularly in supporting more targeted poverty alleviation intervention planning. The mapping results can serve as a basis for local governments in identifying priority areas, allocating resources, and designing data-driven development policies. Thus, this approach offers an innovative solution towards more efficient and evidence-based decision-making in poverty alleviation in Indonesia.